Papers

Naturally irregular time series combine asynchronous observations, missing values, unequal lengths, and nonuniform sampling, while dense adapters can discard temporal structure. We propose a mask-aware state space classifier for irregular time series (MASCIT), which supplies observation masks to the encoder and excludes invalid steps from gated temporal aggregation. Across 34 irregular time series datasets, MASCIT yielded the strongest aggregate point estimate and was the only evaluated neural model with three-seed results on every dataset. MASCIT retained the lowest point rank across six overlapping irregularity indicators, while factorial ablations favored partial over full selectivity. These results support selective state space models as effective, executable backbones for naturally irregular time series classification.

Leveraging the strong generative capabilities and stable training dynamics of diffusion models, diffusion-based recommender systems (RSs) have recently emerged as a novel recommendation paradigm, attracting increasing attention from both academia and industry. However, despite the rapid growth of diffusion-based RSs, a critical issue has emerged: the lack of a unified and systematic quantitative evaluation benchmark, which often results in irreproducible experimental results and unfair comparisons across studies due to inconsistent data processing, training configurations, inference procedures, and evaluation protocols. To address this challenge, we propose Eval4DiRec, the first unified and open-source evaluation framework specifically designed for diffusion-based RSs. Eval4DiRec supports 14 representative diffusion-based RS models across five different recommendation scenarios, providing consistent and reproducible experimental settings to systematically assess their performance. Built upon this framework, we conduct extensive empirical studies to benchmark these models under unified protocols. The results highlight the strong potential of diffusion models for recommendation while also revealing key factors and practical challenges that substantially affect their performance, thereby establishing a solid foundation to facilitate fair evaluation and guide future research in this promising field. Our code and data are available at: https://github.com/wangcong2001/Eval4DiRec.

Agent skill evolution seeks to improve reusable procedural guidance for large language model (LLM) agents through iterative revision. Existing methods base each revision mainly on execution trajectories or feedback, leaving recurring behavioral requirements across tasks implicit and tying revision to the behavior of the current skill. We introduce SkillFocus, which decomposes recurring task requirements into a capability space that remains fixed as the skill evolves, separating what tasks require from how the current skill behaves. SkillFocus maps current task outcomes to this space to identify the capability that leaves the most tasks unresolved, then uses that capability to determine what to revise and which evidence to use. Across four benchmarks spanning heterogeneous tasks, SkillFocus achieves the best held-out accuracy on all four, outperforming the strongest competing result by 5.7 points on average while using 24% fewer evolution tokens on average than the closest iterative baseline. Controlled studies further show that capabilities derived from recurring task requirements outperform task-semantic and execution-derived alternatives, while randomizing task–capability assignments reduces final accuracy by up to 20.2 points. Matching evidence to the selected capability increases candidate gain by 4.4 points under prioritized revision.

Language model agents serving organizations must coordinate requests from multiple users while using knowledge distributed across their interactions. We identify two complementary capabilities for this setting, namely cross-user interaction and decision-making, as well as cross-user memory and knowledge use. Both capabilities are governed by organizational constraints across three aspects: user identity, authority, and access permissions; the attribution and temporal validity of information; and rules for resolving conflicting requirements across users and completion requirements for joint decisions. These constraints shape what information or decisions must be obtained before an action can proceed and what conditions must be satisfied during its execution. Motivated by this, we introduce Org-Agent, a unified constraint-centric reasoning framework that organizes task execution in three stages. Specifically, Org-Agent decomposes a task into atomic subtasks and constructs a task dependency graph whose edges encode the dependencies among them. Building on this graph, it schedules the subtasks in dependency order through topological sorting. It then executes each subtask while accounting for the task's constraints, supported by evidence-acquisition and memory-management tools. Experiments on MUSES-Bench and GroupMemBench demonstrate the effectiveness of Org-Agent on both capabilities, and ablations further support the contributions of dependency modeling and tool use.

