Generative RL 进展

Vision-language agents that crop and zoom are trained with rewards that credit a successful tool call, yet a successful call does not show that the model needed to look or used the pixels it received. On our cold-start checkpoint only 10% to 12% of visual calls were both needed and used, and released agents make spurious calls 36% to 87% of the time on individual benchmarks. Outcome rewards, judge rewards, and branch probes each observe one side of this failure, and about two thirds of what an outcome reward pays goes to calls that were neither needed nor used. CounterCredit asks both questions of every image-returning call at its realized pre-call state, using the policy's own gold-answer score. A decision value compares the realized visual branch with answering immediately; an evidence value compares the returned crop with random same-size patches substituted into the same call. A call verified on both earns cashback and every other executed call pays rent; the price is bounded so that every correct trajectory outranks every wrong one, and a dual-channel GRPO advantage keeps the price in its own units. From the same cold start, prompt pool, and budget, CounterCredit reaches 89.5% on V*, 80.2% on HR-Bench-4K, and 76.4% on HR-Bench-8K, 6.3 to 9.4 points above outcome-only GRPO at 1.78 against 1.84 calls per question, and lowers the spurious-call rate to 31% to 36%, the lowest among the agents evaluated. The same recipe lifts a Qwen3-VL-8B base from 75.4 to 80.8 on average.

Reinforcement learning (RL) is vital for optimizing video generation models, with a robust reward model (RM) serving as the cornerstone. However, existing video reward models often produce unstable scalar scores because they directly map complex, subjective video quality into a single score without explicit evaluation criteria. This leads to scalar drift, where the scoring scale collapses or shifts across different prompts, making the reward unreliable for RL. Drawing inspiration from professional human annotation engineering, we address this problem with RewardVerse, a rubric-based video reward framework that introduces a dynamic rubric as an intermediate representation between the evaluation query and the scorer. Instead of unconstrained direct scoring, RewardVerse first generates explicit evaluation criteria and then performs rubric-guided scoring, providing a stable semantic anchor that mitigates scalar drift. To efficiently optimize this collaborative pipeline, we propose Rubric-Guided Policy Optimization (RGPO), a two-stage training algorithm. RGPO first warms up the scorer using self-evolving seed rubrics and then jointly optimizes the rubric generator to produce query-adaptive evaluation criteria while continuously aligning the scorer with human ratings. Extensive experiments on the 16-dimensional EvalVerse benchmark and external datasets demonstrate that RewardVerse mitigates scalar drift, achieves state-of-the-art performance on both pointwise and pairwise evaluation, and provides a robust and interpretable reward signal for RL in video generation.

Chain-of-thought (CoT) can sound plausible yet be unfaithful to the model's underlying reasoning. Most prior work probes CoT faithfulness through input--output behavior or input attributions, leaving internal computation largely underexplored. We instead cast faithfulness as internal concept grounding: Does a large language model's (LLM) CoT reasoning engage the same internal concepts that support the LLM's direct prediction, and do the shared concepts causally drive its answer? Encoding a prediction pass and a CoT pass with a single shared sparse autoencoder (SAE), a reliable approximator of the latent concepts LLMs use, makes their internal concepts directly comparable. We introduce three correlational metrics of concept-level alignment and a causal metric, $Δp$, which ablates the shared concepts and measures the drop in answer probability. Across five LLMs and four datasets, concept alignment is generally high, as indicated by the correlational metrics; yet these only identify which concepts are shared, not how much they causally contribute. $Δp$ fills this gap: causal faithfulness varies substantially with model depth, peaking at mid-to-late layers rather than the final ones, and model scale reshapes the layer-wise profile. Moreover, causally important shared concepts are not always verbalized in the CoT. These dissociations suggest that faithfulness cannot be reliably assessed from surface-level or representational correspondence alone; assessing it requires causal tests of whether the internal concepts underlying a CoT actually drive the model's prediction.

On-device inference is booming, but the momentum is almost all in language models. Diffusion pipelines are memory hungry, latency-sensitive, and require orchestrating an embedder, a transformer, a decoder, and often further postprocessing that is not as standardized as LLM inference loops are. We navigate the trade-off between performance, quality, and model footprint to reach as many client devices in the wild as possible. We make three contributions: an embedding translator that maps a small text encoder into a large encoder space to cut weight and latency; a reproducible sweep recipe for navigating the speed/quality/memory triangle in diffusion pipelines; and an interactive on-device image generation editor achieving sub-second TTFI on recent GPUs.

