Generative RL 进展

Despite extensive alignment efforts, Large Language Models (LLMs) remain vulnerable to generating unsafe content under adversarial prompting, yet the internal mechanisms by which safety behaviors are implemented remain poorly understood. We study LLM safety from a mechanistic interpretability perspective and characterize a multi-stage *safety circuit* that organizes refusal behavior, consisting of (i) $Harmful Detection Heads$ that respond to harmful inputs, (ii) $Safety Neurons$ that mediate and stabilize safety signals in the residual stream, and (iii) $Refusal Heads$ that translate these signals into safe response generation. Using targeted attention-head and neuron-level interventions, we provide causal evidence consistent with this circuit organization, showing that suppressing upstream Harmful Detection Heads disrupts downstream refusal behavior and that safety neurons mediate this interaction. We validate that this decomposition recurs across multiple LLM architectures and adversarial attack settings, and use simple, architecture-preserving weight scaling as a mechanistic probe to test its functional relevance. Across six LLMs, circuit-guided scaling improves safety rates under attacks by 26.5%, while incurring only a 1.7% accuracy drop across four standard benchmarks. Overall, our results support a circuit-level interpretation of LLM safety and suggest that mechanistic abstractions can reveal stable and transferable patterns underlying aligned behavior.

Infectious disease transmission is shaped by patterns of human interaction, which adapt as epidemic conditions change. Capturing these context-dependent behaviors remains a fundamental challenge for epidemic models. Here, we recast this challenge by using large language models (LLMs) to represent adaptive human behavior within mechanistic epidemic models. We operationalize this idea through Generative Adaptive Behavioral Layer for Epidemics (GABLE), which adapts LLMs to infer behavioral responses to epidemic and policy conditions and translates them into age-structured contact matrices coupled to a mechanistic epidemic model. Applied to COVID-19 in France, GABLE reproduced responses in population mixing and age-specific contact structures that remained epidemiologically informative. In short-term forecasting, LLM-generated contact matrices outperformed mobility-driven matrices derived from real-world mobility data, with the largest gains at longer horizons. GABLE also extends beyond forecasting to prospective policy evaluation by projecting behavioral and epidemic responses to candidate interventions before implementation. When supplied with subsequently implemented policies, GABLE reproduced epidemic trajectories and generated distinct responses to alternative policy timing and composition. By leveraging LLMs as a flexible behavioral layer, GABLE provides a framework for coupling context-sensitive behavioral generation with epidemic dynamics.

Long-horizon story-driven video generation requires a production agent to coordinate narrative decomposition, state tracking, shot design, prompt construction, rendering, and revision across interdependent scenes. Existing adaptive video systems primarily refine requests or reusable skills, leaving recurring production failures disconnected from persistent, stage-targeted improvements across stories. We introduce CineForge, a self-evolving video-production agent framework that couples CineForge-Produce for video generation with CineForge-Evolve for cross-story policy evolution. CineForge-Produce organizes each source story into typed narrative, character, spatial, and cinematic states, uses them to coordinate asset and clip generation, and records the process as a canonical production trajectory. CineForge-Evolve applies Case-to-Pattern-to-Policy Evolution (CPPE) to review trajectory evidence, consolidate recurrent findings into bounded stage-local patches, and deploy validated updates through structural replay and confidence-controlled paired evaluation. To measure complete story realization, we introduce CineScope, which combines a 100-script CineScope-Data suite with a human-aligned, multiscale CineScope-Metric spanning causal state, directorial orchestration, pacing and resource allocation, and character arc. Across CineScope-Data and two public benchmarks, the evolved CineForge policy improves CineScope-Metric from 4.024 to 4.380, outperforms three long-video baselines with consistent gains under ScriptAgent, and reduces review LLM calls by 37.0% on new stories. These results establish production trajectories as actionable experience for video agents that improve cumulatively across long-form storytelling tasks.

