Generative RL 进展

Large language model (LLM) agents integrated with external tools are vulnerable to indirect prompt injections embedded in environmental states. However, existing studies largely rely on manually implemented or reused environments, stochastic LLM-based tool simulation, and predefined injection locations, limiting scalable security research across broader domains. To bridge this gap, we propose **ToolHazard**, a scalable adversarial environment synthesis framework that reduces human engineering and supports expansion with additional seed domains and compute. Through an Environment Simulator, an Attacker Agent, and a User Simulator, ToolHazard synthesizes executable stateful environments, discovers viable injection points and generates environment-specific payloads, and constructs state-grounded long-horizon tasks. Based on ToolHazard, we build **ToolHazard-Bench** for stress-testing agents under complex workflows and diverse environmental attacks. Experiments reveal substantial agent vulnerabilities and show that injection timing and placement affect attack effectiveness. Moreover, ToolHazard-generated alignment data improves security on both ToolHazard-Bench and AgentDojo while preserving benign task utility.

We consider the setting of Normalizing flows with approximate inverses, an established paradigm spanning both full-dimensional ($d=D$) and bottleneck ($d<D$) settings, and group these models under the term flow autoencoders. We present a theoretical investigation into their training dynamics and prove that the proposed loss used by existing approaches is suboptimal; specifically, both encoder and decoder surrogates must be optimized in alignment with reconstruction loss. Guided by these insights, we propose Normalizing Autoencoder (NAE), which employs a novel conditional loss that aligns the surrogate loss gradient with that of reconstruction loss, directly improving upon the current standard. Extensive experiments across molecule generation, tabular data, and image benchmarks demonstrate that NAE achieves state of the art performance. Our work highlights the importance of loss alignment in flow autoencoders and establishes NAE as a powerful generative framework.

Existing Vision-Language Models (VLMs) exhibits a critical bottleneck in robust spatial reasoning. Recent reinforcement learning (RL) methods aim to close this gap with verifiable outcomes, yet they suffer from poor credit assignment across intermediate reasoning steps. Concurrently, structured reasoning approaches overlook the critical depth perception necessary for comprehensive 3D understanding. To address these challenges, we propose SCOUT (Structured Chain-Of-Thought Utilizing Process-Supervised RL Training). Specifically, we design a structured Chain-of-Thought (CoT) framework that explicitly models 3D environmental perception to ensure robust spatial understanding and reasoning. Furthermore, we introduce a novel RL algorithm featuring multi-objective process rewards and a tailored advantage estimation method, facilitating fine-grained credit assignment across distinct segments of the reasoning trajectory. To support our framework, we develop SCOUT-24k, a structured spatial reasoning CoT dataset synthesized through a customized pipeline. Extensive evaluations demonstrate that SCOUT-3B improves upon baseline models by 16.85% and 6.3% on general spatial benchmarks and complex spatial reasoning tasks respectively. Notably, our larger SCOUT-7B even outperforms GPT-4o by a margin of 4.28%. Moreover, despite being trained exclusively on single image, SCOUT-7B exhibits robust out-of-domain generalization to multi-image and video scenarios. These empirical results render SCOUT as a critical step towards next generation of spatially-aware VLMs.

Long-form song generation models continue to improve in duration, structural integrity, and acoustic complexity, making reliable aesthetic rewards increasingly important for aligning these models with human preferences. However, reward models for complete songs remain limited, and existing evaluators typically predict scores in a single forward pass without providing readable explanations. We introduce MUSECRITIC, a semi-scalar reward model that generates a natural-language critique covering five aesthetic dimensions and uses it as an intermediate representation to predict continuous reward scores. MUSECRITIC follows a two-stage training pipeline: a teacher model first provides high-quality critiques for supervised fine-tuning, after which the fine-tuned model generates its own critiques for reward learning, mitigating distribution shift between training and inference. On an in-domain test set of 200 SongEval songs, MUSECRITIC reduces macro-averaged mean squared error from 0.2875 to 0.2316 and improves macro-averaged LCC, SRCC, and Kendall's tau to 0.9068, 0.8838, and 0.7178, respectively. On the out-of-domain Music Arena benchmark with 733 preference pairs, it achieves the highest accuracy of 71.35%. Moreover, using MUSECRITIC with GRPO improves Muse-0.6B on all nine aesthetic metrics from SongEval and Audiobox Aesthetics. These results demonstrate that critique-conditioned reward modeling reduces scoring error and provides an effective optimization signal for song generation. The project repository is available at https://github.com/WuqnEl/MuseCritic.