Self-play methods that co-train a single language model as both coder and test author promise to move code-generation RL beyond fixed test suites, but they suffer from two coupled pathologies: permissiveness collapse, where pass-rate rewards are maximised by trivial, non-discriminative tests, and concentration bias, where i.i.d. sampled tests cluster on modal inputs and inflate estimator variance. We introduce CoVer (Co-trained Coder and Verifier), a single-policy GRPO framework that addresses both failure modes. First, an information-gain (IG) reward scores each self-generated test by the mutual information between its pass/fail vector and a graded, ground-truth-anchored correctness signal y [0, 1] m, gated by the sign of their covariance so that only positively discriminative tests receive reward. Second, a three-stage diversity-aware selection step prunes a candidate pool to a behaviourally non-redundant suite (invalidity, input-string, execution-profile filtering), raising the effective sample size of the IG estimator at fixed execution budget. On five benchmarks (LiveBench, MBPP, LiveCodeBench, CodeContests, Code-Forces), CoVer raises one-shot pass@1 by +5.8 points at 7B and +7.1 points at 14B over the Qwen2.5-Instruct backbone, and achieves the highest macro-average among all compared methods at both scales. As a drop-in backbone inside the CodeT ranking pipeline, CoVer-7B adds +3.5 points, demonstrating the dual benefit of co-training for both generation and selection.

Monocular 6-DoF pose estimation of non-cooperative targets is important for on-orbit servicing and debris removal. A single-image estimator can confuse near-symmetric spacecraft orientations, and tracking can preserve an incorrect pose. We present Hypothesis-Anchored Tracking (HAT), a causal framework that uses inter-frame motion to select among competing CAD-based pose hypotheses before alignment and fusion. Rather than independently choosing the highest-scoring hypothesis in each image, HAT retains competing orientation histories and selects a pose to anchor the relative trajectory estimated by monocular SLAM. Sparse anchors and pose fusion provide per-frame estimates after initialization without revising past outputs. The method requires only a calibrated RGB sequence, a metric CAD model, and target image regions, which can be supplied by detection or segmentation. The pretrained pose and SLAM networks require no target-specific training or fine-tuning. We evaluate two versions, Mega-HAT and Pico-HAT, using MegaPose and PicoPose, on SPARK-2024, SwissCube and SHIRT, with YCB-Video assessing performance outside the space domain. Using one temporal configuration per method, the arithmetic means of the four dataset-wise comparisons show 9.4% lower mean pose error and 3.76 times the sustained input FPS for Mega-HAT relative to independent MegaPose, and 23.9% lower mean pose error and 2.42 times the FPS for Pico-HAT relative to independent PicoPose. Mega-HAT ablations on SPARK and an offline reference examine component contributions and the effect of revising past estimates.

Reference-based metrics for Grammatical Error Correction (GEC) such as M$^2$ and ERRANT assume that the reference set enumerates all valid edits, and therefore often penalize corrections that are grammatical and meaning-preserving but phrased differently. We introduce RM-EVAL, a reward model trained on human preference data from SEEDA, as a reference-free meta-evaluator that predicts human-like quality judgments at both full-sequence and partial-sequence levels. Beyond evaluation, we show that the same reward model can be used as a learning signal to improve GEC generation via Reward-Guided Text Generation (RGTG), which keeps a base GEC model frozen and performs online, reward-driven decoding. Across SEEDA, RM-EVAL achieves strong agreement with human rankings, and RGTG yields consistent gains in reward and external validation, demonstrating a unified framework for both assessing and enhancing GEC systems without relying on gold references.

Reinforcement learning is increasingly used to align image generators with reward signals, and Flow-GRPO recently extended this paradigm to flow-matching models by treating the denoising sampler as a stochastic policy that can be optimized from reward feedback. Training in this setting is unstable in a way specific to multi-step denoising: the policy update changes systematically across denoising steps, with importance ratios drifting below one, becoming increasingly dispersed, clipping at different rates, and leaving fewer usable samples late in training. Prior work treats these effects as separate failure modes and addresses each with a hand-tuned stabilizer. We show instead that they arise from a single per-step quantity, which we call path variance. This quantity is determined exactly by the sampler's Gaussian transition kernel and can be estimated cheaply during training. This reframes instability as a resource that can be measured and budgeted rather than a collection of symptoms to repair. Our method, $λ$-Controlled GRPO, calibrates importance-ratio behavior from this predicted law rather than from noisy empirical statistics, and allocates gradient effort across denoising steps according to their predicted cost. The two scales governing the update are fixed by standard policy choices rather than introduced as free tuning parameters. On a text-to-image model under two reward settings, rendering difficult target text scored by optical character recognition and matching human preferences scored by a preference model, $λ$-Controlled GRPO improves both text accuracy and preference reward over the strongest empirical stabilizer. It also keeps late-step path variance within its intended budget, precisely where the baseline systematically overshoots. The result is a Flow-GRPO update calibrated by its own transition law rather than stabilized after instability appears.

