Hugging Face · Daily Papers

System One models promise fast, general-purpose decision-making without lengthy token generation. However, understanding their capability spectrum and decision reliability requires more than aggregate performance on current benchmarks. Existing evaluations often provide limited insight into the breadth of model capabilities and the reliability of their decisions. We introduce SO-Arena, a benchmark characterizing the capability spectrum and decision reliability of System One models across 11 domains and five capability families. For the capability spectrum, SO-Arena offers evidence-grounded evaluation through reliable sources and manual verification of reference labels against model-visible inputs; fine-grained capability assessment through domain-specific difficulty criteria; and realistic interactive evaluation, where models iteratively execute actions in sandboxes. Reasoning-intensive tasks further enable system one-two comparison. Beyond accuracy, SO-Arena assesses decision reliability through two complementary measures: decision stability under controlled option-set changes and missing-answer recognition when no valid option is available. Evaluations across 21 models reveal that the benefits of scaling and additional reasoning vary substantially across domains. More importantly, strong overall performance does not necessarily imply reliable decision-making: the best-performing System One model ranks only seventh in reliability. Together, these findings highlight the need to characterize capability spectrum and decision reliability as complementary dimensions, providing a systematic view of the strengths and limitations of System One models.

Driving world models aim to simulate how a scene unfolds under different ego actions. This requires plausible motion of surrounding vehicles as well as precise ego-trajectory control. Learning vehicle motion from perspective video alone is challenging because motion is entangled with appearance and vehicles can be occluded. Prior work therefore conditions video generation on layouts that specify vehicle motion, providing control over surrounding vehicles but requiring their future trajectories as inputs. We introduce BirdDreams, an autoregressive driving world model that jointly generates perspective video and a schematic bird's-eye-view (BEV) video, making scene layout part of the predicted future. The BEV, rendered from scene annotations for training, depicts roads and vehicles as simple shapes at a fixed metric scale, providing a top-down view of their positions and motion with reduced occlusion between vehicles. Perspective video, by contrast, retains scene details that the BEV omits, such as traffic signals and object appearance. Joint attention between two video diffusion transformers, one per stream, lets each stream inform the other during generation. The predicted BEV also provides a geometric interface for ego control: we translate and rotate its latest frame along an input ego trajectory and use the resulting frames as a soft condition that shows how the scene would appear to move if only the ego vehicle moved. At inference, BirdDreams generates both videos from a single camera image without requiring an HD map or vehicle-state inputs. On PhysicalAI-AV and nuScenes, BirdDreams follows input ego trajectories more faithfully than prior driving world models and generates more plausible ego motion without ego-action conditioning. Human evaluation shows that BirdDreams outperforms a perspective-only baseline in generating realistic surrounding-vehicle behavior.

Video generation models produce visually realistic videos but still struggle with physical dynamics and causal consistency. Learning from generated videos requires fine-grained physical supervision, yet reliable quality judgments, correction pairs, and controlled variations are difficult to obtain. Large language models (LLMs) possess physical knowledge, but struggle to reliably guide video generation models in producing their training videos. We present PhysLoop, a unified harness designed to support physically grounded self-improvement in video generation models. The harness guides and constrains LLMs throughout video generation, selection, and correction, distilling their physical knowledge into physically consistent videos. The framework supports controllable training-data construction and inference-time video repair. By identifying and correcting failures, PhysLoop marks a key milestone toward recursive self-improvement (RSI) in video generation models with limited human intervention. We release 100K videos spanning diverse, realistic indoor and outdoor scenes, along with generation tools and two evaluation datasets: PhysLoop-Gen and PhysLoop-Repair. We evaluate controllable generation and localized repair on these datasets, respectively, and assess physical behavior on the external Physics-IQ Verified benchmark.

In this report, we introduce the Reciprocal World Model (RWM) and present Duo v1, a full-duplex world model that continuously perceives audiovisual inputs while responding to users in real time. The RWM treats intelligence and physical reality as equally fundamental to what constitutes a world. Existing world models largely focus on simulating the latter, leaving reasoning and decision-making as external controls rather than intrinsic parts of the world itself. In RWM, we identify two equally fundamental capabilities: Intent, which provides the cognition underlying world interactions, and Realization, which generates coherent audio-visual behavior. To realize this formulation, Duo v1 introduces three key technical designs: (1) Leveraging autoregressive modeling for continuous interaction reasoning and participation decisions, rather than restricting it to instruction understanding for diffusion models. (2) Constructing chunk-specific conditions and synchronize tokens across Intent and Realization on a unified timeline. Module-specific attention masks enforce causal dependencies both within and across chunks, preventing access to future inputs. (3) Mitigating distribution shift in both modules through dual rollout distillation, training on autoregressively generated Intent tokens and synthesized audiovisual histories from Realization. Both modules continuously perceive real-time audio, video, and text, enabling full-duplex interaction with spontaneous conversations and interruptions. Beyond conversation, users can share real objects, reshape virtual scenes, guide assistants, or play audiovisual games without mouse or keyboard controls. The release of Duo v1 makes our first step, with the next version coming soon. We look forward to broader exploration of this emerging field.

Decision models answer a question with a typed output, such as a choice among options with a probability for each, so one model can choose an agent's actions, evaluate actions already taken, or score every action to teach a policy with reinforcement learning. Prior work has studied them mainly as actors, judging their usefulness by how well they act. We ask whether a decision model that acts poorly can still teach well. We compare three decision models and three general LLMs as actors and judges on ALFWorld, and use JEV as a teacher on ALFWorld and WebShop. As actors, all six models are pulled away from a subgoal that must come first by an easier one. As judges that see the whole trajectory, the same models credit the steps toward that subgoal, and asking whether each action helped, rather than whether the task will succeed, separates useful from useless actions. As a teacher, JEV raises a Qwen2.5-1.5B policy over GRPO from 72.8% to 82.9% success on ALFWorld and from 56.8% to 71.4% on WebShop. These results suggest that a poor actor can still teach well, because judging an action after the episode ends reveals the consequences that acting must anticipate. Code is available at https://github.com/ZJU-REAL/decision-agents.

Recent work suggests that pretrained image generators encode real-world understanding beyond image synthesis. Vision Banana exploits this by framing vision tasks as conditional image editing. However, its data, weights, code, and implementation details are not public. We present Open Vision Banana (OVB), a fully open and reproducible recipe for converting pretrained generators into unified multi-task perception systems. Our analyses identify the design choices that matter most for generalization, including color encoding, prompt design, parameter adaptation, and task balancing. These choices interact, so individually helpful choices can hurt when combined with others. Starting from 4B and 9B parameter open-weight base models and using 2.1M-3.8M training pairs from public datasets, OVB surpasses Vision Banana's reported Cityscapes zero-shot performance (73.9 vs. 69.9 mIoU) and rivals state-of-the-art affine-invariant depth estimators. Its non-zero-shot variants come within five points of Vision Banana on referring and reasoning segmentation but remain further behind on instance segmentation. All models largely retain the base model's image-editing abilities. With legend ensembling, we introduce a test-time scaling method for generator-based perception. It segments an image under several class-color legends and fuses the results by majority vote, correcting unstable color-class bindings and improving mIoU by up to 8.7 points on Cityscapes and 15.0 points on COCO-Object without retraining. We release the dataset construction and processing pipeline, the training and evaluation code, as well as the final, intermediate, and ablation checkpoints. Project website: https://openvisionbanana.github.io

