Papers - 2026-11-06
吾能观之数千而面色如故
Grounding-driven Visual Reasoning#
GameVerse: A Minute-Scale Gameplay Dataset for Long-Horizon Interactive World Modeling
Recent advances in interactive world models have made it possible to generate controllable, dynamic environments from action inputs. Yet current models struggle to maintain coherence beyond roughly one minute of rollout, a limitation rooted not in model capacity but in the scarcity of long-horizon training data. Existing gameplay video datasets typically consist of clips under ten seconds, with captions that describe whole clips in coarse and undifferentiated text, which is insufficient for controllable long-horizon generation. We address this with GameVerse, a minute-scale gameplay video dataset purpose-built for long-horizon interactive world modeling. GameVerse contains 69.7K clips spanning 1,655 hours and 349.6 million frames across diverse first- and third-person gameplay, including 1,358 hours of continuous two-minute sequences, an order of magnitude longer than prior datasets. Clips pass through a unified pipeline of scene-consistent segmentation, quality filtering, camera pose estimation, and event-based caption annotation. Each clip is paired with a temporally stable camera trajectory that remains coherent across minutes, together with captions that separate static scene context from dynamic foreground events and align to event boundaries rather than fixed time windows. Comprehensive analysis shows that GameVerse surpasses existing public datasets in scale, diversity, and annotation quality. Experiments further demonstrate that training on GameVerse substantially extends the temporal horizon of interactive world models.
3D LLM#
From Pixels to Actions - Towards Native Vision-Language-Action Policies
Vision-language-action models are typically built on a simple premise: perception is first compressed by visual encoders, and action reasoning begins only afterwards. Yet manipulation often hinges on precisely the visual signals most vulnerable to this abstraction, e.g., local geometry, boundaries, depth, and subtle state changes. This raises a fundamental question: should action policies inherit a fixed visual endpoint, or directly access evolving visual computation? We introduce NEO-VLA, a structurally encoder-free vision-language-action framework. Images are tokenized by lightweight patch convolutional layers and evolve jointly with language throughout the backbone, allowing a flow-matching action expert to draw from instruction-conditioned visual representations at different depths. This simple change reveals a consistent pattern. Native architectures outperform encoder-based counterparts under controlled comparisons; full fine-tuning reduces but does not erase the gap, and additional scale alone cannot recover it. More strikingly, layer-wise probing shows that the action-critical signals often peak well before the final visual representation. Across diverse manipulation settings and distribution shifts, NEO-VLA consequently achieves stronger and more robust control. Together, these results suggest that embodied control benefits from access to visual representations across the computational hierarchy, rather than a fixed encoder output.
FrontierPhysics: Evaluating Agents for End-to-End Physics Research
AI agents are playing an increasingly important role in physics research. However, existing physics benchmarks only evaluate fixed questions, isolated sub-tasks, or paper reproduction with given solutions, and rarely check research quality alongside correctness. We introduce FrontierPhysics, an end-to-end benchmark with 56 tasks made by 42 physicists based on their own research, covering nine domains across theoretical, computational, and experimental physics. Each task requires agents to review the literature, develop a research plan, conduct research and experiments, and produce structured outputs and scientific manuscripts. Combining deterministic verifiers with expert-authored rubrics, FrontierPhysics evaluates whether agents' research is correct and publishable. Across 22 models in 26 configurations, no configuration passes more than 46 of its 168 rollouts. Manual review of agent trajectories shows that frontier agents' main weaknesses are insufficiently rigorous reasoning and limited expert intuition, rather than gaps in physics knowledge or execution errors.
What Does Multi-Turn Reinforcement Learning Really Improve in LLM Agents?
While multi-turn agentic reinforcement learning (RL) excels on long-horizon tasks, its performance gains remain highly sensitive to interaction constraints. Evaluating across sampling breadth $K$ and interaction depth $T$ via $Pass@(K,T)$, we show that trajectory-level RL advantages depend on sufficient turn budgets, diminishing significantly when budgets tighten. Besides, agentic RL does not uniformly elevate single-step decision quality; rather, its strength concentrates in error recovery and long-horizon resilience. To mitigate the deployment risks caused by this budget dependence, we provide a preliminary discussion on the benefits and limitations of process supervision signals and introduce BudgetMix RL as an alternative that varies interaction budgets during training. Experiments demonstrate that BudgetMix RL achieves robust performance across diverse execution settings. Project Page: https://rethinking-agentic-rl.github.io/
PyRUA-DataFarm: LLM Agent as a Cost-Efficient Teleoperator for Real-Ready Robot Data
Teleoperation remains the default way to collect robot demonstrations, but every episode costs robot and operator time, and every new task starts from scratch. We ask whether a coding agent can take over the teleoperator's job. In our DataFarm, the agent receives a data request and one calibration episode of the robot in the target scene and rebuilds the task in simulation: it writes the scene, an expert controller, and a success judge. The simulator supplies exact actions, and a video model conditioned on real images renders each rollout as photographic video, generating all camera views together consistently. The output is real-ready data: synchronized camera views labelled with the robot's actions, on which a policy can be trained directly and deployed on the real robot. Independent judges certify the expert and every rendered episode before admission. On five real-robot tasks, policies trained only on farm data match policies trained on as many teleoperated episodes (93 vs. 84 successes in 100 trials), and under an unseen tablecloth and coloured light they keep 87 successes where the teleoperation-trained policies fall to 46. Adapting a task is fully automatic and costs \$21 of API usage on average, paid once per task, plus \$5 for a render contract; every further episode costs \$0.80 on average for simulation, rendering, and review, with no robot or operator. One hour of three-view data for a task costs \$107 on average, these costs included, slightly below a quote of \$117 for teleoperation in China and below \$290 in labour alone in North America.