Reinforcement learning has substantially improved large language model (LLM) agents in verifiable domains, but remains difficult to apply to open-ended agent tasks, where solutions are diverse and reliable scalar rewards are hard to obtain. Recent pairwise evaluation methods alleviate reward discrimination collapse by replacing pointwise scoring with relative preferences. However, they still compress rich comparative feedback into a single trajectory-level reward, obscuring decisive intermediate steps and preventing successful behaviors from being consolidated into reusable skills. We propose ArenaFlow, a hierarchical credit propagation framework for open-ended agent reinforcement learning. ArenaFlow leverages tournament-based relative ranking to derive trajectory-level reward signals. Each comparison is further equipped with structured reflective evaluation, which reveals three types of supervision: pivotal success steps, reusable strategy skills, and usage attribution of retrieved skills. At the step level, ArenaFlow propagates trajectory-level advantages to high-confidence pivotal steps according to tournament survival depth, enabling more targeted optimization of local reasoning behaviors. At the skill level, ArenaFlow estimates skill utility from group-level usage attribution and maintains a global skill memory through utility-aware updating, pruning, and retrieval. The resulting high-utility skills further serve as policy priors for future exploration. Extensive experiments validate ArenaFlow's effectiveness on open-ended agent tasks.

We propose a Verification Reward Model (VRM) framework for training language models to create and repair chip verification artifacts. The central object is a versioned verification contract that binds requirements, permissible stimulus, observation boundaries, reference behavior, evaluation budgets, and acceptance criteria. Compiler, simulator, formal, mutation, coverage, and expert-review evidence are converted into auditable records. Deterministic acceptance checks remain outside the learned model. A learned outcome model predicts expensive future evidence from the specification, the generated artifact, and explicitly masked partial evidence; a semantic critic identifies evidence-supported weaknesses; a deterministic reward composer translates the resulting quality vector into task-conditioned training rewards. Extensions include paired clean/fault interventions, marginal fault-discovery rewards, uncertainty-aware evaluation scheduling, and a quarantined cross-domain learning loop. Evaluation emphasizes independently validated fault detection, false alarms, generalization across design families, and total cost to reach a specified quality level. We describe a first experiment on a small, open-tool-compatible benchmark with lightweight models, with full UVM capability admitted through feature-specific qualification. This is a position paper: we specify the framework and the experiments that would test it, and we report no training, EDA, or silicon results.

Large Language Models (LLMs) are increasingly deployed in applications that must weigh clashing moral values, yet even strong models exhibit hidden biases and brittle instruction-following across languages. We introduce a 12,000-instance dataset of two-option dilemmas covering pairwise three value conflicts: Honesty vs. Justice, Justice vs. Autonomy, and Autonomy vs. Honesty, along with their translations into Hindi, Arabic, Spanish, and Chinese, to probe cross-lingual behavior. Benchmarking on GPT-5-mini reveals that it consistently favors Honesty over Autonomy across all five languages when no policy is given. The Llama-3.2-1/3B models exhibit strong first-option bias; however, both plain fine-tuning and Direct Preference Optimization fine-tuning effectively remove this bias, increasing accuracy to greater than 98%. In order to decouple the effect of learning correlations in the dataset from abstract values, we propose a task vector transfer based experiment where after computing the task vectors for a direction of value preference we orthogonalize it with respect to the general instruction following vector. Our experiment shows that this method is effective in isolating the direction of the specific value preference that can successfully be used to conduct task arithmetic to obtain a model with the opposite stance.