Video-to-audio (V2A) generation faces significant challenges in achieving precise temporal synchronization and high perceptual quality due to the complex, ambiguous relationship between visual and auditory cues. Existing methods typically compress video inputs into single feature representations, leading to significant loss of temporal dynamics and fine-grained visual information. These approaches also rely on reconstruction-based training objectives that poorly correlate with human perceptual judgments of audio quality and appropriateness. We propose HarmoniDPO, a novel framework that integrates preference-based optimization into diffusion-based V2A generation to address these limitations. (1) Our approach leverages a dual video representation: combining global context with frame-wise features to preserve temporal dynamics and semantic detail. (2) Inspired by reinforcement learning from human feedback (RLHF), HarmoniDPO employs online Direct Preference Optimization (online-DPO) to fine-tune a diffusion-based V2A model from preference judgments, enhancing perceptual quality and alignment. (3) Additionally, we introduce Dual-scale Diffusion Search (DDS), a test time scaling algorithm that adaptively optimizes output fidelity during inference. Experiments demonstrate that HarmoniDPO outperforms state-of-the-art methods in audio-video synchronization and subjective audio quality, offering a robust solution for generating realistic, human-preferred audio from video.

Generative models like Diffusion Models and Flow Matching have demonstrated remarkable capabilities in synthesizing high-fidelity driving videos, but are severely constrained by high inference latency due to the requirement of extensive sampling steps. We argue that this inefficiency stems from the prevailing reliance on a standard Gaussian source distribution, where consecutive frames are initialized as independent Gaussian noise. This paradigm disregards the rich spatiotemporal correlations inherent in driving videos, compelling the model to regenerate deterministic scene structures existing in previous frames from noise, which is both computationally redundant and prone to geometric inconsistency. To address this problem, we propose GeoFlow, a novel framework designed to achieve efficient driving video generation by harnessing explicit geometric priors. Instead of sampling from standard Gaussian noise, we leverage multi-view geometry and spatially-adaptive noise injection to construct a Geometry-Aligned Prior (GAP) distribution as starting point. This initialization bridges the gap between source distribution and data distribution, yielding a significantly straighter and shorter sampling trajectory. Extensive experiments demonstrate that GeoFlow can achieve remarkable efficiency of both training and inference: merely several hours of fine-tuning on baseline models can significantly boost few-step generation quality, while fully converged training drastically reduces number of inference steps required for state-of-the-art video generation.

The reliability of Large Language Models (LLMs) for API calling degrades in multilingual settings. A common failure occurs when a model selects the correct tool but generates argument values in an inconsistent language, which we term Argument Language Mismatch (ALM). Although semantically correct, such outputs are operationally invalid and not captured by standard API-calling metrics. We revisit post-training strategies for mitigating ALM and find that, in our benchmark, supervised fine-tuning (SFT) provides a strong baseline, substantially improving argument language consistency and end-to-end function call accuracy. Under consistent model selection, SFT achieves performance comparable to, and sometimes exceeding more complex reinforcement learning (RL) approaches. We further examine whether RL with structured, argument-aware rewards offers additional benefits. While methods such as Group Relative Policy Optimization (GRPO) can improve language consistency and better preserve general reasoning ability, these gains are incremental and most pronounced in generalization and multi-objective trade-offs. Overall, our results suggest that much of the performance in multilingual API grounding can be achieved through careful supervised training, with RL providing targeted rather than fundamental improvements.

Multimodal large language models (MLLMs) have made rapid progress, yet they still exhibit object hallucination, generating plausible but incorrect descriptions that are inconsistent with the visual input. Direct Preference Optimization (DPO) mitigates this by training models to prefer non-hallucinated responses over hallucinated ones, and recent efforts further enrich the preference data with relevant context. However, it remains unclear whether DPO actually leverages such context. To investigate this, we propose Contextual Preference Gain (CPG), a simple metric that measures how much a model's preference strengthens when relevant context is provided. We find that higher CPG consistently corresponds to lower hallucination, yet standard DPO and its variants exhibit only limited CPG, indicating that they underutilize contextual information and thus remain prone to hallucination. To address this, we propose Context-Calibrated DPO (C$^2$-DPO), which directly maximizes CPG while preserving the original preference ordering. Across multiple benchmarks, C$^2$-DPO substantially reduces hallucination without compromising general reasoning, relatively reducing the Object HalBench hallucination rate of Qwen2-VL-Instruct-2B by 36%. Code is available at https://github.com/mlvlab/C2-DPO