EmoReward: Self-Evolving Reward Calibration for Emotional Video Generation RL
Group Relative Policy Optimization (GRPO) has advanced generative model post-training in verifiable domains such as code and mathematics, but evaluating video quality, especially emotional expression, often depends on subjective aesthetics and has no uniquely correct target. Existing approaches typically rely on fixed rewards for visual quality, motion quality, and text alignment that poorly capture emotional intent and can invite reward hacking, while RLHF is difficult to scale and subject to individual aesthetic differences that complicate the calibration of facial, body motion, and chromatic cues across narrative contexts. We introduce EmoReward, a new self-evolving training mechanism that distills evidence reliability from reward traces collected during an actual training run to adaptively guide multi-reward weight allocation in subsequent training by determining which emotional evidence to trust and how much to weight it. EmoReward first plans specific grounding cues from the emotional prompt to guide evaluation, then uses heterogeneous signal primitives to score the generated video during the act stage. Reflective Reward Gating calibrates their contributions through instance-level gates for phase, face confidence, conflict, and subtlety, then renormalizes weights over the active set to produce the sample's fused reward. RewardMemory then estimates each primitive's reliability from rollout scores and outcomes within each emotional regime in the finished training run, then fixes bounded multipliers that calibrate fusion weights in the next training run without updating reward models. Experiments on HunyuanVideo 1.5 show improvements across multiple RL algorithms and compare EmoReward with other agentic reward methods under SAGE-GRPO. Ablations and transfer experiments on SD 3.5 and FLUX further support adaptive reward calibration for visual generation.
StreamDuo: A Live Commentator and a Research Agent for Unseen Games
Streaming video LLMs can now commentate live gameplay, but they are trained and tested on the same games. We study cross-game streaming commentary, in which a model trained on one game commentates games it has never seen, and find that current commentators transfer their timing but not their knowledge of the game world. We propose StreamDuo, which pairs a trainable commentator with a frozen research agent that reads the wiki of the game being played. When the commentator sees something it cannot name, it sends the current frames, and the research agent returns a few wiki entries two seconds later while the stream continues. We train only the commentator, on a single game, in two stages. Information-gain labels teach it to ask only where an answer would help it say what the human caster said, and on-policy self-distillation from a copy of itself that sees the ideal answer teaches it to use noisy answers. Trained on Baldur's Gate 3 and tested on ten unseen games, StreamDuo raises entity recall from 0.082 to 0.116 over Proact-VL trained on the same data, cuts source-game leakage from 9.9% to 1.0%, and achieves the best timing and the highest judged liveliness and overall quality among streaming models.
On the Generalization Limitation of On-Policy Distillation for Post-RL Consolidation
Multi-teacher on-policy distillation (MOPD) extends on-policy distillation (OPD) to consolidate capabilities from multiple teachers. Although prior work reports near-perfect in-distribution (IND) distillation performance, we find that this can mask incomplete capability consolidation beyond the supervised distribution. A preliminary study on synthetic reasoning shows that OPD does not fully transfer teachers' generalization capabilities. We then extend this investigation to real-world tasks, evaluating extrapolative and compositional generalization across mathematical reasoning, coding, agent, and long-context capabilities using Qwen3 models. Across extrapolative settings, OPD and MOPD reach 96.04-102.73% of their teachers' IND scores. Yet their out-of-distribution (OOD) scores remain 2.45-16.25% below the base model's. Across three compositional settings, MOPD improves IND performance over the base model, but its performance on composed tasks as OOD remains 18.19-24.34% below that of the best teachers. To understand these gaps, we examine teacher capability and data coverage as two potential bottlenecks. Jointly alleviating these bottlenecks eliminates OOD forgetting in our experiments and suggests the potential for generalization beyond the teacher. (M)OPD also reduces OOD forgetting on average relative to its reinforcement learning (RL) baselines. Analysis of the associated update geometry reveals smaller and more spectrally concentrated updates on average than those of the RL baselines, along with greater shared update structure under MOPD. Motivated by these findings, we propose geometry-based patching and base-preserving replay to improve OOD performance while largely preserving IND performance. Our findings identify a blind spot in capability consolidation via (M)OPD, highlighting the need for algorithms that explicitly support OOD retention and generalization.