Satellite communications are an essential component of sixth-generation (6G) mobile networks, which provide ubiquitous connectivity for global services. However, the satellite uplink remains a critical bottleneck for ground devices: their limited transmit power and antenna apertures result in low data rates and high packet errors. To overcome this bottleneck, this paper advocates a novel relaying paradigm termed generate-and-forward (GF) communications, where satellites exploit on-orbit generative artificial intelligence (AI) to robustly reconstruct corrupted data prior to forwarding. Specifically, we propose SpaceDiffusion, an over-the-orbit diffusion framework for satellite-assisted image transmission. The core of this framework is a channel-distortion-aware diffusion theory developed using the following approach. By formulating the recovery of compressed and lost image tokens as an inverse problem, this theory incorporates a channel-distortion correction term directly into the conventional denoising diffusion implicit model (DDIM) update. As a result, this design enables a single pretrained diffusion model to adapt dynamically to varying packet-loss patterns and compression distortions without retraining. Furthermore, we analytically characterize the progressive token-reconstruction error and derive a diffusion-step activation threshold that predicts when SpaceDiffusion is expected to outperform conventional decode-and-forward (DF) relaying. Building on these theoretical insights, we further develop an energy-aware early-exit policy to efficiently deploy SpaceDiffusion in orbit. Experimental results demonstrate that SpaceDiffusion achieves lower end-to-end latency compared to DF scheme with retransmission protocol and saves approximately 15 dB of uplink transmit power at a target perceptual quality.

Evaluating live streaming speech synthesis (TTS) requires assessing fine-grained, highly expressive prosody such as emotion, intonation, and energy which traditional MOS predictors fail to capture. While proprietary Large Language Models (LLMs) like Gemini can evaluate these aspects, they are too costly for massive inference and reinforcement learning feedback. To address this, we first introduce Live-ProsodyJudge (LPJ), a cost-effective pairwise evaluator distilled from Gemini into Qwen3-Omni. However, we identify a critical flaw in standard multi-dimensional evaluation: verdict coupling. The judge tends to lazily align all individual dimension scores with its overall preference, collapsing a rich multi-dimensional rubric into a single preference bit. To resolve this, we further propose Decoupled-Live-ProsodyJudge (D-LPJ). D-LPJ eliminates the overall verdict target to prevent blind following, masks uncertain pair-dimensions during Supervised Fine-Tuning(SFT), and introduces a novel span-local GRPO strategy that applies normalized advantages strictly to their corresponding rationale spans. Evaluated on highly curated human-annotated test sets, 10 sample balanced-order LPJ achieves higher point accuracy than a single Gemini call, while D-LPJ successfully produces independent,decoupled dimension judgments. Furthermore, in a Best-of-8 TTS candidate selection tournament, the LPJ-selected utterance falls within the human top-3 in 85.29% of high-confidence cases, demonstrating its efficacy for fine-grained TTS preference optimization.

Cross-view object geo-localization (CVOGL) typically assumes a fixed query image, overlooking the ability of mobile agents to actively acquire more informative observations. To address this limitation, we introduce Active Cross-View Object Geo-Localization (ActiveGeo), where an agent sequentially selects new viewpoints and determines when to stop, aiming to improve localization with minimal observations. We further propose ActiveMoPT, an ActiveGeo framework with three-stage training. First, Multi-View Prompt-Preserving Adaptation enables the model to aggregate multiple query views while reusing the initial prompt. Second, Trajectory-Guided Policy Initialization uses supervised agent trajectories to learn viewpoint selection and initial stopping behavior. Third, Cost-Aware Policy Refinement employs GRPO with a gain-cost reward to jointly optimize localization accuracy and observation efficiency. We also construct ActiveGeo-858, a zero-shot test set containing 858 scenes and 1,716 target annotations. Experiments show that ActiveMoPT achieves state-of-the-art performance on MoP-UAV using only 1.45 query views on average, and substantially outperforms previous CVOGL approaches under zero-shot evaluation on ActiveGeo-858.

We propose SnapPhysics, a training-free framework that reconstructs 3D objects and estimates their physical properties such as mass, friction, and center of gravity from a single image. For physically coherent interactions in mixed reality (MR), such properties are as important as geometry. Prior approaches infer them by analyzing object dynamics in video, which is computationally costly, or by querying vision-language models (VLMs) on single images, which lacks geometric grounding and inter-object relationships. We address these limitations by combining instance-level 3D reconstruction and spatial alignment with a physics-aware scene graph that encodes these relationships and per-object metric geometry as structured context for VLM-based property reasoning. Experiments on 3D-FRONT show that SnapPhysics improves scene-level F-Score by 18.6% over the best learning-based method, and on real captured scenes with ground-truth mass, it reduces the mean absolute log difference error (mALDE) by up to 20.5% and improves log-scale correlation ($r^2_{\mathrm{ls}}$) by up to 19.6% over VLM-only estimation. SnapPhysics enables physically interactive MR experiences without manual parameter tuning. Project page: https://snapphysics-ismar2026.github.io/.