Large foundation models such as GPT-6 Astra and Claude Fable 5.1 have recently been shown to control robots directly from visual observations. They nevertheless remain brittle on long-horizon, contact-rich tasks in real-world environments. Further progress depends on two complementary resources: demonstrations of human behavior that supervise how tasks are performed, and interactive environments in which policies can be rolled out, evaluated, and improved in closed loop. Language models have advanced along well-characterized scaling laws; the corresponding question for physical AI is how capability scales with environments and behaviors, which first requires identifying which data to scale and how to produce and reuse such data continuously. To address this question, we introduce Ropedia Xperience 2, comprising 1,589 released metric-scale 3D scans and a collection of 4,230 egocentric recordings of 74 tasks (131 hours). We further show that each capture can be repurposed for robot learning: scenes are converted into physically interactive simulation environments, human motion is retargeted to whole-body robot trajectories, and additional robot behaviors are obtained through virtual teleoperation and automated long-horizon trajectory generation. We propose this joint capture of environments and behaviors as a gold standard for physical AI data collection, laying a data foundation for establishing scaling laws over environments and behaviors.

Decoder-only Large View Synthesis Models (LVSMs) with KV-cache have recently achieved state-of-the-art quality by regressing novel views from neural networks without reconstructing 3D geometry, while Test-Time Training (TTT) layers save computation but sacrifice quality. We propose Test-Time Baking (TTB), a strategy that learns to bake global scene information from cross-view attention into a lightweight MLP, enabling fast and high-fidelity novel-view decoding. The first key is to repurpose the fast-weight MLP of TTT to learn the cross-view attention mapping explicitly from its input-output pairs, rather than replacing attention entirely, achieving the same rendering efficiency as TTT while delivering substantially higher quality. The second key addresses a modality misalignment between context prefilling and novel-view rendering: we disentangle context tokens into camera-only context rendering tokens to serve as TTB input, and context ground-truth tokens with image content to serve as the TTB mapping target, ensuring the baked MLPs are relieved of the modality generalization challenge between context and novel-view processing. Our method achieves up to **23x** rendering speedup with quality on par with the state-of-the-art across several benchmarks. Project page: https://test-time-baking.github.io

Aerial spatial intelligence (ASI) requires models to understand objects and spatial relationships across changing viewpoints and scales. Evaluating these abilities as models evolve calls for fresh physical observations, yet producing new aerial spatial questions requires repeatedly establishing entity identities, geometric references, and cross-view correspondences. We introduce RSLiveBench-UAV, a continually evolving vision-language benchmark driven by recurring acquisition of real UAV imagery. Its update protocol admits images captured after the preceding release and delays answer disclosure. An entity-centered production workflow uses geometry-assisted association to link repeated observations of physical vehicles, reuse human-verified attributes, and combine them with view-specific annotations and spatial references for traceable question generation. All tasks use deterministic scoring without model judges. We evaluate 30 general-purpose, remote-sensing, and embodied models on 35,491 question records from ten scenes, covering 17 tasks across 2D perception, 3D perception, cross-view perception, and reasoning. All 18 tested embodied models outperform the three selected remote-sensing comparators in Overall, although the strongest embodied model scores 43.29 versus 59.88 for GPT-6-Astra. Fine-grained vehicle recognition remains challenging, and performance varies across heights and camera pitches. Retrospective scene-group comparisons and rolling-window replay characterize score-pattern consistency under scene renewal. RSLiveBench-UAV connects fresh physical observations, entity-centered answer production, and systematic model evaluation to support continual assessment of ASI. Benchmark will release at https://rslivebench.space

Multimodal coding agents decouple 4D world construction from visual generation, repurposing video models from world generators toward generative renderers. This division of labor raises a fundamental question: if code builds the world, can video models just render it? A reliable generative renderer hinges on two complementary criteria: (i) world preservation, retaining prescribed 4D kinematics and camera paths; and (ii) visual realization, persistently binding text-specified target attributes over time. To evaluate this capability, we introduce WorldDraft, a benchmark of 432 strictly paired test cases across 48 executable 4D world families spanning six dynamic challenge categories. Under a 3 times 3 factorial design, WorldDraft orthogonally crosses draft granularity (abstract proxy to previs) and appearance alignment (neutral to conflicting) under identical dynamics. Extensive evaluations across 12 representative models reveal that: (I) world preservation and visual realization remain in stark operational tension; (II) richer draft scaffolding anchors structure but paradoxically impairs appearance fidelity; and (III) conflicting source cues trigger severe source leakage (up to 64%), exposing fragile cross-modal disentanglement. To mitigate these bottlenecks, we propose WorldDraft-Primer, a plug-and-play visual prompting strategy that consistently improves both objectives across frozen video renderers. Together, WorldDraft and WorldDraft-Primer lay a foundation for bridging programmatic 4D simulation with video generation, advancing physically grounded generative rendering.

Second-order optimizers (e.g., Shampoo, SOAP) have emerged as compelling alternatives to Adam/AdamW and promise faster per-step convergence than AdamW in transformer training. However, they still remain costly in practice, and even SOAP still pays substantial optimizer-state memory and eigendecomposition overhead. We ask whether the benefits of SOAP can be retained at nearly the efficiency of AdamW. In this paper, we answer yes with a lightweight SOAP variant, Factored One-sided Adam-Moment (FOAM). The path to practical impact is twofold: first, FOAM employs input-only preconditioning, dropping the output-side Kronecker factor to reduce memory and remove one eigendecomposition. The Kronecker structure of the Hessian shows that this is exact for a diagonal output-side curvature and nearly lossless for a near-diagonal one, as we observe empirically. Second, FOAM replaces SOAP's dense rotated second moment with a factored Adam-moment estimator based on row-by-column statistics. Under the K-FAC Fisher assumption with a diagonal output curvature, this estimator exactly recovers SOAP's dense second moment, and we measure only a 15-20% K-FAC violation, with negligible effect on the update. Our empirical results support the claim that FOAM matches strong SOAP and Muon baselines in validation loss across scales of large language model (LLM) pre-training, uses half of SOAP's optimizer state, runs at Muon-class wall-clock time, and retains robustness across a three-fold learning-rate range. The results hold from 124M to 760M parameters, demonstrating that SOAP's gains need not require SOAP's cost. Code: https://github.com/naver-ai/foam

AAA games offer a rich testbed for evaluating agent perception, reasoning, and control. However, evaluating whether agents can complete tasks in AAA games is challenging and often human-annotated because task outcomes are often difficult to verify automatically, let alone applying it for Reinforcement Learning (RL). We introduce GTAGym, an environment and framework for evaluation and RL on GTA V tasks. Its evaluation benchmark, GTAVBench, derived from GTA V main-story missions, contains 103 online tasks reconstructed across nine task families, alongside 47 image-based shooting and sniping tasks using GTA V imagery. A web server provides access to the online environment and evaluation interface, supporting parallel testing through screenshots and keyboard-mouse actions with automatic engine-state outcome checks. We evaluate six frontier LLMs on GTAVBench with textual task instructions and on its GTAVBench-AGI branch, where instructions appear only in the first screenshot and a harness automatically compresses interaction history. The best configuration achieves 70.9% overall accuracy in GTAVBench-AGI, yet aerial navigation, evasion and interception remain difficult. GTAGym also provides a scalable generator that samples task parameters within user-specified ranges and option sets under fixed templates, creating batches of 3D and image-based shooting environments for RL with automatic reward and termination signals. LoRA RL of Qwen3.8-27B on generated on-foot navigation tasks improves the in-domain performance and transfers to out-of-domain task categories, and native GTA V tasks from 15% to 23%.