Criminal judgment prediction requires models to infer statutory articles, charges, and sentencing outcomes from case facts. Unlike standard classification tasks, it involves a structured reasoning process in which statutes should be matched with facts, charges should be justified by statutes, and sentencing outcomes should remain consistent with charges. Existing approaches optimize final labels, and while some have attempted to evaluate reasoning quality, their evaluations are indirect, often relying on LLM-generated rubrics that reflect model-internal preferences rather than the inherent logical structure of legal adjudication. We propose Juris Policy Optimization (JPO), a post-training framework for structured legal reasoning in Chinese criminal judgment prediction. JPO first uses teacher-generated rationales to supervise a standardized four-step reasoning process, and then applies reinforcement learning with a composite reward over legal prediction quality, reasoning structure completeness, and cross-step consistency. JPO further introduces token-level advantage reweighting and adaptive clipping for legally salient reasoning segments. Experiments on multiple open-source language models and three Chinese legal benchmarks show that JPO consistently improves both judgment prediction and reasoning quality over supervised fine-tuning and reinforcement learning baselines.

Evaluating data-text alignment remains challenging: existing metrics often provide limited explanations for the scores, while prompt-based LLM-as-Judge methods can be expensive and unreliable. We present an end-to-end explainable evaluation metric that fine-tunes a language model to identify omitted, extra, incorrect, and correct data units in a data-text pair. These local judgements are aggregated into precision, recall, and F1 scores, providing both fine-grained diagnostic feedback and an interpretable measure of alignment quality. Across benchmarks, our fine-tuned models outperform LLM-as-Judge methods in error prediction and achieve competitive precision, recall, and F1 scores, while maintaining strong correlation with human judgements. Beyond evaluation, our verifier outputs also provide useful feedback signals for downstream correction and refinement, supporting alignment-oriented improvement of data-to-text and text-to-data. Code and resources are available at https://github.com/guihuzhang/xqdt.

Large language models (LLMs) demonstrate strong Helpfulness, Harmlessness, and Honesty (3H) alignment in English-centric settings, but these gains transfer poorly to low-resource languages due to cultural mismatches. Existing multilingual 3H benchmarks rely predominantly on automated translation or LLM based synthesis, propagating source-language biases while sacrificing local relevance. To address this gap, we introduce Pak3H1, the first human-validated, culturally contextualized Urdu benchmark suite for 3H alignment, comprising PakAlpaca (helpfulness), PakBeaverTails (harmlessness), and PakTruthfulQA (honesty). Our multi-stage pipeline integrates manual cultural adaptation and dictionary-guided post editing to prioritize native speaker judgment, ensuring both semantic fidelity and contextual authenticity. Zero-shot evaluations across multiple open and proprietary LLM architectures reveal systematic cross-lingual alignment gaps: helpfulness win rates decline under localized contexts, harmlessness guardrails break down against regional safety risks, and composite honesty metrics degrade substantially due to localized factual constraints. These findings expose structural limitations in current alignment approaches, underscoring the necessity of human-guided localization for equitable multilingual evaluation.

The same Persona behavior can be beneficial in one context but harmful in another, causing static Persona elicitation to perform inconsistently across tasks. We introduce the Persona Selection--Realization Framework, which models behavior generation through a latent Persona state and decomposes it into Persona Selection and Persona Realization. The discrepancies between static Persona elicitation and an ideal Persona policy in these two components define the Selection Gap and Realization Gap, respectively. Building on this framework, we propose R$^2$A, a two-stage approach for learning Persona policies. Persona Representation Learning uses structured Who--How--What presentations to encode the target Persona's objective, conditional behavioral principles, and trajectory-level manifestations. Persona Runtime Alignment then removes the explicit Persona specification and jointly calibrates behavior selection and trajectory realization using task feedback. Across 12 evaluation settings covering the four principles of the Accountable-Professional Persona studied in this work, R$^2$A overall outperforms both the base model and static Persona elicitation. Ablation results further show that Persona Representation Learning is critical for preventing Runtime Alignment from producing behaviorally imbalanced policies and for achieving more stable Persona policy learning.

Rubric-based reinforcement learning extends RL beyond tasks with exact answers or rule-based verifiers by scoring responses against instance-specific criteria. However, this makes reward computation expensive: training requires repeated rubric judging, often with proprietary APIs or local generative LLM judges with 7B parameters or more. We study whether smaller language models can serve as efficient and reliable rubric-based judges. To make this question measurable, we construct PointRubric and RaR-Science-Static, two pointwise rubric-based evaluation datasets with instance-specific criteria and itemwise satisfaction labels. We compare three ways of extracting criterion-level judgments from small models: Generative verdicts, Yes/No Logprob margins, and Probe judges. Across both datasets, the Qwen3-1.7B Probe judge achieves the strongest criterion-level agreement among these methods, outperforming Generative and Logprob judges. Used as a GRPO reward model, it trains a policy from 0.232 to 0.643 on RaR-Science rubric score, compared with 0.594 for an 8B Generative judge baseline, while the baseline requires 10.7$\times$ more reward-judge time. Task and domain transfer experiments further suggest that Probe judges preserve criterion-level reward structure across settings.