AIVC (AI Virtual Cell) is a learned simulator of cellular behavior across conditions. Predicting how a cell population responds transcriptionally to a genetic perturbation is a core task. Perturb-seq records that response by destructive sequencing, so a control cell and a perturbed cell are never observed as a pair, and cells under one condition remain heterogeneous and noisy. Regression on individual cells absorbs sampling variation into the estimated effect, whereas interpretation requires the reproducible population effect. P2P (Perturbation-to-Perturbation) takes a stochastic cell-set view as its supervision unit. Two views drawn from the same condition share a reproducible population effect and differ by view-specific variation. A permutation-invariant set encoder summarizes the control population, a structured encoder represents perturbation tokens, cellular context, dose, and combination interactions, and a gate blends empirical condition-effect memory with a neural residual. A heteroscedastic head predicts the population mean and gene-wise response variance. Under one protocol and five seeds, P2P attains the lowest expression RMSE and the highest Effect Pearson, DEG F1, and DEG average precision on each of Adamson, Norman, Replogle K562, and Replogle RPE1 relative to GenePert, LinearPert, SLIM, Scouter, and scPILOT. On Replogle K562, Effect Pearson rises from 0.643 to 0.702 and DEG F1 rises from 0.067 to 0.178 relative to Scouter, the strongest baseline on both metrics.

Memory is critical for AI agents. Many existing agent-memory systems follow an Ahead-of-Time (AOT) design, constructing memory before a specific request arrives. While this reduces online serving cost, such request-agnostic memory construction can discard fine-grained information that later becomes important. To address this limitation, we propose Just-In-Time Agent Memory (JAM), a trainable framework for query-conditioned context construction at runtime. A Memorizer preserves complete raw histories in a hierarchical page-store with compact navigational summaries, while a Researcher iteratively retrieves, inspects, and integrates evidence for each request. To train these memory-use behaviors, we introduce Memory-Gym, an evidence-grounded data synthesis pipeline covering nine task types across six domains, and optimize the Researcher through verified-trajectory supervised fine-tuning followed by Hint-guided Group Relative Policy Optimization. We demonstrate the effectiveness of JAM across a variety of benchmarks on agent memory and long-context processing, where it achieves stronger task performance than AOT-style memory systems while remaining substantially more efficient than prior trained agentic memory approaches. To support reproducibility and future research, we release our anonymized source code at https://github.com/VectorSpaceLab/general-agentic-memory.

Existing LLM serving systems virtualize and optimize KV-cache memory, but treat model-weight memory as fixed throughout execution. Recent work on multi-precision model representations challenges this design by allowing a single stored model to support both full-accuracy and lower-precision execution, making the effective weight footprint runtime-dependent. This creates an opportunity under bursty workloads, where temporary spikes in KV-cache demand often determine throughput and SLO compliance. We present DPS, a dual-precision LLM serving system that turns weight memory into an elastic resource: under normal load, DPS serves the full-accuracy model; under KV pressure, it switches to a nested, lower-precision variant and repurposes unused weight memory for KV cache blocks. DPS is built on Semi-Unified Memory (SUM), which partitions the weight region into a persistent lower-precision sub-region and a shared region that alternates between residual weight tensors and KV-cache blocks, preserving compatibility with paged KV-cache management. We implement DPS on top of vLLM and evaluate it across both dense and MoE models and various production workload traces. Our results show that \sysname improves sustained throughput by $2.1$–$3.3\times$ and effective pass@1 by up to $+41$\,pp over Static FP16, while preserving FP16-class accuracy.

Tool-using agents can initiate consequential infrastructure changes, yet evidence required for admission may expire while other checks run or depend on a shared fault domain. We formulate evidence acquisition as joint witness selection and scheduling under quorum, diversity, freshness, deadline, and resource constraints. Assurance-Aware Semantic Scheduling (AAS) combines integer-program selection, dispatch-aware temporal scheduling, bounded diagnostic expansion, and receipt-aware repair. Formal results state the assumptions needed for dispatch-time freshness and finite diagnostic expansion. In three generated infrastructure workloads, AAS produces 1,075/1,200 valid candidates versus 647/1,200 for constraint-aware forward scheduling; stale candidates fall from 440 to 12. Paired sensitivity studies reuse the same instances and operation latency draws across parameter settings. A corrected timeout intervention finds 18/20 admissions with repair or full resynthesis versus 0/20 for a static plan, with lower committed cost when receipts are reused. On 20 constructed cases requiring a certified decomposition cut, refinement recovers an oracle-matching feasible plan every time. These are controlled simulation results; the bounded oracle shares a temporal search component, and transfer to deployed systems remains untested.