We present UniSpark-9B, a unified multimodal model that investigates hybrid attention for visual understanding and generation. Built on a Mixture-of-Transformer-Experts backbone with approximately 9B parameters in each branch, UniSpark combines full attention for explicit token interactions with GatedDeltaNet linear attention for compact context propagation. Autoregressive text prediction and visual flow matching jointly support visual question answering, text-to-image generation, instruction-guided editing, and interleaved text-image prediction. In interleaved generation, noisy visual targets are isolated from subsequent predictions, while clean visual representations provide context for later outputs. A blockwise protocol enforces this distinction through attention masks in full-attention layers and recurrent-state retention rules in linear-attention layers. The selected training sources contain 159.7M generation samples, 213.4M understanding samples, and 624.8K interleaved sequences. A four-stage curriculum jointly trains multimodal understanding and visual generation, refining answer accuracy and instruction following while extending image synthesis to 2K. UniSpark scores 2470.3 on MME-S and 0.83 on GenEval. Qualitative examples further illustrate its capabilities in interleaved image generation, video synthesis, and video editing. Together, these results demonstrate that hybrid attention provides an effective backbone for unified multimodal understanding and generation.

While search agents demonstrate impressive capabilities in multi-step question answering, their robustness to poor-quality evidence remains under-explored. Such evidence is rare in benchmark corpora, and resulting failures are hard to reproduce. We propose DeepStress, a dynamic stress-testing framework that subjects agents to degraded evidence at controlled rates. It replaces the retrieval module with a synthetic environment where documents are generated dynamically for each query, under independent control of their trustworthiness, relevance, and factuality. To characterize behavior under stress, we go beyond binary accuracy with a five-way classification of final-answer outcomes and two new reliability-aware metrics, RAS and CoPRAS, which credit cautious behavior and account for computational cost. Evaluating 12 systems on questions from HotpotQA and BrowseComp-Plus, we find that several open-weight agents answer confidently when sources contain obvious credibility red flags, and that many HotpotQA questions are answered from parametric knowledge even when all returned documents are factually wrong.

How multimodal language models should encode visual input remains up for debate. Most models use a pretrained vision encoder, while some recent recipes train the encoder from scratch with the language model or remove it entirely. Vision encoders bundle together three design choices: a visual prior from upstream pretraining, dedicated vision capacity, and pre-fusion placement of that capacity (i.e., before the multimodal backbone). Because published models also differ in data, architecture, and compute, comparing them cannot isolate the effect of each choice. We disentangle these choices via a controlled framework where we pretrain all models from scratch on common data and compare them on visual question answering and language benchmarks at matched lifetime compute. We find that the visual prior is not necessary; an encoder trained jointly from scratch is on par with a pretrained one, without a separate training stage. What matters instead is dedicated vision capacity: an encoder-free model trails on visual understanding even when its backbone is enlarged by the encoder's parameter count. Yet that capacity need not precede multimodal fusion, as vision-only modules added in-fusion to early backbone layers perform comparably to a pre-fusion encoder and can lower time to first token. These results also hold across backbone sizes, when encoder pretraining is treated as free, and when starting from a pretrained language model. Overall, our findings favor dedicated vision capacity trained jointly with the backbone and placed early within it.

GPT-6 Astra can drive a robot arm directly through a fixed harness, and scores of such robot-use agents on RoboDojo have risen quickly since. Reported systems differ in harness and test subset and often change several designs at once, so it is hard to tell which choices produce the gains. We take these designs apart and test them in one framework, along the harness around the model, the action interface, the reuse of experience and the scene geometry given to the model. A lean harness that keeps the full episode history and stops each motion early goes a long way. Reusing experience changes how the model acts but is not a free lunch, and scene geometry helps most, even when depth is predicted from RGB alone. JustRUA starts from the official L3 harness and keeps only what matters most. Even with no learned policy, self-improvement loop or task-specific tuning, this simple harness raises the score on the RoboDojo simulation benchmark substantially, from 29.7 to 50.3.

Transferring LLMs into System One decision models typically replaces the language modeling head with a decision head while retaining the full backbone, making every decision pay for all layers. Yet it remains unclear how much depth such decisions require. We find that decision information is concentrated in intermediate layers, with early layers building essential features and the final six layers removable without loss, while independently trained smaller models largely mimic the full model rather than developing specialized decision capabilities. These findings suggest that efficiency requires not simply smaller backbones, but shared computation across different depths. Building on this, we propose , which jointly trains a full model with prefix sub-models that share its final layers and decision head through self-distillation. A lightweight gate reads the hidden state at each exit to determine whether additional computation is needed, allowing one model to operate at multiple layers. Across seven decision benchmarks, our approach matches or outperforms independently trained models at every depth, while achieving better accuracy-efficiency trade-offs through adaptive computation.

Reference-frame stabilization registers every frame of a hovering-drone video onto one reference frame, keeping object positions comparable. The registration pipelines it relies on are selected on feature-matching benchmarks in which only the camera moves. Over a busy intersection, vehicles moving together supply mutually consistent correspondences that can draw the robust homography estimate toward their motion: consensus capture, which inlier statistics cannot flag. A two-part benchmark measures it and the object masking meant to prevent it. A static protocol scores ten pipelines under exactly known homographies on 4K aerial frames, tuned on a site-disjoint split at a matched resolution and feature budget. A dynamic protocol scores them at an offline and a real-time budget on hovers selected from the video alone to span the amount and coherence of moving foreground, plus a stationary control. It measures background alignment, object position error and the share of a moving group's displacement an estimate follows, against a pseudo-ground-truth reference audited with manually annotated ground control points, which also score every candidate. On static scenes, tuned AKAZE, RootSIFT, SIFT and BRISK lead, retuning the matcher thresholds raises the mean average accuracy (mAA) by at most 0.0062, and masking costs less than 0.005. Unmasked ORB is captured on 28 of 304 real-time dynamic pairs and on 6 offline; its real-time captured estimates have a median corner error of 18.8 px yet typical inlier counts and ratios. Masking removes capture for every pipeline and leaves the leading pipelines unchanged. Code, data and results are released.

Distribution matching distillation (DMD) has emerged as a leading paradigm for distilling diffusion models into few-step students, where an auxiliary fake score model tracks the student's distribution and the discrepancy between the teacher and fake predictions drives the student toward the teacher. In multi-step distillation, although only the student's final output is delivered, existing methods train the fake model on intermediate states or across all denoising steps. Consequently, the student's update signal becomes entangled with intermediate distributions, which can distort the matching direction and remain nonzero even after the student's final outputs already match the teacher distribution. To resolve this mismatch, we propose Endpoint Distribution Matching Distillation (eDMD), which trains a single shared fake model exclusively on the student's final outputs and uses it to correct every denoising step. Our theoretical analysis proves that this endpoint reference guarantees that the distribution matching term vanishes if and only if the student's final output distribution matches the teacher's, regardless of intermediate distributions. Beyond the fake reference, we further strengthen teacher supervision through a precise teacher (PT) that replaces single-step guided predictions with short multi-step trajectories on the final noise interval, paired with a calibrated query density for intermediate steps. Experiments on SD3-Medium, FLUX.2-klein (4B and 9B), and Qwen-Image-2.1 demonstrate that eDMD consistently outperforms existing distribution matching methods for 4-step generation.