Autonomous recovery of a micro unmanned aerial vehicle (UAV) onto a moving airborne carrier enables reusable deploy-mission-recover operation, but couples long-range rendezvous, close-range perception, carrier motion, aerodynamic interaction, and a discontinuous contact event. This paper presents an RTK-vision-guided reinforcement-learning framework in which a child UAV is physically transported by a larger carrier, takes off from the carrier while airborne, executes an independent sortie, returns to the carrier's current position, redocks, and subsequently descends with the carrier. Both vehicles carry RTK-GNSS, and the carrier continuously shares its navigation state with the child. Near the recovery deck, RTK remains active while a downward-facing camera with a fiducial marker detector provides marker-relative alignment cues. A proximal policy optimization (PPO) policy governing the terminal recovery phase is trained in a physics-based MuJoCo simulation environment with explicit sensor noise models, an aerodynamic disturbance surrogate, and marker-latency randomization, then transferred to hardware. PX4 retains low-level stabilization, and a deterministic safety gate authorizes descent independently of the learned policy. The PPO checkpoint achieves 99.55% success over 2,000 held-out randomized terminal episodes, compared with 78.4% for a tuned PD baseline under identical conditions, with a median planar terminal error of 6.62 cm. Across 14 outdoor trials, the full mission succeeds in 13 trials (92.9%), spanning both near-region recovery and recovery after the carrier translates away from the release point. The results demonstrate a complete autonomous aerial deployment-and-recovery cycle rather than an isolated landing maneuver, establishing a practical basis for reusable carrier-child operation in inspection, surveillance, and mobile-logistics applications.

Training embodied foundation models typically requires massive-scale datasets and extensive computational resources, yet often suffers from three critical limitations: (1) inefficient sample utilization due to low-informative samples; (2) imbalanced gradient contributions across heterogeneous tasks; and (3) severe credit assignment problem in long-horizon planning, where trajectory-level rewards indiscriminately penalize all tokens. To address these issues, we propose an efficient training paradigm that achieves state-of-the-art average performance through strategic data selection and hierarchical policy optimization. Our approach consists of three synergistic stages. First, Rejection Sampling-based Fine-Tuning (RSFT) filters out low-informative samples to establish robust behavioral priors while preventing distributional collapse. Second, Iterative Rejection GRPO (IR-GRPO) employs task-specific queues stratified by difficulty to keep datasets balanced across reinforcement learning iterations, coupled with a hybrid reward mechanism for precise cross-task feedback. Third, to enhance long-horizon task planning, we introduce Trie-GRPO, a novel reinforcement learning algorithm based on action prefix trees, which enables step-level advantage estimation. This resolves the credit assignment problem by isolating intermediate correct decisions from downstream errors, while effectively balancing exploration efficiency and depth compared to conventional search trees. As a result, EmbodiedMind achieves a state-of-the-art average performance of 70.02% across 18 benchmarks, and significantly outperforms other embodied foundation models in long-horizon task planning accuracy. Our project will be released for reproducibility.

Tool learning enables large language models (LLMs) to use external tools for tasks beyond parametric knowledge. Reinforcement learning can optimize tool-call behavior from feedback, but current methods still face two problems: fixed-threshold curricula can become misaligned with the policy's evolving capability boundary, and additive rewards can leak argument-level credit when the predicted tool is wrong. To address these problems, we propose MATCH, a closed-loop framework for model-aware tool learning with curriculum scheduling and hierarchically gated rewards. Model-Aware Curriculum Learning (MACL) maintains reward-derived sample difficulty that co-evolves with the policy, and each epoch selects samples near the current capability boundary together with a top-k pool of harder cases. Hierarchical Tool-call Gated Reward (HTGR) scores tool name, argument key, and argument value as a gated chain, granting credit at each level only when prerequisites hold. The same HTGR rewards drive both GRPO updates and MACL's difficulty refresh, closing the loop between policy optimization and sample scheduling. On API-Bank and BFCL V3, MATCH reaches 72.19% and 62.87% overall accuracy, outperforming the main supervised and RL-based baselines. Backbone experiments further show consistent improvements across four backbones from two model families.

Self-evolving methods reduce the need for human-annotated trajectories by allowing tool-using agents to generate their own training data. Yet existing methods typically separate trajectory generation from evaluation, relying on static verifiers that cannot adapt to emerging failure modes or self-consistency signals that may reinforce errors shared across trajectories. Jointly adapting planning, execution, and evaluation offers a promising alternative, but introduces a fundamental coordination challenge: each component continuously changes the data or feedback used to train the others. We address this challenge with UnifiedPlayers, a cooperative framework comprising a Planning Player that generates tasks, an Execution Player that produces multi-turn trajectories with Python tool calls, and an Evaluation Player that constructs executable verifiers. We design role-specific rewards that coordinate the three players toward a shared learning objective under GRPO. Across two model backbones and twelve reasoning benchmarks, UnifiedPlayers outperforms the strongest prior baseline by at least 3.5% on mathematical reasoning and 3.9% on general reasoning tasks. Moreover, the learned verifier achieves 84.2% adversarial detection accuracy, while its reward signal exhibits 2.03$\times$ higher per-question variance than a self-consistency baseline, providing more discriminative verifications. These results highlight cooperation among specialized players as a promising path toward self-enhanced tool-integrated agents.