Flat retrieval-augmented generation treats a corpus as a bag of chunks, discarding document hierarchy and cross document structure. We introduce SearchWiki, a harness framework that synthesizes a corpus into a hierarchical, typed, navigable wiki and trains an agent, WikiResearcher-9B, to retrieve information through multi-turn tool use. The wiki organizes knowledge into three layers - document overviews, cross- document topic pages, and page-level source records; enabling progressive refinement of retrieval when initial lookup misses. We optimize the agent's navigation policy with on-policy reinforcement learning with a multi-component reward function balancing answer correctness, retrieval quality and trajectory efficiency. Evaluation on ViDoRe-V3 (8 domains), FinanceBench, and memory benchmarks (LoCoMo, LongMemEval, PersonaMem-v2) shows that WikiResearcher- 9B which is our RL-tuned Qwen 9B model, significantly outperforms same-size untrained baselines and exceeds or matches larger external models. SearchWiki paired with WikiResearcher-9B demonstrates that learned navigation over structured corpora is a superior alternative to flat retrieval.

Larger batches reduce the variance of stochastic gradients per update and are therefore often expected to accelerate training. Yet whether this statistical benefit translates into lower wall-clock time-to-target remains unclear, because each update consumes more samples and may take longer to execute. We study this tradeoff in reinforcement learning for large language models. We separate its algorithmic and systems effects by comparing learning and execution along their natural axes. At the algorithmic level, we compare configurations at equal cumulative sample counts while retuning batch-dependent hyperparameters. Over a bounded range of batch sizes, this procedure yields an approximately batch-size-invariant family whose members follow similar sample-indexed learning trajectories. At the systems level, we exploit the computational asymmetry between rollout generation and training: autoregressive generation is often memory-bandwidth-bound at low concurrency, whereas training work scales approximately with the number of processed tokens. Combining these two views yields a direct decision rule: a larger-batch configuration reduces time-to-target only when its throughput gain exceeds its samples-to-target penalty. Experiments with GRPO and PPO support both sides of this decomposition. At the algorithmic level, square-root learning-rate scaling with Adam produces approximately batch-size-invariant learning curves over a bounded range of batch sizes. At the systems level, larger batches improve generation throughput by up to 2.29x on fixed hardware. In GRPO, combining higher throughput with learning-rate retuning reduces time-to-target by up to 29%, whereas increasing the batch without retuning is slower despite its higher throughput.

Large language model driven search engines such as Google AI Overviews and Perplexity have created new opportunities for Generative Engine Optimization (GEO) the practice of refining content to increase its likelihood of being cited or summarized by generative systems. We demonstrate Agent2UCB, an agentic GEO system that autonomously improves content visibility through customized, feedback-driven optimization. For each content item, the system evaluates nine GEO strategies, identifies the most effective method, and accelerates selection using a bandit-based Agent2UCB policy that integrates LLM priors with online reward signals. To monitor side effects, the system also provides a lightweight, text-only SEO readiness evaluation covering readability, topical coverage, and EEAT-style credibility. Experiments on GEO-Bench show consistent visibility gains while preserving SEO quality. The demo allows users to choose the websites of interest, observe the optimization workflow, and compare GEO/SEO outcomes across methods.

Reinforcement learning with verifiable rewards (RLVR) substantially improves single-sample accuracy (pass@1) but causes the policy's solution space to contract, diminishing the returns of test-time scaling. In this work, we investigate where inside a reasoning trajectory this breadth is lost: does the policy fail to access a valid solution family, or does it fail to execute computation once initiated? To disentangle access from execution, we analyze the Countdown task, whose solution space can be exhaustively enumerated into discrete entrance families defined by the first operand and operator, across PPO on Qwen2.5-3B and GRPO on Qwen2.5-3B-Instruct. Across both training setups, solution coverage falls by up to 67%, halving even on problems solved across all checkpoints. We show that this contraction is heavily concentrated at the entrance: per-token likelihood shifts are 11x--16x larger prior to the first arithmetic operation than during downstream reasoning. Supplying only an unselected entrance prefix restores completion rates in low-access families by over an order of magnitude (0.018 -> 0.212 under PPO), demonstrating that alternative solutions remain executable but are no longer initiated. Guided by this localization, we find that while surface prompting fails to recover diversity, entrance-targeted interventions succeed: late-layer parameter interpolation with early checkpoints increases solution coverage by 37% at no loss in pass@1. Finally, we show that early-step entropy collapse recurs across six math benchmarks with 7B and 14B models, but is not an inevitable byproduct of reasoning optimization: an SFT baseline preserves more than double the coverage, and staged SFT--DPO--RLVR pipelines retain early-step entropy. In summary, reasoning breadth is lost at the door, not inside the room. Code: https://github.com/ershiyidian/early-branch-locking.