A small fine-tuned parser that emits a well-formed but wrong tree fails silently: nothing downstream can tell. We ask whether the output format, a free design choice, controls how often this happens. On the PIZZA ordering benchmark we fine-tune two sub-billion models (Qwen3.5-0.8B and LFM2.5-350M) with LoRA on six lossless encodings of the same target tree, from 44.3 down to 10.9 tokens per order, including two matched-length pairs that vary surface form at fixed length. Across 300 runs (5 training pool sizes, 5 seeds) with a hierarchical bootstrap and Holm correction fixed before the results were read, shortening alone never helps, but writing the target in ordinary words does. The word encoding is the most accurate format on both models (87.5% vs. 85.8% exact match for the native s-expression on Qwen) while emitting 4.1x fewer tokens, and it cuts the silent error rate by 2.3x on Qwen and 1.8x on LFM2, separating at every pool size. Error analysis locates much of the gain in negation. All runs also learn to refuse off-menu input (held out: 95.9-99.0% handled) and to apply in-utterance self-corrections. Strict refusal has a measured price: 2.1-10.1% of real, mostly polite, orders are refused.

Existing adaptive-reasoning methods shorten the explicit reasoning a model writes, or skip it by switching modes; either way the answer is still written out token by token. Rather than deciding how much to think, we explore a different paradigm, zero-or-T thinking: a single model either answers from the implicit thinking inside one forward pass, generating no token, or performs full autoregressive reasoning at a cost of T tokens when that is worth paying. We realise it in a hybrid decision model (HDM), where one frozen backbone serves both modes and a lightweight self-routing block chooses between them. The block estimates the value of thinking, the utility the slow mode is expected to add over the fast one, from what that implicit thinking has already produced, and the slow mode runs only where that value exceeds a price λ set per request. Where no per-decision label exists, as in episodic agentic tasks, we introduce Calibrated Decision Policy Optimisation (CDPO), a reinforcement-learning recipe that trains the block from episode-level returns. We evaluate HDM on MMLU-Pro, AGIEval, GPQA-Diamond, JevBench, BANKING77 and MiniWoB. Specifically, on JevBench HDM-4B scores 0.857 against Jev's 0.851 (+0.7%) at a median 0.179\,s against 0.652 (-72.5%). On the agentic venue MiniWoB HDM-27B cuts token cost by 80.1%.

Full-duplex spoken dialogue models learn turn-taking, overlaps and backchannels from conversations recorded with one channel per speaker, but most available conversational audio is a single mono mix. We present Kezek, a model that turns a mono recording of exactly two people into one track per person with the original length and timing; overlaps and short replies ("yeah", "mm-hm") stay on the track of the person who said them. Kezek decides who speaks when and who is who once for the whole recording (stage 1: a two-voice activity model guided by a speaker fingerprint trained to be insensitive to phase), and a conditioned time-domain separator only carries out that decision (stage 2), with outputs that always sum to the input. Standard separation metrics miss the failures that matter for dialogue data, so we report short-reply retention, leakage onto the other person's solo, channel swaps and coverage. Against the open pipeline pyannote -> DialogueSidon on the same recordings, Kezek keeps more short replies (0.83 vs. 0.69 on Open Yap test; 0.80 vs. 0.71 over the other's speech on unseen telephone speech from CallHome English), mostly because it returns the whole recording while the external pipeline drops part of it (on the audio both return, reply retention is about the same: 0.859 vs. 0.865 on Open Yap); it also leaks less, swaps channels less often, and runs 4-5 times faster (about 230 hours of audio per GPU-hour). We also report negative results and an ideal-mask ceiling that locate the remaining gap: quiet replies under a louder voice. Code and weights are released under permissive licences.

We present EvenVLM, the first text and multimodal (VQA) dataset for Even (Lamut), an endangered Northern Tungusic language spoken by approximately 6,800 people in Siberia. Our dataset comprises 50,521 structured Even-Russian entries extracted from dictionary and phrasebook sources via GPT-4o Mini-based parsing, from which we synthetically generate 107,543 training examples spanning translation, grammar, reading comprehension, and semantic classification. To address tokenization inefficiency on Even's agglutinative morphology, we propose a three-step tokenizer expansion algorithm that adds Even-specific graphemes (eng, oe, o-umlaut), whitelisted vocabulary, and BPE merges, reducing average tokens per word by 27% (4.99 to 3.64). Using LoRA fine-tuning of Qwen3-VL-4B, we adapt the model for text tasks (0.53% trainable parameters, rank 32) and separately for visual question answering (rank 64, frozen vision encoder). Our best configuration achieves 11.4% exact match and a 13.6 weighted score on text benchmarks, 0.1283 chrF for sentence translation, and 65% accuracy on VQA (up from 0% for the base model and text-only fine-tuning). We release our datasets, evaluation benchmarks, and training pipeline to support further research on Even and other low-resource languages.

ZONOS3-Mini is an ultra-compact, ultra-fast, zero-shot voice-cloning TTS model. The model can run in real time on a laptop or smartphone. For its latency and size, it achieves state-of-the-art accuracy and cloning fidelity. To achieve this we propose a latent MoE backbone with only 886M parameters (100M active per token) and a new audio codec, Zyphra Audio Codec 1 (ZAC-1), which represents 44.1 kHz audio at a 15 Hz frame rate. On the new ZTTS2-Eval benchmark, ZONOS3-Mini reaches a 6.71% macro-average WER over 55 language sets (35 clean and 20 in the wild), which is competitive with larger and slower state-of-the-art models. We release the model weights, inference, and training code.

Data-driven surrogate modeling has emerged as a powerful paradigm for accelerating computational fluid dynamics, and hierarchical U-shaped networks, which resemble the structure of classical Geometric Multigrid (GMG) methods, have become prevalent designs. Modern surrogates built on these designs achieve competitive accuracy at a fraction of the cost of numerical solvers; however, accurate prediction of complex flows remains challenging, and no existing architecture leads across them. In such networks, the multigrid analogy mainly concerns the multi-resolution topology, while their components largely follow common vision practice. To address this, this study introduces SUMNO, a multigrid neural operator for fluid flow fields on regular grids that spatially unrolls a V-cycle into a feed-forward U-shaped network. Parameter-free down- and upsampling stand in for restriction and prolongation, and additive skip connections for coarse-grid correction, with normalization-free transfer and skip pathways. Within this macro-architecture, convolutions are used at the fine and intermediate resolutions and shifted-window attention at the coarsest, whose window partitioning is generalized to handle periodicity independently per axis. Evaluated on unseen test configurations across four unsteady and steady flow benchmarks against five representative baselines, SUMNO is best in 5 of 7 settings and ranks second in the remaining two. It reduces the 10-step rollout MSE by about 30% relative to the strongest baseline, and outperforms three baselines with over 130 million parameters using 9.4 million parameters.