Task-driven 3D affordance grounding aims to localize the functional region in a cluttered 3D scene that enables an action specified by a natural-language instruction. Existing methods either predict 3D masks directly or construct them by selecting and fusing intermediate 2D/3D regions. However, they remain vulnerable to two intertwined failure modes: the predicted or selected regions may miss the target interaction area or have unsuitable granularity, while language grounding may confuse visually similar alternatives under relational instructions. To this end, we introduce ThinkAfford, which decouples high-recall affordance proposal generation from instruction-grounded reasoning. Specifically, the Affordance Proposal Generation module first uses learnable affordance prompts and multi-level visual features to predict interaction-conditioned heatmaps, extracting a variable number of fine-grained proposals without parsed object or part names as segmentation prompts. Visual-Prompted Affordance Reasoning then reasons over labeled proposal overlays using the full instruction, returning identifiers in a structured "think-then-answer" response. Moreover, Group Relative Policy Optimization uses proposal-level rewards from lifted 3D overlap to align VPAR selection with final 3D grounding. On the SceneFun3D validation split, ThinkAfford achieves 10.69% AP50 and 25.46% AP25 under the official evaluator, outperforming comparable 3D open-vocabulary and vision-language-model-based 2D-to-3D baselines. Module-level diagnostics further show that APG attains 77.5% recall at 25% intersection-over-union, while GRPO-trained VPAR achieves 72.1% selection accuracy on APG-covered queries, compared with 63.4% under supervised fine-tuning.

Evaluating detailed image captions from Vision-Language Models (VLMs) requires going beyond surface-level semantic similarity. Reference-based metrics (e.g., CIDEr and SPICE) and LLM-as-scorer protocols struggle to verify dense factual claims, while existing QA-based alternatives generally offer lower probe density, narrower domain coverage, or no explicit alignment between individual questions and segmented image regions. We introduce CapProbe, a full-scene dense QA benchmark that turns detailed caption evaluation into region-aligned factual checking. Each image is decomposed into coarse semantic regions covering both foreground and background elements; for every retained region, we generate multiple-choice questions spanning 10 semantic categories, forming a dense checklist of probed visual facts. Guided by a two-tier taxonomy of 37 L1 domains and 219 L2 sub-domains, CapProbe comprises 346 images, 1,868 regions, and 25,650 questions, averaging 74 QA pairs per image. A language judge answers from the caption alone; an Uncertain option and Effective Accuracy provide a judge-dependent proxy for distinguishing unanswered probes from incorrectly resolved ones, while density-based metrics penalize verbose yet uninformative captions. The protocol is cost-effective: by converting unconstrained scalar scoring into structured MCQ reading, it reduces open-ended scoring bias while remaining judge-conditioned and yields relatively stable model rankings under a fixed reader. Experiments on 13 VLMs show large Coverage gaps across models, a clear competency-efficiency trade-off, and failure modes that sparse or overlap-based evaluation often misses. The benchmark data, annotations, and evaluation code will be released soon.

Existing Multimodal Large Language Models (MLLMs) predominantly rely on image-text pairs for modality alignment pretraining, mapping global image representations to long textual descriptions. However, this image-level alignment suffers from referential ambiguity: models struggle to infer the correspondences between multiple visual objects and textual entities from the global representation, leading to data inefficiency and suboptimal semantic grounding. To address this, we propose MultiModal Code-Switching (MMCS), a novel pretraining paradigm that provides explicit object-level supervision. Inspired by the linguistic phenomenon of code-switching, MMCS interleaves vision and language by replacing textual entities with their corresponding visual objects, enforcing local vision-language grounding. We further develop a scalable data synthesis pipeline to generate a pretraining dataset of 773K samples with accurate object-entity correspondences. Experiments show that MMCS is highly data-efficient: with only 50K samples, it matches or surpasses models trained on 600K image-text pairs. Furthermore, MMCS consistently improves visual grounding and perception capabilities across varying model scales.