Arabic large language models must refuse harmful prompts without over-refusing benign or sensitive prompts, yet a single refusal rate hides this trade-off. We evaluate it using benign refusal B and harmful-prompt refusal H, where H measures refusal rather than harmful compliance. Across five Arabic-capable models and 130 runs on the full human-written AraSafe set, refusal-only supervised fine-tuning (SFT) collapses toward blanket refusal, whereas selected mixed-SFT configurations reach H = 90% to 93% at B = 14% to 23%; four selected configurations exceed the H = 90% target in all three runs, while Fanar does so in two of three. Direct Preference Optimization (DPO) and inference guards change B and H differently across models rather than acting as uniform upgrades. In a blinded 300-response audit, annotator binary-refusal agreement is 89.0% (kappa = 0.78); Qwen3Guard and Aya Expanse 32B reach 88.7% and 91.0% accuracy, respectively, with no conclusive paired difference. Selected SFT raises H on Arabizi for all five models, but none reaches 90%, showing only partial transfer from Modern Standard Arabic. Overall, the results support model-specific operating-point selection: set a deployment target and retain only interventions that improve it.

Standard clothing asset generation---restoring forward-facing flat-lay garment images from diverse real-world contexts---holds immense commercial value yet demands both macroscopic topological accuracy and microscopic physical fidelity. Although our previous work RAGDiffusion effectively eradicated large-scale structural hallucinations via retrieval-augmented macro-constraints, achieving industrial-grade micro-texture realism remains an unsolved bottleneck. We formally identify this limitation as High-Frequency Trajectory Collapse: supervised fine-tuning (SFT) converges to the conditional mean of the training distribution, which is dominated by smooth, low-frequency textures, causing high-frequency patterns (e.g., fabric weaves, intricate logos) to become nearly un-sampleable. Naively applying Reinforcement Learning (RL) post-training further triggers Artifact Hacking, where models exploit semantic biases in generic reward models by generating deceptive checkerboard noise. Our key insight is that RL can fundamentally reshape the sampling distribution of flow models---elevating the probability of high-fidelity trajectories under accurate reward guidance---while adversarial regularization prevents exploitation of reward blind spots. Realizing this principle requires three prerequisites: (i)inherent capacity, established through a 27,725-pair high-complexity garment dataset (STGarment-Plus) and a Dual-Image-Stream FLUX architecture upgrade; (ii)perceptive reward, provided by a novel attribute-aware reward model (Garment-RM) trained on 500K images via fine-grained contrastive learning, achieving 84.67% human preference accuracy; and (iii)hacking prevention, enforced by our Adversarial-Regularized GRPO (AR-GRPO) strategy that integrates a dynamic discriminator into the RL sampling trajectory to penalize artifacts while enriching authentic high-frequency details.

Multimodal search agents answer visual questions by interleaving image understanding, web retrieval, tool use, and evidence synthesis. Strong systems exist, but in two expensive regimes: proprietary frontier models such as GPT-5 and Gemini, or large open vision-language backbones trained with substantial agentic data and reinforcement learning. We ask a different question: when released agent trajectories are distilled into much smaller backbones under a single-node budget, what is actually transferred? We study this with LiteSearch-VL, a low-compute recipe for Qwen3-VL-2B and Qwen3-VL-4B that uses only released OpenSearch-VL trajectories, parameter-efficient LoRA adapters, and synthetic step-level preferences: DPO on GPT-5-generated hard negatives targeting five local failure modes (premature answer, wrong tool, weak query, repeated query, ignored image). Across 12,400 GPT-5-judged rollouts on SimpleVQA, FVQA, LiveVQA, and VDR-Bench-testmini, the dominant effect is behavioral rather than a uniform accuracy lift: full-trajectory supervised fine-tuning transfers the agent contract, taking the 2B model from almost never emitting a usable answer (1,237/1,240 no_answer rollouts) to 28.4% macro Pass@1, matching or slightly exceeding the off-the-shelf 4B base (25.6%). Synthetic preference learning and compact tool distillation act as refinements rather than phase transitions (best 4B configuration: 30.8% macro Pass@1). Finally, a controlled VDR step-budget ablation shows that extra search turns convert abstentions into wrong_entity errors rather than correct answers, identifying answer verification, not search depth, as the next bottleneck for small multimodal agents.