Compact JEPA world models enable efficient latent-space planning, but low-dimensional representation trained under reward-free self-supervision must encode both action-conditioned dynamics and predictable visual context. This competition can entangle controllable state with high-rank nuisance appearance and degrade planning as scenes become more complex. We introduce LRC-JEPA, a lightweight end-to-end world model that routes information into a compact predictive latent $\mathbf{z}$ and learned-query residual-context embeddings $\mathbf{u}$. Only $\mathbf{z}$ is propagated by the dynamics model and used for planning, while $\mathbf{u}$ captures temporally persistent information for cross-attention reconstruction; a differentiable residual connection encourages the latent to retain complementary dynamic content. Under explicit assumptions, we show that the resulting representation is sufficient, minimal, nuisance-invariant, and disentangled. Across four simulated control environments, LRC-JEPA improves average planning success over a parameter-matched JEPA baseline by 9 percentage points and matches or exceeds substantially larger pretrained models. On the real-world Bridge-v2 set, its 5.5M-parameter active encoder outperforms DINO-WM (22.1M) and V-JEPA2 (303.9M) encoders while also enabling faster planning. Physical-state probes, reconstruction interventions, and ablations confirm the effectiveness of LRC-JEPA's representation disentanglement.

Language model post-training is often bottlenecked by the need for human-collected preference data, which is expensive and difficult to scale. Reinforcement learning from AI feedback (RLAIF) style approaches that leverage pseudo labels offer an abundant alternative but introduce systematic biases that degrade downstream alignment. Recent general-purpose semi-supervised methods correct for teacher bias using a small set of human-labeled examples, but suffer from high variance especially when human annotations are scarce. To this end, we propose ABC-Align, leveraging abundant pseudo label signal to minimize variance and applying a lightweight, adaptive correction grounded in the human-labeled subset. The correction strength is tuned automatically during training using plug-in estimates of the relevant bias–variance quantities. On LLM alignment with RLHF, DPO, and GRPO where human feedback is scarce, we empirically demonstrate that ABC-Align achieves superior performance over prior semi-supervised baselines in a series of experiments on an increasing scale. Our code is available at https://github.com/SewoongLab/abc-align .

The profile of a role-playing agent usually depends on the pre-defined personality in a system prompt, whereas its memory processing pipeline, including prioritisation of stored memories and subsequent retrieval, remains independent of this personality. This separation causes the agent's memory processing to be inconsistent with the pre-defined personality, and makes it difficult to validate whether agent behaviours follow this personality. In this paper, we propose Personality-Integrated Memory (PersMem), which integrates personality into the agent's memory processing pipeline, making it consistently personality-dependent. PersMem processes memory using four steps, where the personality is mapped to operation-specific parameters controlling: (i) affective appraisal annotating emotion states of the user input; (ii) retention of previously stored memories along with the current input; (iii) passive affect-driven memory retrieval exploring memories similar to user input in semantics and personality-guided emotions; and (iv) active goal-driven memory retrieval that refines and selects passively retrieved memories for the reply. Consequently, consistency with the pre-defined personality can be examined by inspecting memory-processing traces during human-agent interactions. We evaluate these personality-dependent differences in attachment and Big Five settings. PersMem exceeds the chance baseline for four-way attachment classification by 23.1 percentage points. In Big Five dialogue comparisons, PersMem achieves 67.5% accuracy, 6.7 percentage points above a baseline using uniformly sampled memories. On CoSER, PersMem achieves an average score of 66.13, with scores of 69.33 for Character Fidelity and 84.33 for Storyline Quality. Together, these results show that PersMem produces distinguishable personality-related memory-processing patterns.

On-policy distillation (OPD) specializes pretrained video diffusion models through teacher supervision along the student's own generation trajectory. Although large video models are natural teachers, developing specialized video experts can require costly video data and training, while querying them incurs substantially higher latency than querying image experts. More readily available and cheaper to query, image experts offer a cost-effective alternative, particularly for largely temporal-agnostic capabilities such as aesthetics and OCR that admit frame-level supervision. However, heterogeneous image and video latent spaces prevent direct supervision of intermediate student states, while image experts lack cross-frame motion supervision, making temporal consistency vulnerable to frame-level improvements. In this paper, we propose MILD, a Motion-Preserving Image-to-Video Latent Distillation framework that transfers specialized image expertise while preserving pretrained video dynamics. MILD uses a learnable linear connector that aligns student latent states and predicted updates with those of image experts, enabling supervision transfer across heterogeneous latent spaces. We further constrain image-guided corrections around the pretrained student's predictions to preserve video dynamics and incorporate an optical-flow-based motion reward to improve motion quality and temporal consistency. Across specialized image experts and multiple video-student backbones, our method consistently outperforms video-teacher OPD baselines, with further studies demonstrating effective transfer across connector designs and heterogeneous architectures. These results establish image-to-video distillation as an effective route to improving video generation by drawing on the diverse and evolving capabilities of the image-generation ecosystem.