Reinforcement learning for LLM agents involves two distinct optimization di- mensions: how environment feedback is exploited within a trajectory, and how complete trajectories are aggregated across a batch. We formulate these dimen- sions as Intra-Trajectory Feedback Attribution and Inter-Trajectory Objec- tive Aggregation, and introduce BATON (Bayesian Attribution and Trajectory Objective Normalization), a dual-axis policy optimization framework. BATON instantiates the first axis with Bayesian Feedback Attribution, which constructs a feedback-conditioned posterior over sampled actions, and the second with Trajec- tory Mass Normalization (TMN), which assigns equal optimization mass to com- plete trajectories. Experiments with GRPO and GiGPO on ALFWorld, WebShop, and SearchQA show that both axes provide independent gains and that their combi- nation consistently achieves the strongest overall performance across model scales.

With the widespread industrial deployment of Large Language Models (LLMs), DeepSearch has emerged as the dominant paradigm for resolving complex user queries. It typically operates through an iterative closed-loop workflow consisting of planning and reflection, information retrieval, and answer generation. However, existing reward models (RMs) and evaluation benchmarks are primarily designed for static single-turn tasks, failing to capture the full-pipeline complexity of DeepSearch workflows. To address this limitation, we propose F2DR, a fine-grained full-pipeline DeepSearch reward framework. F2DR evaluates DeepSearch workflows across three dimensions: Content, Trajectory, and Answer, enabling comprehensive process-level assessment. We further construct DeepSearch RM-Bench, a dedicated benchmark for evaluating RMs in DeepSearch scenarios. Extensive experiments demonstrate that F2DR achieves significantly higher evaluation consistency than self-evaluation-based baselines, while DeepSearch RM-Bench exhibits strong discriminative capability across existing open-source RMs. We will publicly release the complete DeepSearch RM-Bench dataset soon.

Video diffusion models repeatedly process long spatiotemporal token sequences during denoising, making attention a major computational bottleneck. Linear attention offers an appealing alternative and has been widely adopted in recent large language models, but directly applying it to video models often fails to preserve the fine-grained interactions required for high-quality generation. We present Video DeltaNet (VDN), which combines local Softmax attention with bidirectional linear memory for long-range video context. Its linear branch introduces Video Delta Attention (VDA), which updates memory once per frame by jointly incorporating its spatial tokens. Separate output projections and learnable gates calibrate the two branches, while a staged teacher-alignment recipe progressively introduces the new pathway into pretrained models. We instantiate VDN on MiniMax H3, applying the hybrid to video-to-video interactions while retaining Softmax for interactions involving text or audio. With eight-step distillation and an optimized SGLang serving stack, VDN-H3 completes DiT denoising for a 14.3-second, 768p video in 6.70 seconds on eight NVIDIA B200 GPUs, corresponding to a 14.5x speedup over the 50-step dense H3 baseline on the same GPU count. GitHub code available at: https://github.com/OpenVDN/vdn-minimax-h3. Weights available at: https://huggingface.co/OpenVDN/vdn-minimax-h3

Hateful memes spread abusive content through implicit interactions between images and text, posing serious threats to the safety of online communities. In recent years, multimodal large language models have been widely used for hateful meme detection and are increasingly adopted to generate explainable detection results. However, we find that existing explain-then-detect methods often couple explanation generation and label prediction within the same training process. This coupling causes interference between task objectives, leading to limited detection performance and even worse results than simple SFT baselines. To address these challenges, we propose ProKDA, a progressive knowledge-to-decision alignment method for explainable hateful meme detection. Inspired by the human annotation training process, ProKDA first uses an agentic background knowledge construction pipeline to obtain external knowledge related to meme understanding. It then adopts a three-stage training strategy that sequentially performs background knowledge learning, hatefulness detection learning, and hatefulness boundary alignment. Unlike prior explain-then-detect methods that jointly optimize both tasks, ProKDA focuses on a single training objective at each stage. This design reduces interference between the two tasks and progressively transforms background knowledge into robust detection decisions. Experiments on three public hateful meme benchmarks show that ProKDA achieves state-of-the-art detection performance and provides accurate, explainable, and evidence-supported decisions for hateful meme moderation. Project page: https://meizhiyuan88666.github.io/prokda.