It is common to train Large Language Models (LLMs) for agentic tasks with Reinforcement Learning (RL) from scalar rewards. However, this throws away the fine-grained information about the environment that the agent experienced during the trial (e.g., error logs). To utilize this rich information, recent works wrap the agent in a harness that refines its own inputs through reflection or compaction and hands past experience to future rollouts, but the construction of the handed-off artifacts (e.g., a playbook) is itself not learned. We want the best of both worlds: gradient updates from a scalar reward, while also utilizing environment information to hand off experience. As a concrete testbed, we fix the Agentic Context Engineering () harness and train one model to act, reflect, and curate task-local playbooks. Our method, , evaluates candidate experience-handoff artifacts by their downstream effect on future actor rollouts, and co-optimizes the role-specific behaviors (action generator, insight reflector, and playbook curator) under a fixed scaffold. Across agentic benchmarks, achieves higher held-out performance than baseline methods. The results suggest that agent training should optimize not only the actions in the current rollout, but also the model-written experience that future rollouts inherit.

Streaming video understanding aims to continuously process incoming visual observations, making the retention and retrieval of relevant information increasingly important as the video history grows. Although existing benchmarks include memory-related questions, question-answering accuracy alone does not reliably measure memory capabilities. Incorrect predictions may arise from failures in visual understanding rather than forgetting, whereas correct predictions may rely on linguistic priors or recent observations without requiring historical memory. To address these ambiguities, we introduce MemInStream, a benchmark dedicated to evaluating memory capabilities in streaming video understanding. We develop an evidence-controlled validation protocol that associates each question with temporally localized evidence intervals. The protocol validates evidence solvability and memory necessity by retaining questions that models can answer from the annotated evidence but cannot resolve using recent observations, post-evidence observations, or textual input alone. This helps disentangle memory limitations from understanding failures while reducing memory-independent shortcuts. MemInStream provides a systematic evaluation of memory capabilities through four major categories, covering Episodic Retrieval, State Tracking \& Updating, Multi-Evidence Integration, and Contextual Memory across diverse video scenarios. Overall, MemInStream provides a comprehensive and reliable benchmark for systematically evaluating memory capabilities in streaming video understanding.

Humanity has long sought to explain natural phenomena. Reliable prediction of experimental outcomes is a central test of our understanding of the processes that produce them. If artificial general intelligence (AGI) is to surpass human scientific capabilities, it must also predict experimental outcomes more reliably than humans. Such capabilities could accelerate scientific discovery across disciplines, help address humanity's most pressing challenges, and improve human well-being. We introduce Science Prediction Universe (SPU), a benchmark of hidden experimental outcomes with 100 public verified tasks from chemistry, materials science, biology and medicine, and 16 private tasks from unpublished wet-lab experiments. Given an experimental setup and candidate conditions, an agent works with any resources it needs to rank the candidates by a measured outcome. We evaluate thirteen agent configurations spanning twelve frontier models across two harnesses (Codex and Claude Code). Current agents remain far from saturating wet-lab outcome prediction, with substantial performance disparities across models: on the public benchmark, configurations separate into two distinct tiers, with task-adjusted Spearman correlations ranging from 0.227 for the lowest-performing configuration to 0.478 for the best. Analysis of 1,155 scored public trajectories (227,103 tool calls) suggests that the bottleneck lies in scientific judgment rather than tool execution: agents rely on shared monotone priors, agreeing far more with one another than with experimental reality (0.52 vs. 0.33), and frequently compute trivial physical or chemical proxies that do not influence the final outcome.

Adaptive video streaming optimizes Quality of Experience (QoE) metrics by selecting appropriate bitrates according to varying network bandwidth and user demands. In practice, however, real-world network bandwidth often exhibits heterogeneity relative to training environments. Current methods predominantly tackle this problem through learning-based approaches designed to improve generalization performance. Our systematic investigation reveals a critical limitation: neural networks suffer from plasticity loss, significantly impeding their ability to adapt to heterogeneous network conditions. Through theoretical analysis of neural propagation mechanisms, we demonstrate that existing dormant neuron metrics inadequately characterize neural plasticity loss. To address this limitation, we have developed the Silent Neuron theory, which provides a more comprehensive framework for understanding plasticity degradation. Based on these theoretical insights, we propose the Reset Silent Neuron (ReSiN), which preserves neural plasticity through strategic neuron resets guided by both forward and backward propagation states. Moreover, we establish a tighter performance bound for ReSiN under non-stationary network conditions. In our implementation of an adaptive video streaming system, ReSiN shows substantial gains under non-stationary conditions in which the available bandwidth alternates between high and low regimes, improving the QoE by 178%, and it remains beneficial under stationary real-world traces, where it improves the QoE by 1.6%. Code will be available: https://tinyzqh.github.io/ReSiN/

Memory decides which LLMs an edge device runs, and how fast: capacity on a 4 GB Jetson Nano, bandwidth on an 8 GB Jetson Orin Nano Super. Deployed formats (GPTQ, AWQ, llama.cpp's K- and I-quants) spend equal bits on every integer code of a matrix, although the codes are far from uniform. LazyWeights keeps these formats but tunes GPTQ so the codes compress, Huffman-codes them losslessly in interleaved lanes, and decodes them on demand in a small unit between DRAM and L2, so GPU kernels run unchanged and DRAM moves only coded bytes. At 1% and 3.5% perplexity budgets, coded Phi-3-mini, Mistral-7B and SmolLM-1.7B are 5.7-31.4% smaller than the best llama.cpp format. A bit-exact 0.21 mm^2 ASIC decoder unit (1 GHz, 7 nm) cuts Orin latency and energy per token by up to 36% and 38% (54% and 44% with prompt-lookup speculation). On the Nano, LazyWeights needs 58-64% less time and 53-57% less energy per token than CPU llama.cpp, and runs a streamed Mistral-7B 6.6times faster at 3.89 bits per weight. A Huffman decoder beats an rANS decoder in rate and area at equal throughput; the ASIC decoder unit runs bit-exact on a Zynq UltraScale+ FPGA (AMD ZCU102).

Frontier video generation models work in a VAE latent space that compresses video in both space and time. This compression made training tractable, but now limits high-resolution, controllable, and interactive generation -- doubling resolution quadruples token count, conditions apply to groups of frames rather than individual ones, and each response must wait for the rest of its multi-frame block. We present PixVid, a simple yet effective method that lifts a pretrained latent video model to directly generate every frame in pixel space. PixVid keeps the diffusion transformer and replaces the VAE decoder with a convolutional head that renders its outputs into pixels. We find that predicting clean video rather than velocity, together with a larger noise shift, is key to quality. Finetuned from Wan2.1-14B, PixVid is competitive with the latent baseline within sim 1.3M training clips. Working framewise in pixels, it generates at 2times higher resolution with the same per-frame token count, keeps depth-conditioned video editing aligned under fast motion, and responds to each action at the next frame. Project page: https://chenshuo20.github.io/PixVid/

General-purpose language models are trained mostly on English: over open models we review, at most, non-English pretraining data makes up only a 12% share. Models built for European languages give them far more, but score well below general-purpose models of the same size. A common concern is that, in small models, languages compete for capacity. We present Lara Mini base, a model trained from scratch on 6T tokens in English and 44 other languages, which together cover all 24 official languages of the European Union. As training progresses, the 44 other languages together come to outweigh English. To give these languages enough capacity at low cost, Lara Mini base uses a mixture of experts architecture with 7.2B total and 0.6B active parameters. Among open base models with up to 2B active parameters, it achieves the highest average score across standard English, multilingual, math, and code benchmarks, ahead of models trained on 4 to 6 times as many tokens. Its margin is largest on multilingual benchmarks in the languages it supports and on math benchmarks. It supports contexts of up to 128k tokens. We find that our MoE architecture reduces capacity competition between languages. For a single user on one GPU, without quantization, it generates text at least 1.8 times faster than any model with a higher average score. We release the weights of Lara Mini base under the CC BY-NC-SA 4.0 license. In addition, we will release T-HellaSwag, a higher-quality machine translation of the commonsense reasoning benchmark HellaSwag into 110 language variants.