We describe our system for DocSem, the document-grounded quantitative reasoning shared task at DocInsights 2026, and analyze why it succeeded on labeled data and failed on the test set. The pipeline pairs hybrid block retrieval with Program-of-Thoughts (PoT) generation executed in a sandboxed interpreter, self-consistency sampling, and entity enrichment from chunk-level knowledge graphs. On our held-out split, application architecture moved the metrics far more than model scale did: PoT added 0.282 joint accuracy to a compact 7B model but at most 0.005 to a 72B model, and a 27B model with the full harness matched the 72B (0.884 vs.\ 0.873) at roughly 2.7$\times$ fewer parameters and a quarter of the CO$_2$. We read this through a distinction between world knowledge, which scales steeply with parameters, and language knowledge, which scales gently, and show that structured-output training makes a compact model harness-ready rather than merely small. On the raster, watermarked test PDFs the same system collapsed to 13.58% joint (rank 149 of 163); a controlled re-rendering of the validation set reproduces the OCR half of the collapse while bounding what the simulation misses. Auditing the physical nature of evaluation inputs precedes architecture, and the leaderboard's bimodality is consistent with reading quality, not reasoning, having separated the field.

Recent omni-modal models demonstrate strong perception of audio and visual inputs, yet often struggle to connect what they hear with what they see at the same moment. This weakness in temporal correspondence can cause models to associate spoken cues with the wrong visual scenes, producing plausible answers grounded in incorrect audio-visual pairings. We diagnose this problem through controlled temporal swaps, revealing that model answers do not reliably follow changes in these pairings. To address it, we propose Synchrony-Guided Representation Alignment (SyncRA), a lightweight method for strengthening temporal correspondence between audio and vision. Specifically, SyncRA contrasts intermediate audio-visual representations within each video, aligning matching moments while separating mismatched ones to capture local temporal correspondence within a shared global context. The objective derives supervision directly from existing input timing, requiring no additional annotations and leaving inference unchanged. We evaluate SyncRA across four open omni-modal models spanning different sizes and architectures on five public video benchmarks. SyncRA consistently outperforms answer-only fine-tuning across all model-benchmark combinations, while substantially improving the ability to track changing audio-visual pairings in controlled evaluations. These results demonstrate that lightweight, targeted supervision can effectively strengthen temporal correspondence and translate into broad improvements in audio-visual question answering.

Split Federated Learning (SFL) enables resource-constrained clients to participate in collaborative training, but vanilla SFL exchanges smashed data and gradients at every batch, which incurs significant communication overhead. Recent methods reduce this overhead with an auxiliary network at the client-side cut layer. However, we identify that this approach makes the client optimize a local objective that differs from the end-to-end objective, which fundamentally limits the collaborative training between the client and the server. We propose Compensated Feedback based SFL (CoeF-SFL), a communication-efficient framework that retains the end-to-end objective without any auxiliary network. In CoeF-SFL, the client and the server exchange the smashed data and the gradients once per round and reuse them during local training. Since this reuse makes the gradients stale on the client side, we compensate them with a curvature-based correction in the activation space and develop two variants. CoeF-D approximates the Hessian with a diagonal gradient outer product, while CoeF-J exploits the tractable Jacobian-based Hessian of a surrogate loss that upper-bounds the true loss. We provide the theoretical background of each method, characterizing its compensation. Across vision and language tasks, model capacities, cut layers, and data distributions, CoeF-SFL significantly outperforms auxiliary-network-based methods under the same communication frequency, and the improvement is most substantial on vision tasks. Code is available at https://anonymous.4open.science/r/CoeF-SFL-2686/README.md