Large vision-language models (LVLMs) remain vulnerable to jailbreak attacks that exploit visual inputs to bypass safety alignment inherited from their language backbones. We propose SafeCap, a reinforcement-learning framework that aligns LVLMs through learned self-captioning. SafeCap trains a policy model to first generate a safety-relevant image caption and then produce a final answer; the caption is further optimized by whether it enables a frozen LLM to reach a safety-aligned decision. This caption-mediated objective encourages the policy to expose visual cues relevant to safe response generation rather than relying solely on direct refusal supervision. Across five multimodal safety benchmarks and six vision-utility benchmarks, SafeCap substantially improves aggregate safety performance under its intended DirectCap protocol, with gains of 3.7-19.0 points in safety average across four model settings while maintaining comparable or improved vision utility. Under controlled comparisons on matched backbones and data, SafeCap outperforms safety SFT, DPO, and SafeGRPO, demonstrating the effectiveness of caption-mediated reinforcement learning for multimodal safety alignment.

While foundation models have significantly advanced human recognition across diverse modalities, they predominantly rely on static, geometric feature extraction. This approach fundamentally diverges from human perception. Consequently, current models often suffer from "semantic blindness," overfitting to transient noise while failing to leverage invariant soft biometrics, and struggle to capture temporal motion signatures. To bridge this gap, we propose SapiensID 2.0, a human recognition framework enriched with both semantic and temporal awareness. To overcome the lack of soft-biometric annotations, we transfer zero-shot semantic knowledge from Multimodal Large Language Models (MLLMs) into a discriminative embedding space. We resolve the dimensional mismatch between these spaces using Invariant Trait Alignment (ITA) to distill core persistent traits, and Transient Noise Disentanglement (TND) to decouple artifacts like clothing. Furthermore, we design a Kinematic Semantic Attention Head (K-SAH) that extends spatial attention across temporal windows. By tracking semantic patches over time, K-SAH captures rich kinematic signatures without requiring large-scale video datasets. Extensive experiments demonstrate that SapiensID 2.0 achieves state-of-the-art performance across image- and video-based person re-identification and gait recognition, while maintaining robust face recognition capabilities.

Large Language Model (LLM)-based simulators often reproduce observable actions but fail to capture the underlying reasoning behind them. In education, where student simulation is increasingly used for various applications such as evaluating tutoring systems, this gap is especially pronounced. Two students may submit identical submissions for entirely different reasons. We present INTERNAL STUDENT DIALOGUE (INSIDE), a student modeling framework that fine-tunes LLMs not only to act like students but also to think like them. INSIDE generates internal dialogue grounded in Bloom's Taxonomy across cognitive, affective, and action dimensions, and fine-tunes models on paired think traces and actions. We baseline against different prompting frameworks and evaluate on two axes: fidelity of simulated actions and quality of generated internal dialogue. Our evaluations show that INSIDE improves simulation fidelity in both action fidelity, matching code generation of real students, and reasoning alignment, achieving the highest alignment across models up to 57.9%.

Feedback signals used to train Large Language Models (LLMs) are the primary driver of their behavior and our main lever for instilling alignment with human values and objectives. However, a key limitation of current post-training methods is the inability of human annotators and automated reward functions to faithfully capture the feedback we would like to give. We introduce Evaluation-Conditioned Training (ECT), a post-training framework that uses natural language to condition each training sample on the fidelity of the feedback we provide and then elicits the desired behavior by conditioning the LLM on a high-fidelity monitor in deployment. ECT is aimed at improving performance under imperfect feedback and works as an add-on to existing algorithms such as SFT and PPO. We first provide a conceptual framework for ECT and discuss its potential to address persistent sources of reward mis-specification. Then we motivate ECT in the context of the eliciting latent knowledge (ELK) problem. Finally, we evaluate ECT on two proof-of-concept experiments: increasing even-handedness in news article generation and reducing sycophancy on an arithmetic task. In each setting, we utilize imperfect feedback, rewarding bias and agreement with the user, respectively. In both settings, ECT improves the targeted behavior relative to direct training.