TAC-CoT: Tool-Empowered Implicit Chain-of-Thought via Interleaving Explicit Tokens
LLM-based agentic systems are increasingly combined with CoT reasoning to solve tasks that require accurate reasoning, external knowledge, and multi-step deliberation. Yet, explicit CoT incurs substantial generation overhead, while implicit CoT, which replaces textual rationales with compact continuous tokens, does not naturally support the explicit communication interface required for agentic tool-use. We propose TAC-CoT, a tool-empowered adaptive implicit CoT framework that replaces explicit reasoning traces with compact latent reasoning segments while retaining textual tokens only for tool calls, observations, and final answers. This interleaved design enables agentic tool-use without verbalizing every intermediate reasoning step. Experiments on agentic math and QA benchmarks show that TAC-CoT improves accuracy by up to 8.8 points while reducing model forward passes by up to 6.8x. Furthermore, analyses show that TAC-CoT improves tool-use quality in both single- and multi-tool settings.
Multimodal World Model#
Turns Matter: Advancing Streaming Video RL via Turn-Level Credit
Streaming video understanding requires models to repeatedly decide whether to respond or continue observing. Recently, Group Relative Policy Optimization (GRPO) has been applied to this setting, with existing approaches emphasizing reward design while retaining a single trajectory-level advantage across turns. Such shared credit can reinforce redundant replies after earlier successes and discourage useful follow-ups after earlier failures. We propose TurnWise, a turn-level GRPO framework that aligns learning with individual streaming decisions. TurnWise addresses the mismatch between trajectory-level feedback and turn-level interaction, promoting informative responses and reducing redundancy. Specifically, turn-level credit assignment uses silence-completed prefix scoring to construct group-relative advantages from each decision's remaining return, separating its learning signal from earlier rewards. Turn-mean policy optimization averages token losses within each turn and then across turns, giving response and waiting decisions equal weight regardless of output length. These complementary designs improve streaming policy learning without a learned critic or additional rollouts. Extensive experiments show our method outperforms prior RL approaches for streaming video understanding.
Halley: A Predictive Control Plane for Optical AI Data Center Networks
A frontier AI model is trained in an AI data center, on tens of thousands of accelerators joined by a dedicated network, and every training step ends with all of them exchanging data at once. Between racks, that network is optical, and each of its four layers, from lasers to switches, runs a fixed or after-the-fact policy. Lasers sharing an enclosure, cooling, and power can fail together. At a fixed spare ratio, correlated failure provably caps a pool's availability however large it grows. In this paper, we present a comprehensive predictive control plane framework, called Halley, that reads each resource's next state and acts before the failure occurs. It has four components, each carrying a formal guarantee: LightReserve that sizes the laser pool to provide a coverage bound, FoldFit that schedules the switch within $11/9$ of the optimal, ThermaLead that holds a microring with one-step-ahead control, and TriageGuard that protects the highest-risk modules first. Halley gives each layer a predictive controller that defers to the deployed policy when unsure. Composed from simulations at that point, effective goodput, the share of ideal throughput kept, rises from 65.2\% to 76.5\% if a window that exhausts its spares is lost. If only the dark channels' bandwidth is lost, it reaches 67.4\%. Per megawatt, that is up to 1.03--1.17 times the training; for the same training, up to 3.3--14.7\% less data center capacity. \textsf{Halley}'s fallback circuits are formally verified and are shown to achieve close timing on a commercial FPGA, and pass a 130~nm signoff.
Solv3R: 3D Looped Transformer as a Multi-view Solver
Classical structure-from-motion reconstruction proceeds iteratively, refining camera poses and scene geometry from a current estimate. This allows maps to easily expand without recomputing already computed views, or geometry estimates to improve as more observations are provided. Most geometric foundation models, however, predict entire geometry in a single forward pass or update a recurrent state as frames arrive sequentially, making this expansion and refinement non-trivial. We introduce Solv3R, a looped transformer that learns to continue solving multi-view geometry from its current state. Importantly, its update operator is independent of the iteration count, allowing previously refined and newly introduced views to be jointly updated regardless of how ``solved'' they are. We train the operator to preserve accurate reconstructions under continued iteration and correct inaccurate states, enabling refinement of fixed image collections and integration of new observations. Representations of solved views can also be cached once and reused as geometric context across iterations, enabling compact and efficient key-value caching. We analyze the solver dynamics and compare it with the state-of-the-art, demonstrating that a single learned solver supports offline reconstruction, incremental reconstruction, relocalization, and streaming without task-specific training or modules. The code and models will be made publicly available at https://ubc-vision.github.io/solv3r/.