Diffusion-based visual generative models deliver strong image and video synthesis quality but incur high inference costs because sequential samplers repeatedly evaluate large networks. Caching-based methods reduce inference latency by reusing intermediate computations across adjacent timesteps. However, existing cache controllers rely primarily on local temporal variation and overlook the trajectory-level consequences of cache reuse. We introduce Error-Propagation-Aware Cache (EpaCache), a training-free caching policy that adaptively allocates the reuse budget on timesteps with lower downstream impact. Experiments on image and video synthesis models demonstrate that EpaCache consistently improves the latency--fidelity trade-off over existing caching methods. On FLUX.1-dev, EpaCache outperforms the prior state-of-the-art caching method in both latency and fidelity, reducing inference time from $11.7$ s to $11.3$ s while improving PSNR from $21.4$ to $22.8$. On HunyuanVideo, EpaCache achieves a $2.63\times$ speedup over uncached inference and improves SSIM from $0.891$ to $0.905$ over the prior state-of-the-art method at matched latency.

Large Language Models (LLMs) still exhibit limited capability in following complex instructions. While existing approaches often rely on preference learning to enhance this ability, they typically overlook the relationships between the permissible response spaces of different instructions, which restricts a model to align with subtle and diverse constraint variations. To address this, we propose Cross-Relational Preference Learning (CRPL), a novel framework for constructing preference data that explicitly models inter-instruction relationships through two key techniques: Cross-Relationship Perturbation and Cross-Region Pair Sampling. This enables the generation of more diverse preference data that captures a wide spectrum of constraint variations. Additionally, we introduce an atomic constraint-based verification mechanism to rigorously assess response satisfaction, ensuring high-quality preference pair construction. Extensive experiments across multiple preference learning methods (e.g., DPO, KTO), LLM backbones and four instruction-following benchmarks demonstrate that our approach achieves substantial improvements over prior baselines and exhibits strong generalization.

Urban construction governance requires early decisions that connect workplace safety, permitting requirements, and community impact, yet the relevant evidence is often scattered across separate municipal and regulatory data sources. This paper presents PermitGPT, a unified generative artificial intelligence framework for converting unstructured construction permit descriptions into structured decision-support outputs across three domains: safety hazard identification, permit requirement specification, and community impact assessment. To address data fragmentation, we spatially and temporally align records from the New York City Department of Buildings, Occupational Safety and Health Administration, and NYC 311 service requests, producing 90,000 structured prompt-response pairs derived through rule-based alignment and domain-informed spot checking. We fine-tune three open-weight language models using parameter-efficient adaptation and evaluate them on 2,833 held-out test cases. The results show complementary model behavior: Gemma-3-1B provides the most efficient inference at 3.07 samples per second with low memory usage, Llama-3.2-3B gives the highest lexical overlap for regulatory-style outputs with a BLEU score of 0.0091, and 4-bit Mistral-7B-Instruct-v0.3 achieves the strongest semantic alignment with a BERTScore-F1 of 0.7747. Because the task involves open-ended structured generation, low BLEU values are interpreted alongside semantic metrics and qualitative output structure rather than as standalone indicators of utility. Overall, PermitGPT provides an initial step toward AI-assisted construction governance while identifying directions for stronger task-level evaluation and real-world validation.