Large language models receive limited explicit training in reasoning about runtime program states. We study whether training models to reason about runtime program states improves downstream software-engineering capabilities. We introduce two complementary program-state reasoning tasks. Buggy input-output reasoning requires a model to generate a concrete input that exposes a behavioral difference between a buggy program and a hidden correct implementation and to predict the resulting execution behavior. Precondition-postcondition reasoning requires an agent to symbolically characterize a bug-triggering precondition, predict the expected postcondition, explain their causal connection, and instantiate this reasoning as an executable regression test. By incorporating these two tasks into a staged post-training pipeline, we develop Comet-9B, a 9B language model based on Qwen3.5-9B Base. We evaluate the resulting checkpoints on repository-level patch generation, regression-test generation, and security PoC generation. Adding both program-state reasoning tasks to supervised fine-tuning (SFT) on issue resolution improves success rates by 7.25 percentage points on SWE-bench Pro and 9.70 points on SWT-Bench Verified. Sequential reinforcement learning on the two tasks yields further gains of 7.25, 26.79, and 4.67 percentage points on SWE-bench Pro, SWT-Bench Verified, and CyberGym, respectively. Despite having only 9B parameters, Comet-9B achieves a score comparable to the reported GPT-5.2 result on SWE-bench Pro and matches the reported success rate of a GPT-4o-based agent on SWT-Bench Verified.

Physical interactions can create future hazards that are not apparent from the robot's current geometric surroundings. We present a framework termed Predictive Semantic Safety (PSS), which connects visual physical reasoning to backup-based safety filtering. A vision-language model (VLM) predicts physical events and their timing or directly predicts object displacements. An explicit motion model converts event hypotheses into object trajectories. Split conformal prediction calibrates position errors jointly across specified objects, observation times, and future times; geometric shape bounds convert the resulting position regions into predicted object occupancy. PSS evaluates a prescribed backup maneuver against this occupancy and derives input-affine constraints for minimally modifying the nominal input while preserving backup feasibility under the robot dynamics and input limits. MuJoCo experiments with a Unitree Go1 consider falling fixtures, impact-driven support loss, and contact propagation. PSS achieves a safe episode rate of 99.3%, compared with 43.3% for a Backup Control Barrier Function baseline that only uses current obstacle geometry.

Vehicle-to-Everything (V2X) cooperative perception improves 3-D detection by sharing intermediate features, but dense remote features may repeat context that the ego agent can infer locally. Most communication-efficient designs optimize masks or codes empirically, leaving a more basic question open: which remote evidence is indispensable given the receiver's own observation? We introduce a closure-fidelity perspective on ego conditioned remote perception. Under a finite deductive abstraction and explicit conditions, its rate–distortion function decomposes over an irredundant core, and the exact zero-distortion rate becomes $P_A H(\pi_A)$. This analysis suggests a concrete design principle: transmit compact evidence and recover derivable context with bounded receiver-side inference. Guided by this principle, SemRD-V2X is an operational neural proxy that combines exact-budget BEV support selection, pointwise channel compression, and masked shared-weight reconstruction before standard fusion. Experiments on simulated V2XSet and real-world DAIR-V2X validate the resulting design. In a controlled five-run V2XSet comparison against a locally reproduced V2X-ViT-v1 baseline on one Tesla V100, SemRD-V2X reduces the analytical feature payload by $26.6\times$ while improving AP@0.5/AP@0.7 by 4.13/8.57 points, with 3.81% additional mean compute latency. These results position closure fidelity as both an analytical lens and an actionable design principle for communication-efficient cooperative perception.

Synthetic tabular data generation provides an effective alternative for the training of machine learning models when real-world data is limited or inaccessible. However, the heterogeneous, non-smooth, and incomplete nature of tabular data poses fundamental challenges to conventional probabilistic and deep generative models, where their interpretability remains limited. This paper proposes a novel fuzzy distribution modeling methodology for synthetic tabular data generation based on fuzzy sets theory. Feature distributions are represented using fuzzy sets and feature dependencies are modeled through Fuzzy Cognitive Maps, resulting in a low-parameter, and an interpretable data representation. Synthetic samples are generated by sampling fuzzy concepts rather than raw values, enabling native support for mixed data types, missing values, and domain constraints. The methodology further supports linguistic queries and IF-THEN reasoning, facilitating transparent simulation of decision-making processes. Experimental results on benchmark datasets demonstrate competitive performance with respect to utility, fidelity and privacy compared to state-of-the-art methods, while offering substantially improved interpretability. These results establish fuzzy distribution modeling as a principled and effective approach for synthetic tabular data generation in fuzzy systems and decision support applications.