Reinforcement Learning from Human Feedback (RLHF) aggregates heterogeneous preferences into a single reward model, assuming preference homogeneity. When preferences are heterogeneous, this aggregation induces a procedural fairness failure where majority preference groups dominate reward learning while minority preferences are systematically under-represented. This work defines procedural fairness in alignment as preserving distinct preference signals during reward modeling and shows that standard RLHF violates this via preference averaging. Preference-Aware RLHF (PA-RLHF) is introduced, separating optimization across preference modes at the reward learning stage. In a controlled setting, PA-RLHF improves overall alignment accuracy from 46.9% to 67.9% and reduces the fairness gap between best and worst aligned groups from 15.9 to 9.6 percentage points. These results show that procedural fairness failures in alignment can arise from structural design choices in reward learning, even in controlled, noise-free settings, with direct implications for large language models and agentic systems, where biased reward models can compound inequities across sequential decisions.

Neural language models trained on large crowdsourced corpora frequently exploit spurious surface patterns tied to target labels without true linguistic or causal relevance, boosting benchmark performance while failing on adversarial or out-of-distribution inputs. Existing approaches either require manual specification of the feature vocabulary or automate discovery only partially, leaving the gap between dataset-level correlation and model-level exploitation unaddressed. We present U N M ASK, a fully automated pipeline that discovers, causally verifies, and mitigates spurious correlations in text classifiers without additional human annotation. Given unlabeled training examples, U N M ASK generates candidate surface patterns as executable boolean expressions, filters them through a statistical validation protocol with independent replication, and establishes causal model dependence via verified counterfactual interventions. Causally confirmed features then serve as annotation-free group definitions for Deep Feature Reweighting, eliminating the group labels that standard DFR requires. Applied to BERT and RoBERTa trained on MNLI, our pipeline independently rediscovers established lexical-overlap and negation biases, verifying 9 of 10 features on BERT and 6 on RoBERTa, and improving HANS accuracy by up to 12.58 pp. On CivilComments-WILDS, programmatic groups match the 70.1% worst- group accuracy of hand-labeled DFR (Kirichenko et al., 2023) without demographic annotation. We further demonstrate that the discovery and validation stages generalize to reward model preference data, surfacing interpretable spurious correlations in RewardBench2.

Despite significant advances in instruction-following and auditory comprehension, the evaluation of Emotional Intelligence (EI) in Spoken Language Models (SLMs) remains confined to rudimentary paralinguistic perception, lacking a systematic, theory-driven cognitive framework. We introduce EmoSBench, the first comprehensive EI evaluation benchmark for SLMs constructed upon the four-branch theoretical model, covering Perceiving, Understanding, Using, and Managing Emotion across ten sub-tasks. Preliminary assessments on EmoSBench reveal a substantial gap: even leading proprietary models like GPT-4o-Audio achieve only 52.6%, significantly trailing human baselines. To bridge this gap, we develop EmoS, a specialized evaluator model optimized via Supervised Fine-Tuning (SFT) and Group Relative Policy Optimization (GRPO). To facilitate its effective training, we curate EmoDialogue, a bilingual dataset providing necessary fine-grained supervision through response pairs with rigorously defined EI gradations. Concurrently, we introduce a reward mechanism integrating a Steep Exponential Accuracy Reward (SEAR) and a Rationale Fidelity Reward (RFR) to enforce precise ordinal scoring and valid reasoning. Experiments demonstrate that EmoS reaches 83.8% accuracy, approaching human-level performance. Furthermore, evaluations on authentic, unconstrained spoken interactions validate its robust real-world generalization, establishing a foundational framework for advancing emotionally intelligent dialogue systems.

Autoregressive language models align training and use: generation conditions on a clean prompt, and training predicts future tokens from clean left context. Diffusion language models offer parallel denoising, but native dLLM pretraining can randomly corrupt prompt and continuation tokens together, weakening the clean-prefix interface needed for prompt-conditioned generation. We identify this mismatch for prompt continuation and propose PCD (Prefix-Conditioned Diffusion), a pretraining objective that combines AR prefix supervision with no-shift suffix denoising. At the training-objective level, PCD changes the attention mask, corruption mask, and label construction in continued pretraining; it does not require an autoregressive decoder, verifier, or new inference mode. By supervising the clean-prefix side autoregressively and applying diffusion only to the unknown continuation, PCD makes the local training interface resemble how block-diffusion models are queried at evaluation time. We further separate intra-sample prefix conditioning from inter-sample objective mixing, allowing us to identify the local alignment signal separately from the optional batch-level mixing knob. Across LLaDA2-Mini and Qwen-1.7B backbones, PCD consistently improves over same-family native dLLM stable baselines, reaching a 4.2% relative gain on the main LLaDA2-Mini six-benchmark average (+2.56 points) and a 14.2% relative gain in the primary Qwen mechanism comparison (+4.86 points). These results suggest that aligning the pretraining context distribution with prompt-conditioned generation can recover a measurable part of the dLLM continuation gap without changing inference.