Joint-embedding predictive architectures (JEPAs) have primarily been developed for self-supervised representation learning. Denoising JEPA (D-JEPA) recently demonstrated strong generative capabilities on natural images, yet the applicability to 3D medical imaging remains unexplored. Building on the D-JEPA framework, we present Med-D-JEPA, a systematic adaptation and evaluation of joint-embedding predictive generation for 3D brain MRI. Med-D-JEPA operates on continuous latent tokens produced by a 3D KL-regularized adversarial variational autoencoder, and combines masked context prediction, representation-level alignment, per-token diffusion, and iterative next-set-of-token sampling. We evaluate unconditional and class-conditional generation quality on BraTS2019 and OASIS-1 datasets; downstream classification utility; and preliminary whole-tumor segmentation on BraTS2020. Across different generation settings, Med-D-JEPA achieves superior or competitive performance compared to several strong baselines on fidelity and diversity metrics. Compared to training with real samples, Med-D-JEPA-based synthetic pretraining improves classification AUC from 0.63 to 0.85 on BraTS2019 and from 0.78 to 0.87 on OASIS-1. In the segmentation study, pretraining on Med-D-JEPA samples improves Dice from 0.74 to 0.80 and reduces HD95 from 13.40 to 9.56 mm. These findings establish joint-embedding predictive generation as a promising direction for 3D medical image synthesis and encourage further research in this direction.

In financial markets, a sequential policy that reacts systematically to price movements may become predictable to other market participants. This paper studies whether large language model (LLM) agents exhibit such directional structure through RetailAgent, an experimental framework in which an LLM observes anonymized intraday equity price histories and permitted state, then repeatedly chooses long (hold the stock) or flat (stay out) before the subsequent interval return is revealed. We compare returns during long and flat intervals along the same stock's intraday path after removing the overall fraction of long decisions. This exposure-matched measure reveals persistent negative timing across modality, horizon, state, and model family. Shuffling saved action sequences substantially attenuates the effect, showing that alignment between actions and subsequent returns drives the negative score. Feeding self-authored memories into decisions further increases policy persistence, while timing becomes more negative among stock-days on which the agent uses both actions. These results reveal stable, recoverable directional structure in sequential LLM financial decisions and a behavioral signal for studying how another participant could respond to a predictable policy.

Large Vision-Language Models (LVLMs) remain prone to hallucinations, producing responses that are irrelevant or inconsistent with the multimodal input. Existing mitigation methods mainly rely on external supervision, output calibration, or attention regulation, leaving the internal representation dynamics of autoregressive generation underexplored. We identify an inference-time failure mode in which cross-modal representations degrade across decoder layers and drift across generation steps, destabilizing token prediction and increasing hallucination risk. We propose Dynamic Alignment Compensation (DAC), a training-free inference-time method that detects representation divergence and selectively applies lightweight residual compensation. DAC combines Layer-wise Semantic Compensation to mitigate inter-layer degradation with Sequential Semantic Correction to constrain temporal drift. Experiments on nine hallucination-focused and general-purpose multimodal benchmarks across multiple LVLM backbones show that DAC consistently reduces hallucinations while maintaining strong overall performance.

Multi-trait Automated Essay Scoring (AES) requires rubric-grounded reasoning across interdependent traits, rather than isolated score prediction. Existing feedback-enhanced methods often decouple feedback from scoring or assess traits independently, weakening score--feedback consistency and rubric alignment. We propose HiFTS, a unified autoregressive framework that generates hierarchical CoT feedback before predicting trait-level and holistic scores. HiFTS distills rubric-grounded hierarchical CoT feedback from a teacher LLM and trains student models to jointly generate feedback and scores. HiFTS further applies Group Relative Policy Optimization with a composite reward balancing score agreement, calibration, feedback quality, and structural validity. At inference, a lightweight global prior provides holistic guidance to reduce drift during long-form reasoning. We also introduce CFMS-34, a Chinese multi-trait AES dataset with 951 essays annotated with holistic scores and 34 rubric-based traits. Experiments on CFMS-34 and ASAP++ show that HiFTS achieves strong holistic and trait-level scoring while producing coherent, rubric-aligned feedback.

Recent generative approaches to geometry estimation adapt pretrained image diffusion models and treat the task as image-conditioned generation. Leveraging off-the-shelf image diffusion models, they either (i) train task-specific geometry models (for depth and surface normal estimation) independently, losing the opportunity of exploring the intrinsic correlation of these geometric targets, or (ii) jointly fine-tune modified image diffusion backbones (e.g., altered self-attention), which typically demands substantial labeled data. To overcome these limitations in a principled fashion, we repurpose pretrained video generative models as a unified and data-efficient framework for geometry estimation, formulated innovatively as a next-frames prediction task. Our method, GeoNeXt, inherits naturally structured knowledge and richer priors from the video model, while further adapting them for joint modeling of images and geometry targets (image <-> geometry), enabling more data efficient and effective learning of geometry. Extensive experiments validate our method for zero-shot monocular depth and surface normal estimation across diverse datasets, outperforming both previous task-specific and unified generative competitors while using substantially less training data. Notably, our method rivals discriminative state-of-the-art approaches trained on over 100x more data and even standouts on several benchmarks.