Gradient updates on different data generally do not commute: training a language model on two data sources in opposite orders gives different weights, even with the same data and total exposure. Loss or benchmark deltas show that the models differ, not where. We ask whether this path dependence leaves a parametric training-history memory: a weight component that flips sign when the two sources are swapped, is localized in output space, changes the held-out loss gap between the two orders under targeted interventions, and reveals which trained model came from which order. For one small SGD step of size $\eta$ on each of sources $A$ and $B$, the weight difference $\theta_{AB}-\theta_{BA}$ is, to leading order, $\eta^2 b_{AB}$, where $b_{AB}=H_Bg_A-H_Ag_B$ is the Lie bracket of the two gradient fields at the base model. We define commutator memory by projecting the bracket through the logits into one score per vocabulary token; the scores sum to the bracket's prediction of the gap. The scores are localized: on three models, the same readout of the measured $\theta_{AB}-\theta_{BA}$, or of a bracket from disjoint batches, shares 82-99% of the original top-20 tokens, versus 35-49% for norm-matched random directions. They are causally actionable: in Qwen-3-4B SFT, downweighting the ten tokens with the largest predicted share of the gap closes a median 32% of the measured gap, while frequency-matched tokens with near-zero scores have almost no effect. The weights themselves carry the component: projecting the difference between the two trained models onto $b_{AB}$ identifies which came from which order in 92% of cases across four LLMs (chance 50%). Controlled tests also cover matched-batch DPO, a frozen-rollout GRPO-style objective, and an AdamW endpoint check. The memory is defined per source pair, not per example, and its projection on $b_{AB}$ decays with further training.

Spatial audio large language models (LLMs) enable embodied agents, wearable assistants, and immersive systems to recognize sound events, localize sources, and reason about their spatial relationships. However, existing spatial audio LLMs often rely on early fusion of acoustic and spatial features and source-agnostic token representations. These designs make it difficult to preserve the correspondence between individual sound events and their spatial attributes, particularly in multi-source scenes. To address this limitation, we propose SAIL, a Spatial Audio Intelligence framework with LLMs that preserves acoustic-spatial structure and source-level correspondence from audio encoding to LLM alignment. SAIL introduces a Disentangled Spatial Audio Transformer that represents Mel-spectrogram and interaural phase difference features as separate acoustic and spatial streams. Source-discriminative task queries further learn event, direction, and distance information for each source. A Dual-Stream Q-Former then aligns the two streams with the LLM using acoustic and spatial queries organized by source slots. Compared with the early-fusion baseline, SAIL achieves consistent improvements in dual-source sound event detection, direction and distance estimation, and spatial reasoning. These results demonstrate the importance of structured, source-discriminative audio representations for multi-source spatial understanding and reasoning.

Reinforcement learning (RL) has become a key paradigm for enhancing the reasoning of large language models, yet the high dimensionality of parameter updates makes its training dynamics hard to analyze. We study reinforcement learning with verifiable rewards (RLVR) and use vector steering to identify a low-dimensional effective manifold in activation space associated with RL-induced gains. We uncover two geometric properties. (1) Effective Manifold Capacity: the capacity needed to reproduce RL gains can be very small but is not infinitely compressible; at extremely low capacity, intervention dimensionality and input-dependent expressiveness become key constraints, and this requirement varies with injection depth. (2) Control Manifold Separation: effective control directions lie mainly in the low-variance complement of the activation principal subspace. Within a task and base model, the learned geometry stays largely consistent across training configurations, and across tasks geometric alignment correlates with capability transfer. Experiments on 5 LLMs and 6 verifiable-reward tasks support these findings. We then propose Alpha-Stabler, a plug-and-play framework with a Predictor that monitors principal-subspace intrusion for early collapse warnings, and a Controller that removes the principal-subspace component of activation gradients during backpropagation while preserving the orthogonal complement. Alpha-Stabler stabilizes training for 2,000 steps and consistently improves RL gains, offering practical insights for robust post-training. Code: https://github.com/caiyuchen-ustc/On_Policy_Vector_Training