Speech recognition (ASR) systems are usually compared by word error rate (WER) on recorded files, after punctuation, capital letters and number format are removed. In dictation, the audio is streamed live and the text goes into the user's message as written, so what matters is which mistakes the user has to fix. The test set, streamed to eleven ASR systems, comprises 106 scripted English dictations from one speaker (read in a quiet room, with a fan on, and whispered), 20 clips in each of eleven other language and dialect sets, and 400 recordings that join clips in two languages. Every error was sorted by a fixed rule into four levels: recognition, output formatting, language handling and, for the self-hosted models, audio dropped before the model. Systems with similar WER failed in different ways. Zoom Scribe and OpenAI's GPT-Live-Transcribe have almost the same WER on quiet English (7.4% and 7.3%), but 34.7% and 8.8% on whispered English. Zoom and both AssemblyAI models left out 26-28% of whispered words, and still 21-26% with their end-of-turn detection relaxed or off; seven other systems left out at most 3%. Qwen3-ASR's 32% came from the speech detection in our own deployment (6% when relaxed). Across 107 EARS speakers, the same four systems lost 7-31% of whispered words, while six others lost under 2%. With only English selected, Gemini, Deepgram and our self-hosted Whisper lost the non-English part of a two-language recording in 86-99% of recordings, against 33-46% with both languages selected, while GPT-Live-Transcribe, OpenAI's realtime Whisper and Meta lost it in at most 20%. Running Whisper behind a Silero voice activity detector that drops audio judged to be silence raised whispered WER from 10.1% to 51.4%. Most English results come from one read speaker. The code, the recordings, both references and every transcript are released.

Post-training of large language models typically combines supervised fine-tuning (SFT) with preference-based fine-tuning such as Direct Preference Optimization (DPO). Human-annotated and synthetically generated preference data differ substantially in per-sample cost, but prior work typically holds sample count fixed when comparing the two. In this work, we explicitly study the interplay between cost and performance by comparing the effect of post-training with human and synthetic preference data under a fixed, dollar-matched annotation budget. We train and evaluate multiple base model families across helpfulness and summarization tasks. We find that synthetic data is far more cost-efficient at low-to-moderate budgets, achieving high win rates with as little as \200--400. On helpfulness, human data closes the gap only at budgets roughly two orders of magnitude larger; on summarization, it does not close the gap at any tested budget. We also find that human-annotated data varies in reliability: one widely used human preference source substantially underperforms another, and the gap persists even after standard remedies such as domain restriction and deduplication.

We present Kosmo-GFM, a scene reconstruction system that integrates a geometry foundation model (GFM), a pose optimization engine (PoseEngine), and reconstruction-derived supervision from KosmoWorld. GFM predicts camera parameters, depth, cross-view correspondences, and dynamic 3D points from images and videos. PoseEngine combines these predictions with geometric optimization to recover scene-level camera poses and geometry for downstream 3D and 4D Gaussian Splatting. KosmoWorld generates images and geometric supervision from reconstructed real-world scenes to improve GFM's geometric predictions. On the Sparse and Medium subsets of SpatialBench, GFM achieves the lowest absolute relative depth error and the highest relative-pose AUC at 5 degrees among the compared methods. It also achieves the lowest mean endpoint errors for dynamic points and scene flow across the five evaluated datasets. Across 34 scenes completed by all compared reconstruction systems, PoseEngine achieves the highest scene-averaged pose AUC at thresholds of 1, 3, and 5 degrees, with a cumulative runtime of 14.07 hours, approximately 20% of GLUEMAP's 69.67 hours. PoseEngine Flash provides a further accuracy-efficiency trade-off, with a measured mixed-precision runtime of 2.62 hours, approximately 4% of GLUEMAP's runtime. Separate scaling experiments show generally favorable trends with increased data, model computation, and training budgets. These results support the use of learned geometric priors within a scene-level reconstruction pipeline.

Speculative decoding accelerates text generation with large language models by letting a cheap draft model propose tokens that a large verifier checks in a single forward pass. We introduce Olifant, a draft model built on a memory-based classifier. It predicts the next token from a four-token window in under 0.3 ms on CPU, requiring no GPU, no backpropagation, and no accelerator memory. We measure throughput and energy under two verifier regimes on consumer hardware. Without a KV cache, drafting is nearly free: a domain-matched Olifant model reaches 38.1% acceptance on EU legal text, delivering 2.43x throughput and 2.51x lower energy per token than the neural verifier model alone. With a KV-cached verifier, itself a ~10x energy saver, drafting costs matter: a TinyLlama draft model halves throughput and triples energy despite high acceptance, whereas Olifant still yields 1.25x throughput and the lowest energy in the study (93 J per 100 tokens, 14x below the uncached baseline). We further show that KV caching decouples time from energy: verification of longer drafts raises power draw, shifting the energy-optimal draft length below the speed-optimal one. Speculative decoding is not dead in the age of fast verifiers, provided drafting is nearly free.

Image-to-3D generative models now produce asset-quality textured meshes, but at minutes per asset they remain slow to iterate with and costly to deploy. Prior work has mostly accelerated the sampling of the networks, and where it also optimizes the mesh processing, it trains its own generators. We instead start from a publicly released pipeline and ask how much of its end-to-end time can be saved. Profiling the public Pixal3D pipeline shows that its latency is dominated by what happens after the networks: the four rectified flows take 14% of the wall-clock time and mesh post-processing takes 85%, most of it in CPU-based UV unwrapping. We therefore first rebuild the mesh post-processing on the GPU and take it off the critical path, then cut the per-evaluation overhead of the flows and distill them from 76 to 8 network evaluations. The resulting pipeline, Sprint3D, reduces the median time per asset from 133 s to 4.05 s on one H200 GPU at the same export settings (33times). At the same time, on 2,653 held-out Toys4K objects it improves every input-view normal metric of the Pixal3D protocol, with a slightly lower mean silhouette IoU, lower point-cloud alignment scores and less fine texture detail. The rebuilt post-processing alone also speeds up TRELLIS.2, which uses the same export code, by 6.7times without a detectable change in the evaluated metrics.

When language models revise a decision after another agent disagrees, neither reversal counts nor final accuracy shows whether they are fixing mistakes or abandoning correct answers. We introduce a controlled two-pass audit that holds the question, the evidence, and the model's initial answer fixed and varies only the revision context: neutral reflection, anonymous disagreement, or disagreement attributed to a domain expert. For binary decisions the accounting is exact: revision raises accuracy only if it is selective enough, reversing wrong answers more than a/(1-a) times as often as correct ones, where a is initial accuracy (nine times at 90%). Across five models and 1,989 evidence-grounded decisions in medicine, science, and law, no model clears this bar under expert disagreement, and accuracy falls by up to 16.1 points. The expert label, a prestige cue, adds reversals for four of the five models, yet its effect on accuracy ranges from a 5.2-point loss to a 2.4-point gain, so reversal counts are a poor proxy for harm. With identical arguments from both sources, the label still raises reversal for three proprietary models and in a pooled analysis of four newer models. Models also respond to stated source reliability. Revision pipelines should therefore be evaluated with paired, conditional metrics, per model and protocol.