This paper introduces and operationalizes summarization bias: a proposed systematic tendency of large language models (LLMs) to represent narrative meaning as an abstract summary label rather than as the reconstructable inferential structure that produces it. Within the Bulut Doctrine, narrative effect is theorized along a told-shown axis: in told mode, emotional and informational content is declared explicitly and requires little reader reconstruction; in shown mode, that content is suppressed at the surface and must be reconstructed from physical cues and indirection (Objective Projection). Shown mode is the higher-load condition the doctrine is designed to measure. The claim is that LLMs fail along this axis in a specific direction. Summarization bias is hypothesized to operate in two regimes: (i) a generative regime, in which a model asked to render an emotion through Objective Projection defaults to declaring it instead; and (ii) an evaluative regime, in which a model judging narrative quality rewards told-mode explicitness and under-detects shown-mode suppression. The evaluative regime is the more consequential, since LLMs increasingly serve as judges and reward models, and a directional bias toward told mode would impose a selection pressure degrading prose toward flat declaration. This report does not claim the bias is validated. It defines the construct, situates it against LLM-as-judge biases, rereads a completed independent reliability study as directional evidence consistent with it, and pre-registers a two-regime test with decision rules under which the construct would be abandoned.

Semantic parsing (SP)-based knowledge base question answering aims to answer natural language questions by generating executable logical forms (LFs) over knowledge bases (KBs). When applying Large Language Models (LLMs) to this task, a key challenge over large, heterogeneous KBs is selecting question-related schema elements (i.e., relations and classes) and composing them into complex LFs. Recent LLM-based methods often make early discrete commitments to schema elements during intermediate reasoning, allowing incorrect intermediate schema decisions to propagate and finally result in incorrect LFs. To overcome this limitation, we propose SALR, a schema-anchored latent reasoning method for LF construction. It performs multi-step reasoning by generating continuous thoughts in the model's hidden states, thereby delaying the explicit commitment to LF decisions. To ground this latent reasoning process in the corresponding KB schema, SALR aligns continuous thoughts with a codebook of KB schema elements through an alignment objective supervised by schema traces deterministically derived from gold LFs. It then incorporates the aligned schema codes into inputs for subsequent reasoning steps. This schema-mediated feedback guides LF generation without requiring the model to emit an explicit textual reasoning trajectory. Experiments on GrailQA and WebQSP show that SALR achieves consistent overall gains over strong baselines. Notably, on compositional questions from GrailQA, SALR outperforms TIARA, a strong SP-based baseline, by 2.86 F1 points. Further analyses show that schema-mediated feedback affects LF generation and that schema information is recoverable from the latent states.

Safety tuning pipelines judge only the final answer, which makes it difficult to distinguish robust refusal from two undesirable shortcuts: blanket refusal on benign requests and polished but unfaithful safety rationales that do not actually constrain the answer. We propose AUDITPLAN, a single-model plan-then-answer approach where the model first emits a compact structured safety plan and then answers conditioned on it. The plan records a threat label, intended action, and explicit constraints, enabling machine-checkable auditing while remaining hidden from users at deployment. We train this behavior with supervised fine-tuning followed by reinforcement learning with FAITHGATE, a reward-gating objective that grants answer reward only when the safety plan is correct. This discourages safe-looking but unfaithful behavior and promotes tighter plan-answer coupling. Across Qwen backbones, AUDITPLAN improves both robustness and auditability: on Qwen2.5-3B-Instruct, FAITHGATE reduces ASR from 24.0% to 11.6%, LSR from 1.0% to 0.36%, and over-refusal from 11.0% to 2.0%, outperforming answer-only RL, free-form explanation, and weighted-sum structured rewards. Similar trends hold for Qwen2.5-1.5B-Instruct. Larger-model confirmation runs on Qwen-3-4B-Instruct and Qwen2.5-7B-Instruct preserve the same trend suggesting that explicit internal commitments can make safety alignment more faithful, robust, and auditable.