A strong LLM strategic agent should reason prospectively over uncertain futures, adapt its strategy to opponents' behavioral tendencies, and continuously recalibrate its decision process from interaction experience. However, incorporating these sources in free-form reasoning could lead to unsupported strategic assumptions, inconsistent opponent estimates, and harmful interference from irrelevant historical interactions. To address these issues, we propose SAGE, a training-free inference-time framework that structures LLM strategic reasoning around three coordinated operations: anchor, adapt, and recalibrate. SAGE first anchors reasoning to an equilibrium policy that provides a strategically valid prior. It then conditions deviations from this anchor on a soft belief over opponent behavioral tendencies, enabling opponent-specific exploitation. Finally, SAGE distills strategically related interactions into counterfactual hypotheses about previously missing considerations, allowing past experience to recalibrate the model's reasoning. We evaluate SAGE on three repeated imperfect-information games: Leduc Hold'em, Liar's Dice, and Goofspiel, against various opponent types in each game. Compared with reasoning-intensive LLM agents, including Suspicion-Agent, ReTA, Agent-Pro, EMO, and Hypothetical Minds, SAGE achieves up to a 127.6% payoff improvement in Liar's Dice while reducing input and output token usage by up to 80% and 90%, respectively. In direct match-up play, it attains non-negative mean payoff against 5/10, 8/10, and 8/10 evaluated opponents in Leduc Hold'em, Liar's Dice, and Goofspiel, respectively, while using relatively fewer tokens. Code is available at https://github.com/chenzhwsysu57/SAGE.

Achieving high-quality, cinematic results in text-to-video generation remains challenging for non-experts, whose prompts often lack professional narrative and creative design. We propose SkillPE, a prompt engineering (PE) framework that evolves reusable cinematic skills from expert-authored seeds. SkillPE represents shot logic, composition, lighting, sound design, and other filmmaking cues in a fine-grained format, and retrieves movie references categorized as resonators (good matches), dissonants (weak matches), and divergents (creatively useful near-misses). The first two refine when and how a skill should be applied, while divergents inspire alternative cinematic realizations at different degrees of modification while preserving the user intent. Candidate skills are assessed through generated videos along prompt fidelity, cinematic quality, narrative appeal, and creativity to construct the final skill libraries. Experiments on StoryEval and VBench show improvements of up to 1.40 points over the strongest external baseline and 0.51 points over seed skills on 7-point four-dimensional evaluation, while remaining competitive on benchmark-native metrics. Overall, SkillPE offers a practical approach to balancing fidelity and creativity in cinematic text-to-video generation. Code is available at https://github.com/Ais0n/SkillPE .

Scaling test-time computation is a powerful way to improve language-model reasoning, and is particularly appealing for small reasoning models (sRMs) that are cheap to serve. However, is additional thinking always the right operation? By intervening at intermediate reasoning states across two model families and multiple scales, we find that self-refinement largely consolidates probability mass onto solutions already reachable from the current state, rather than making new ones reachable. These interventions reveal two failure regimes: execution bottlenecks, where the correct path is reachable and reflection can recover it, and knowledge bottlenecks, where relevant external information makes it reachable. Motivated by this distinction, we introduce FlyBy, a selective querying framework, and train 4B and 8B variants to reason first, diagnose what remains unresolved, and, at a knowledge bottleneck, query stronger models whose parametric knowledge extends beyond its own. Supervised fine-tuning bootstraps a multi-depth query action, and cost-aware reinforcement learning calibrates whether to query, what to ask, and how much to spend. On 1,158 hard problems across six benchmarks, FlyBy-4B achieves 45.96% pass@8, surpassing Qwen3-14B (41.64%) at 2.7 times lower serving cost, while also exceeding Qwen3-8B in pass@1 (16.85% vs. 15.31%). Scaling to FlyBy-8B further improves pass@8 to 51.81%.

Verifying LLM-generated answers under a shared computational budget requires jointly deciding which candidates to inspect and which verification attributes to evaluate. We formulate this problem as multi-attribute good-arm identification under a global budget: each candidate is an arm evaluated along several costly attributes, and the goal is to certify as many candidates as possible whose mean scores exceed the prescribed thresholds on all attributes. We propose BMA-GAI, an algorithm that combines cost-aware arm selection with adaptive sampling of attributes. Every observation serves both to guide adaptive allocation and to support anytime-valid certification, which removes the need for a separate confirmation stage. We establish an asymptotic coverage guarantee for BMA-GAI and derive a matching information-theoretic converse that characterizes the intrinsic complexity of the problem, thereby proving that BMA-GAI is first-order optimal away from critical budget levels. Experiments on synthetic benchmarks and an LLM answer-verification task show that BMA-GAI allocates the verification budget more efficiently and certifies more high-quality candidates than competing methods.