Planning for a heterogeneous robot team from a scene image and an instruction involves two kinds of decisions. Some must be judged from the scene: which goals are already met, how the task splits, and which robot does what. Others follow from these and can be computed: prerequisite actions, handoffs between robots, and timing. We argue that each kind should go to the component suited to it: a small vision-language model learns the first, and a deterministic solver computes the second. In NSDL (Neuro-Symbolic Division of Labor), the trained model writes its judgments as a plan sketch, which lists the goals that still need action and the intended steps in order, each with a robot. The solver expands the sketch into an executable joint plan. The model calls two tools, a detector for evidence and the solver, and is trained with the solver in the loop. We evaluate on the public VIKI-L2 benchmark and on LOCUS-Bench, a new benchmark that grades the two kinds of difficulty separately. NSDL outperforms end-to-end planners and formalize-then-plan baselines on both benchmarks and with three backbones from 3B to 8B; the 7B model reaches 97.0% out-of-distribution accuracy on VIKI-L2. Ordering and scheduling errors nearly disappear once the solver writes the structure; the remaining failures are misjudged states, and most goals the model misses are ones it never lists. Code, model weights, and LOCUS-Bench are publicly available at https://github.com/Xincheng-Tan/NSDL.

Advances in vision-language models (VLMs) have improved remote sensing image analysis and supported geospatial decision-making. However, fine-grained understanding of remote sensing imagery remains challenging for VLMs. Large variations in object scale, complex geospatial relations among instances, and limited visual detail hinder precise instance localization, visible attribute identification, and spatial relation understanding, potentially leading to descriptions that lack visual grounding. Existing datasets and benchmarks insufficiently assess end-to-end instance localization and captioning without target-specific textual cues. To support this evaluation, we introduce RICapBench, a dataset and benchmark that unifies instance localization and fine-grained captioning in remote sensing. RICapBench contains 245,466 instance-level annotations and 498,005 spatial relation tuples spanning general remote sensing, UAV remote sensing, and dense small-object scenarios. Each annotation pairs an object's oriented bounding box with a natural language description of its visible attributes and spatial relations to surrounding instances and environmental elements. To evaluate the accuracy and completeness of spatial relation captioning, we propose a remote sensing spatial relation parser and the RelCap-F1 metric. We combine RelCap-F1 with instance localization and attribute captioning metrics into a composite metric, RICScore. Our evaluation of 15 advanced open-source and closed-source VLMs identifies instance localization as a major bottleneck for these models. LocateAnything achieves the highest RICScore of 31.68, but scores lower on spatial relation captioning than other task-fine-tuned open-source VLMs. Code and dataset are available at https://github.com/Xldln/RICapBench.

On-policy distillation (OPD) has emerged as an effective post-training approach to improving fine-grained perception in multimodal large language models by amortizing costly inference-time zooming operations into a single-pass recipe. However, vanilla OPD is inherently susceptible to linguistic momentum: the teacher can follow the student's erroneous prefixes, thus weakening corrective supervision. To address this, we propose Blink On-Policy Distillation (BlinkOPD), in which the teacher "blinks" at every token: it reads the same prefix once with its eyes open and once with its eyes closed. A first-glance gain further strengthens the correction on rollouts that conflict with the teacher. By extracting more effective supervision from each on-policy trajectory, BlinkOPD achieves state-of-the-art average performance across six fine-grained benchmarks using only 500 training images (12x less data) than existing OPD methods. Beyond fine-grained perception, BlinkOPD also demonstrates strong zero-shot improvements across ten held-out benchmarks spanning general visual perception, visual hallucination, and multimodal reasoning abilities. Code: https://github.com/Supltz/BlinkOPD

Calibrated multi-view hand reconstruction has advanced rapidly, yet the effectiveness of existing methods depends strongly on real captured data, which is costly to acquire, limited in hand-geometry diversity, and often affected by imperfect annotations. Synthetic data offers a scalable alternative with accurate supervision, but a substantial synthetic-to-real appearance gap hinders its use. To address this, we introduce Hand in Silhouette (SilHand), a calibrated multi-view hand reconstruction framework that operates on binary hand masks, obtainable at low cost either automatically or manually, to reduce the synthetic-to-real observation gap. Rather than relying on RGB features for 3D lifting, SilHand projects metric 3D hypotheses into each view to query and fuse mask-ray evidence for hand reconstruction. This mask-centric formulation enables training on a large-scale synthetic dataset comprising 46.8k grasping poses and 60k free-hand samples, providing broad coverage of hand geometry while directly transferring to real captured hand geometry. Across five real-world benchmarks, SilHand improves relative reconstruction accuracy from 49.2% to 63.4% using masks rendered from reference hands. Moreover, reconstruction accuracy improves consistently as the synthetic training set grows, indicating that broader geometric coverage translates into better real-world performance. While achieving such diversity is difficult with real captured data, it becomes practical under our mask-centric formulation. Finally, SilHand remains robust under mask corruption, with reconstruction error increasing by at most 1.63 mm. Together, these results support binary masks as an effective interface for learning hand reconstruction from synthetic geometry.

Robots handle much more than rigid objects: towels, cables, water, sand. The simulators used to train robots at scale are strongest on rigid bodies, and adding another material usually means a second engine and hand-written glue between the two. We present Alloy, an open-source physics engine for robot learning in which rigid robots, cloth, cables, liquids, granular material and soft bodies share one scene. Each object is described once; Alloy assigns it to a solver suited to its material, couples the solvers inside every control step, and runs the scene in many parallel worlds on one GPU, or on a CPU. Every world carries a health grade, to help tell a failure of the physics from a failure of the policy. Alloy is built on Newton, MuJoCo Warp and Warp, and it is written to be driven by coding agents: its documentation ships inside the package, every error names its fix, and the engine is exposed as tools over the Model Context Protocol. In a cold test, agents that had never seen Alloy installed a pre-release from the package index and trained policies for walking, cable handling and pouring on a single GPU. We also report how throughput grows with the number of worlds, depth-camera throughput, agreement with closed-form references, and a reliability test over a thousand worlds.

Achieving strong optimization generalization across diverse optimization problems while requiring limited training resources remains a challenging problem for optimization-oriented large language models (LLMs). Existing learning-based methods, however, typically require large-scale supervised data, expensive intermediate step verification, and extensive computational resources, while exhibiting limited optimization generalization across diverse optimization structures. In this paper, we propose MiniOpt, a lightweight optimization-specialized LLM trained through a reasoning-to-model-and-solve paradigm and a two-stage training framework with reinforcement learning with verifiable rewards (RLVR). MiniOpt encourages structure-oriented optimization learning by reinforcing the completeness of optimization formulations and optimization outcomes. Extensive experiments on eight optimization benchmarks demonstrate that MiniOpt consistently outperforms existing learning-based optimization models while requiring substantially fewer parameters and training data. To further understand optimization generalization, we introduce a complementary parameter-decoupled evaluation protocol that disentangles optimization structure understanding from numerical parameter processing. The results reveal that MiniOpt preserves optimization capability under parameter-decoupled settings significantly better than previous optimization-specialized models, suggesting that its performance gains primarily originate from stronger optimization structure understanding. These findings demonstrate an effective pathway toward efficient and generalizable optimization-specialized LLMs. The code is available at https://github.com/Hsiang-1/MiniOpt.