Text-based person retrieval under a sim-to-real gap (synthetic training data, a real-image gallery) is usually tackled with costly fine-tuned cross-encoders. We ask whether a frozen-encoder system can compete. We present SCOUT, which casts cross-modal retrieval as prediction in embedding space. A trainable predictor maps the patch tokens of a frozen video encoder into the embedding space of a frozen text encoder under a bidirectional InfoNCE objective, and no encoder is fine-tuned in the base model. The video encoder is V-JEPA, the text encoder is EmbeddingGemma, and the predictor is initialized from a Qwen3.5-0.8B decoder. We make three findings. First, the best frozen text encoder is simply the one whose geometry best matches the video features. A training-free alignment score ranks three candidate text encoders in the same order as their retrieval accuracy on our held-out split (Spearman $ρ= 1.0$); a fourth, LLM-based encoder shows the rule is metric-dependent, holding for a neighborhood-overlap score ($ρ= 0.8$) but not for a linear probe ($ρ= -0.2$). Second, two precision-targeted levers, parameter-efficient ExPLoRA adaptation of the video encoder and a training-free attribute-decomposed reranker built on a vision-language model, improve the top-rank precision that otherwise limits the frozen system, adding 2.2 points of leaderboard R@1. Third, a local-versus-public calibration study explains which interventions transfer to the real domain. On AI City Challenge 2026 Track 4 the full retrieve-fuse-rerank system reaches 84.25 mAP@10 on the final leaderboard, while a single frozen model submitted alone reaches 60.63. Our trained components cost about 95 GPU-hours. CMP, the dataset authors' fine-tuned cross-encoder that trains for sixteen GPU-days, is one fusion member of the full system, not an alternative. Code and annotations: https://github.com/abtraore/SCOUT-ECCV

Multi-modal talking avatar synthesis aims to generate realistic talking videos from a reference portrait and speech. Despite rapid progress in diffusion-based methods, existing approaches still struggle to jointly achieve accurate lip articulation, human-preferred motion aesthetics, and efficient inference. We observe that phonetic accuracy and motion aesthetics arise from fundamentally different sources and should be addressed at complementary stages rather than learned implicitly by a single generator. Based on this insight, we propose VibeAvatar, which disentangles these two objectives through a Phonetic Kinematics Adapter (PKA) that converts recognition-oriented speech features into phonetic-kinematic conditions at the conditioning stage, and an Aesthetic Motion Policy (AMP) that optimizes a flow-consistent stochastic sampling policy via Group Relative Policy Optimization (GRPO) at the post-training stage. With a lightweight flow-based motion generator operating in a compact 1D warp-based latent motion space, VibeAvatar achieves state-of-the-art results in articulation, aesthetics, and efficiency on both objective metrics and user studies, while generating a 10-second 512px video in under 10 seconds with only $\sim$3GB VRAM.

Large language models (LLMs) often suffer from capability stagnation in self-improvement training because fixed difficulty levels fail to adapt to their evolving proficiency. To address this issue, we propose STRETCH (Self-Taught Reasoning Evolution via Targeted CHallenge), a unified framework inspired by cognitive scaffolding theory. STRETCH introduces a dynamic Stretch Zone mechanism that continuously aligns question difficulty with the model's solving capability. Within a single parameter space, the model alternates between a Scaffolder that generates adaptive, boundary-pushing challenges and a Learner that that optimizes its solving trajectories through reinforcement learning. This dual-loop co-evolution effectively stabilizes training, mitigates reward hacking and promote progressive reasoning growth. Experiments on both negotiation and operation research benchmarks demonstrate that STRETCH consistently outperforms strong prompting and domain-specific baselines. Further scaffolder configuration analysis shows that dynamic difficulty alignment is critical for sustained capability improvement and synchronized reasoning evolution.

Streaming automatic speech recognition (ASR) must be judged jointly on what it transcribes and on how quickly it commits each word. Delayed streams modeling (DSM) has become the dominant paradigm for streaming large audio-language models, exposing a structural delay $τ$ that bounds the decoder's lookahead. We show that $τ$ is a poor proxy for user-perceived latency, and that the alignment-based supervision of DSM leaves latency on the table: the same forced-aligned transcript is used at every $τ$, forcing the model to withhold words it could already commit. We introduce AWED, a word-level emission-delay metric defined relative to the acoustic end of each word, and post-train a DSM recognizer with GRPO under a reward that scores transcription accuracy and measured delay jointly. Trained at a single operating point ($τ=6$ frames), our model dominates both its supervised fine-tuning initialization and the Voxtral Realtime backbone across all evaluated lookahead budgets: it cuts WER by 30.8% relative at an 80\,ms structural delay, and by 5.7% relative at 480\,ms while lowering median AWED from 1.17\,s to 1.04\,s. Latency-rewarded post-training thus advances the accuracy--latency Pareto frontier of streaming ASR without architectural change.