This work studies a central gap in interpreting time-series foundation models (TSFMs): a dynamical property may be accessible in a hidden state even when the forecast fails to respond correctly as that property changes. We formalize these properties as Dynamical Parameters, including trend slope, oscillation frequency, and autoregressive dependence. We compare their representation accessibility, measured by recovery from hidden states, with their forecast response, measured by agreement with the expected forecast change. Across nine frozen TSFMs and thirteen laws, 42 of 63 model-parameter cells achieve accessibility above 0.95, whereas their median reference-aligned response relative to the conditional reference is only 0.46. To explain this gap, causal geometry compares the hidden-state change required to produce the reference response with the change induced by the parameter intervention. Directly modifying the hidden state recovers the reference response, but the parameter intervention often moves the state in a different direction. These results show that accessible parameter information need not be expressed in forecasts when input changes miss the required hidden-state direction.

Video-language models are increasingly used as judges of video understanding, both for evaluating model outputs and for training reward models. Whether their judgments remain reliable when the evidence is buried in day-long videos has yet to be established. Existing benchmarks cannot answer this. Their videos are typically only a few minutes long, many answer pairs can be separated from the transcript alone, and collecting human judgments does not scale to ultra-long videos. We introduce PlaylistEval, an agentic framework that builds video-language judge benchmarks over 100-hour playlist collection without human annotation. It automatically generates questions with paired answers whose differences are controlled by causal degradation, so that every pair demands retrieval across the collection. The resulting benchmark contains 630 pairs across seven domains spanning both static and dynamic knowledge, and on a stratified subset of 152 pairs it agrees with human judgments 93.0% of the time (IAA 0.781). Evaluating 17 omnimodal and multimodal models from eight families reveals that frontier judges reach only 75.4% pairwise accuracy, while open-source judge models perform far behind. We further show that both retrieval and final judgment depend on using multiple modalities, and that judge accuracy degrades as the playlist set grows. We release our pipeline, benchmark, and evaluation code at https://playlisteval.github.io.

How much of a running workflow should a language model agent revise? ControlScope compares continuing generated code, editing the next tool call's data arguments, and replacing the unfinished workflow from the same public execution state. The nested permissions separate available repairs from the actions an agent selects. We evaluate one-time and repeated reviews across filesystem tasks, ALFWorld, and AppWorld. Across two source programs per task and three reasoning-reviewer draws on 20 filesystem tasks, FULL completes 15-16 tasks versus 13 for KEEP; across four fast draws it completes 10-13 versus 13. Fresh student-record confirmation reproduces a batch-read repair. ALFWorld fast panels yield KEEP/ARG/FULL scores of 85/86/87 on 87 tasks across 52 scenes and 134/134/127 on 134 tasks across four scenes; reasoning on the 87-task cohort also yields 85/86/87 with substantial review cost. An AppWorld V1 official-test panel of 585 task instances from 195 scenario templates shows small net differences. Frozen replays expose viable agent-written replacements interrupted by later revision in two failed file-organization runs. An offline source-trajectory midpoint comparison shows later reviews completing an insufficient repair. Five-call protection saves 19.4% of logged model output and loses one success across 20 fresh source runs. An argument-only shortcut shows that the broader sampled policy can overlook a cheaper successful edit available in both operation sets. These outcomes tie repair access to actual choices and subsequent execution.

Drug design couples property evaluation, conditional generation, structure-based design, and local optimization, yet machine learning systems typically address these capabilities with separate task-specific models. We introduce CAGenMol-2, a masked diffusion molecular language model that represents molecules, continuous scalar properties, and 3D protein pockets within a single wrapped sequence. Within this pretrained interface, downstream operations are selected by which sequence regions are observed or masked at inference, allowing one checkpoint to perform property prediction, property- and pocket-conditioned generation, and partial-constraint design without task-specific architectures or backbone fine-tuning. We further propose Adaptive Fragment Optimization (AdaFO), a gradient-free mask-and-refill search that turns the masked decoder into an iterative local molecular optimizer. On CrossDocked2020, AdaFO increases Success Rate from 30.2% to 70.8%, the best reported under this protocol, while largely preserving drug-likeness and diversity. Finally, scaffold-preserving directional editing and CRBN/VHL case studies demonstrate its use in compound design workflows spanning local molecular editing, structure-based prioritization, and downstream simulation-based screening.