Spatio-temporal video grounding (STVG) requires models to identify when a referred event occurs and localize the target entity throughout that interval. Existing multimodal large language models typically serialize dense localization trajectories autoregressively, causing decoding latency to grow with tube length and allowing localization errors to propagate across time. We introduce Parallel Tube Decoding (PTD), a generative formulation that decomposes grounding into a temporal block followed by time-conditioned spatial blocks decoded simultaneously. This removes both token-level and trajectory-level dependencies, reducing the sequential decoding depth to a fixed $1 + 1$ rounds, independent of tube length. To enable parallel spatial generation, we introduce Decoupled Block Attention, which preserves access to shared video-query context while eliminating cross-box dependencies, together with localization-aware policy optimization for temporal boundaries and spatial geometry. On VidSTG, PTD reduces Tube Completion Latency by 79x and increases spatial decoding throughput by 92x over standard autoregressive decoding, while also improving grounding accuracy. With a compact 4B backbone, our model performs favorably well on VidSTG and HC-STVG, and generalizes zero-shot to temporal grounding, grounded VideoQA, and referring video object tracking. Our results show parallel tube generation is an efficient and effective alternative to autoregressive localization in videos.

Scripted and rule-based non-player characters (NPCs) in combat video games often exhibit predictable behaviors that experienced players can exploit, while reinforcement learning (RL) agents typically retain a fixed policy after training and cannot readily adapt their strategy to different opponents. We investigate a runtime strategy-selection framework in which a large language model (LLM) guides a trained RL policy without modifying its underlying behavior. To demonstrate this, we train five NPC agents with a shared PPO policy in Unity and compare a baseline configuration, in which the policy acts independently, with an LLM-augmented configuration in which a locally hosted Mistral 7B model, accessed through Ollama, reads the live game state every five seconds and assigns one of four tactical tags. We evaluate both configurations against three scripted opponent types across 600 episodes and analyze outcomes using the Mann-Whitney U test. Against a Balanced opponent that changes tactics during an episode, the LLM-augmented agents more than doubled their win rate from 11% to 24% and produced significantly longer episodes. Against an Evasive opponent, the augmented agents achieved a higher win rate and faster kills, although their shorter episode duration did not satisfy the strict hypothesis definition. Against an Aggressive opponent, the LLM's near-constant preference for encirclement was counterproductive. Analysis of 2,430 strategy selections showed that Surround was selected in 83.8% of cases regardless of opponent type, indicating limited zero-shot strategic differentiation at this model scale. These results demonstrate both the potential and limitations of LLM-guided runtime strategy selection for adaptive multi-agent game AI.

Long-horizon video generation is evaluated with whole-frame metrics that reward motion and temporal consistency. For fixed-camera nature scenes this creates an ambiguity: motion of water, fire, smoke, or rain is desirable, whereas motion of the background is an error. A system can therefore score well on motion while its scene drifts, or on consistency while its flow stagnates. We introduce SNF-Bench, an evaluation framework for long-horizon fixed-camera generation that partitions each scene into static support and dynamic flow and reports static fidelity, flow persistence with absolute magnitude, and drift leakage separately, never as one score. Drift leakage is interpretive context rather than a headline measurement. Each factor is validated mechanistically rather than by correlation with preference: we inject global translation, rotation, and scale drift and progressive late freezing at known severity into real generations, and require each factor to respond in its stated direction and to remain selective against corruptions it does not target. Auditing publicly released long-horizon text-conditioned checkpoints under one recorded common inference configuration, plus an image-conditioned track with released-pipeline references and a deployment-sensitivity panel, we find that whole-frame motion and static-region drift induce near-opposite orderings of the same outputs. At maximum controlled translation, fBD and NBF rise to $1.86\times$ and $1.32\times$ baseline, but whole-frame Dynamic Degree reaches only $1.07\times$---rewarding the corruption. SNF-Bench measures where motion occurs and whether it persists; it does not measure physical realism. Project page: https://minar09.github.io/snfbench/.