Document understanding systems are typically evaluated as separate components via layout analysis, optical text recognition, reading-order prediction, and image captioning. This separation leaves scientific figures, captions, and visual semantics weakly connected, and fails to measure whether a model can produce a single page-level parse that is both structurally faithful and semantically accessible. We introduce Read-Parse-Describe, a unified dataset for document parsing with visual element description generation. Given a page image, models are required to output an ordered list of layout elements containing bounding boxes, semantic labels, reading-order indices, recognized textual content, and alternative text descriptions for visual elements. The benchmark is built from 2,402 documents, comprising 18,163 continuous pages, 125,218 layout elements, and 11,338 visual instances. Besides, we further propose a plug-and-play space-awareness enhancer that reuses multi-layer visual features and injects grid-aware spatial guidance into vision-language tokens for layout-aware generation. Experiments across end-to-end VLMs and pipeline parsers show that existing systems still struggle to jointly recover reading order and visual semantics, while our enhanced model achieves strong joint performance. These results highlight the need to move beyond isolated document parsing modules toward unified models that jointly read, parse, and describe complex documents. The new dataset and code will be made publicly available at https://github.com/yuheng98chen/RPD-Unidoc

A quantized model is usually judged by comparing it with its BF16 original, but that comparison changes more than the weights. It often changes the serving engine, it always runs on a different server instance, it may change how requests are batched, and on some stacks the library selects a different kernel depending only on which packages import. We measure each of these sources on the same 2,400 public decisions for two of Cloudflare's decision models, Clef-flash (9B) and Clef (27B), on an NVIDIA GB10, and state every quantization effect as the excess over them. Sources that are not quantization move between 0 and 0.92% of decisions. llama.cpp restarts are bit-identical on every record for both models. vLLM restarts flip 0, 5 and 20 decisions at BF16, FP8 and NVFP4 on Clef-flash and 4, 0 and 22 on Clef; a third instance of Clef's quantized builds (exploratory) leaves FP8 at 0 and puts NVFP4 at 4 to 22, depending on which two instances are compared. An engine change at BF16 flips 6 to 16, a batch-size change 8 to 15, and the Gated DeltaNet kernel transformers selects unasked 10 to 14. Quantization inside one engine moves more: for Clef-flash and Clef, 0.88% and 1.04% at Q8_0, 2.12% and 1.62% at FP8, 5.79% and 3.50% at Q4_K_M, and 5.79% and 5.67% at NVFP4. On both models, every 4-bit interval excludes every non-quantization source. Two of our five pre-registered predictions failed, and the same two fail when the tests are rerun on the 27B: FP8 is not inside restart noise, and changing engine moves fewer decisions than 8-bit quantization does. Accuracy hides most of this. Over the full 31,275 public questions, Clef's Q4_K_M build moves accuracy by a net six questions (-0.02 points [-0.20, +0.16]) while 436 of its decisions get worse (exploratory), and NVFP4 costs Clef-flash 0.95 points and Clef 0.65. We release the code, quantization recipes, per-record outputs and the pre-registration.

Most action tokenizers for autoregressive vision language action (VLA) models compress motion into chunks that must be decoded in full before their actions become executable. We introduce ActionCodec2, a non-neural streaming tokenizer that releases actions incrementally while keeping trajectory reconstruction error within a prescribed tolerance. Its encoder quantizes each step relative to the decoder's reconstructed state, so the selected motion primitive corrects residual error rather than allowing errors from independently quantized increments to accumulate. Finding the smallest such primitive set is a geometric covering problem; we derive the exact minimum for translation and constant-factor-tight bounds for rotation. Set-BPE then losslessly merges co-occurring primitive symbols across coordinates and neighboring steps, without a fixed coordinate ordering. A universal version fitted on 20k hours of data from 24 embodiments averages 8.8 tokens per second for 15 Hz end-effector motion on this corpus. With the same VLA backbone, ActionCodec2 outperforms the compared tokenizers on three simulation benchmarks and in zero-shot real-robot evaluation. Its streaming representation allows policies to incorporate state feedback within an action chunk, substantially improving performance over open-loop execution under randomized dynamics, external disturbances, and moving objects. The same tokens also can serve as a direct control interface for VLM agents, which compose successful action sequences into reusable macros. These agent-built macros reduce control decisions and thinking-token usage across attempts and transfer to a held-out task. More results are available at https://zibindong.github.io/actioncodec2-page/

Language moved from many task-specific models to one unified model: a single generative model with a single generative training task that handles all of them within the same backbone, and keeps improving with data, parameters and compute. We believe the bitter lesson applies just as much to computer vision, 3D and spatial understanding, in the end to visual intelligence: the effort should go into general architectures that solve all these tasks in a unified way, and into scaling them, rather than into task-specific designs. We test this on monocular geometry, a task that has belonged to specialised architectures for years. Iris-Z is a general image generator that produces appearance and geometry together. A single backbone denoises, reasons about and understands the image and its depth at the same time, and one set of weights generates images with their depth, estimates depth from photographs and paints images onto a given depth. At enough scale it surpasses state-of-the-art specialists on established depth benchmarks while keeping the generative abilities of the backbone it started from without specialized layers. We release two models for the community.

A fully miscible interface must disappear at equilibrium, but its lifetime need not be governed by a bulk instability. We construct a spatially resolved, dissipative two-field model in which a prepared correlation-excited interface undergoes a localized saddle-node while its homogeneous excited surroundings remain metastable and the relaxed mixture is strictly one phase. The interfacial soft mode produces a deterministic first-passage time that grows as the inverse square root of the distance from the fold and couples directly to the evolving concentration-profile width. The quartic structural potential admits an exact reduction from competing positive harmonic energies on a constrained configuration manifold; sixth- and eighth-order examples provide additional numerical robustness tests. The result is an interface-specific kinetic singularity without equilibrium bulk criticality, subject to explicitly stated microscopic and continuum limits.

Comparing soccer teams that rarely or never meet is a fundamental challenge in sports analytics: clubs play mostly within their national leagues, and national teams within their confederations. We propose a rating framework that places clubs and national teams worldwide on a common probabilistic scale. Building on a recent extension of Glicko with order effects and strength-dependent draw probabilities, we adapt the model to soccer with goal-scoring adjustments, a decomposition of home advantage, extra time, league-aware rating initialization and rules of promotion and relegation that do not inflate the levels of leagues; national teams start from the levels of their confederations. The models are evaluated on about 1.2 million club matches of 252 leagues from 166 countries (2000-2026) and 12.6 thousand national team matches (1998-2026). On held-out seasons, the modified Glicko model is clearly more accurate than Elo and as accurate as score-matrix models built on the goal expectations of a gradient boosting model, including the Dixon-Coles and Markov chain models, while it is substantially more accurate in international club matches. Averaging the predictions of the rating model and the Markov chain model is more accurate than either model alone. For national teams, the model is significantly more accurate than all baselines, especially in the final tournaments, where teams of different confederations meet. The resulting ratings are interpretable and can be used for player evaluation, transfer analysis and match forecasting. The match data, the ratings and the code are publicly available at https://github.com/andreyshelopugin/SportsRatings