GUI agents trained with reinforcement learning (RL) have showcased strong environment learning capabilities on mobile platforms. However, RL typically demands extensive real-environment interactions, leading to high resource costs and instability, especially in GUI scenarios. To address these, we propose WM-R1, the first reinforcement learning framework that trains mobile GUI agents with world models instead of real environments. Specifically, world models serve as the source of state transitions during all rollouts, replacing the real Android environment within the training loop. WM-R1 also embeds world models directly into the thinking process, enabling agents to reason about the consequences of candidate actions before committing to the final action. Crucially, WM-R1 eliminates the need for real-environment interaction, supports massively parallelized and step-level granularized trajectory generation grounded in world models, and introduces a multi-dimensional rule-based reward that jointly optimizes task success, trajectory efficiency, and world model utilization. For efficient training, we curate a high-quality dataset of 2000 challenging tasks. Experiments on Android mobile benchmarks demonstrate that WM-R1-trained agents significantly outperform GRPO-only baselines and inference-time simulation methods. Code is available at https://github.com/genalyu/WM-R1 .

Rerankers, reward models and multi-document QA scorers score candidate documents or responses in one LLM prompt, so each score depends on their order. Such scorers are selected on ranking quality, but their scores determine a decision: what a score threshold retains, a reader answers, or a preference model selects. However, equal ranking quality does not imply equal decisions: on passage reranking, five trained scorers within 0.010 nDCG@10 retain sets that overlap by only 0.66-0.84 when reordered. A published reranker takes the highest retained-set F1 in our comparison and still overlaps by only 0.667. No prompt-time change we test removes that order dependence: the only one that gains ranking quality leaves all three decisions unchanged. Order-consistency SFT (OC-SFT) attenuates it in the weights, training a candidate's score not to depend on the order. It holds ranking quality and leads every decision-stability measure among trained scorers on all three tasks: it flips the reader's answer on 0.125 of permutation pairs against 0.149-0.164 for three other objectives that target order. It is more stable than order-averaged distillation on 12 base models, and one OC-SFT permutation retains sets that overlap more than ten averaged off-the-shelf permutations. A comparison should therefore report what a threshold retains and a reader answers, not ranking quality alone. Code is available at https://github.com/thomsonreuters/presentation-dependence.

While Multimodal Large Language Models (MLLMs) exhibit strong generalization, visual instruction tuning for downstream tasks inevitably causes catastrophic forgetting, impairing overall generalization. While existing methods regulate weight updates to reduce forgetting, they overlook the fundamental cross-modal alignment in MLLMs. Based on prior work and our observations, we argue that cross-modal alignment is implicitly captured in the information-compression trajectory. To preserve the alignment flow embedded in the trajectory, we propose LLaVAFlow, an information-theoretic distillation framework. First, we compress the mutual information between the extracted relations and MLLM embeddings, encouraging a learnable module to produce a refined alignment flow that benefits downstream tasks. Second, we maximize the mutual information between the extracted alignment flows of the pretrained and fine-tuned MLLMs, enabling the transfer of compact alignment information. Extensive experiments show that LLaVAFlow is an effective plug-and-play framework that preserves alignment flow and enhances both downstream performance and generalization.

Reinforcement Learning with Verifiable Rewards (RLVR) significantly improves LLM reasoning but often causes a drop in policy entropy, leading to narrowed reasoning coverage and degraded pass@$k$ for large $k$. While existing methods mitigate this entropy collapse through algorithmic regularizations, cross-model non-parametric perturbation is also neglected. In this work, we propose a simple yet effective approach to preserve the generative diversity of LLMs during RLVR. Instead of relying solely on internal exploration, we force the target model to generate answers based on partial reasoning trajectories generated by a smaller, weaker language models. These unfamiliar prefixes effectively disrupt over-confidence and encourage the exploration of distinct reasoning paths. We empirically study the potential of outer prefixes, revealing the mechanism of the impact of distributional discrepancy to the exploration dynamics in RLVR training. Experiments across multiple mathematical benchmarks show that our method consistently outperforms vanilla RLVR. Notably, the performance gain becomes increasingly pronounced as $k$ scales up, demonstrating a substantial expansion of reasoning coverage. Furthermore, our approach efficiently mitigates entropy collapse without requiring additional SFT, intricate reward designs, or complex prompting.