Monocular 3D object detection spans two regimes: closed-set detectors operating within a fixed category vocabulary, and open-vocabulary detectors that localize arbitrary categories by leveraging depth foundation models for 3D geometry. We find that current depth foundation models, despite their strong zero-shot generalization, lack the object-level precision 3D detection demands: substituting a state-of-the-art depth foundation model for a strong detector's predicted depth degrades accuracy, even falling below the detector's own prediction. Rather than pushing detectors or depth models to be more accurate end-to-end, we treat object-level depth refinement as a stand-alone task and present RefineAny3D, a vision-language model that corrects depth without ever predicting a numerical value. Our key insight is that depth error has a direct visual signature in image space: when projected onto the image, a correctly placed box tightly encloses the object, while a too-far box projects too small and a too-close box projects too large. Depth refinement thus reduces to a visual alignment problem rather than a metric regression problem, which we instantiate by extending the VLM's vocabulary with action tokens that replace numerical depth output with categorical decisions, and by supervising the model on a large-scale chain-of-thought dataset that grounds each decision in explicit visual evidence. Applied as a single post-hoc step, RefineAny3D delivers consistent gains across closed-set detectors, open-vocabulary detectors, and 3D auto-labeling tools, and generalizes to novel categories, scenes, and cameras without retraining.

Aligning Large Language Models (LLMs) for open-ended tasks is challenging because responses must satisfy multidimensional criteria without following a single correct generation trajectory. Existing rubric-based reinforcement learning (RL) methods compress fine-grained criterion-level feedback into scalar rewards, making persistent capability gaps difficult to target under limited on-policy exploration. We propose $RISE-RL$ (Rubric-Informed Selective Exploration), which uses repeatedly missed rubric criteria to elicit privileged trajectories that are difficult to discover through unguided exploration alone. RISE-RL retains only trajectories whose complete-rubric reward exceeds the mean reward of natural rollouts, and then re-evaluates them under the original prompt to emphasize behaviors that remain weakly supported by the natural policy. The resulting guidance signal is optimized through a separate auxiliary objective and removed once its additional benefit diminishes. Experiments with 4B and 14B models across writing, chat, health, and science show that RISE-RL achieves the highest mean score on every evaluated benchmark under guidance-free evaluation. Compared with standard Rubric-RL, it improves the average score by 1.3 points at the 4B scale and $3.3 points at the 14B scale$, including a $6.0-point$ gain on CreativeWriting-V3. It also improves creative-writing diversity and yields gains on objectively scored medical and scientific benchmarks. These results indicate that selective internalization through reward filtering and policy support shaping is effective for open-ended reinforcement learning.

Aligning large language models (LLMs) with human privacy preferences requires capturing individuals' disclosure boundaries beyond general privacy norms. However, a gap remains in eliciting such nuanced preferences to evaluate alignment in realistic settings. We introduce CIDER, a dataset of 14,850 human annotations from 169 users, forming 1,650 contextual disclosure boundary sets across 60 interpersonal communication scenarios involving information sharing that violates privacy norms. Each boundary represents a real user's disclosure decisions over 9 sharing variants in a scenario, for a given communication role and AI-mediated condition. We formulate a task in which models predict a user's disclosure decision from historical boundaries, with varying levels of contextual information. Across 12 open and proprietary models, in-context personalization improves prediction accuracy by up to 11.41 percentage points using only 6 historical examples. Larger models such as GPT-5.4 (with medium reasoning effort) and Claude Sonnet 4.6 are better at leveraging semantic context to understand user-specific, context-dependent disclosure preferences for more accurate predictions, while smaller models tend to rely on structured heuristics based on disclosure granularity and identifiability. Personalization generally improves prediction accuracy, but the improvement is often accompanied by imbalanced shifts in false-positive and false-negative rates across models, with only Claude Sonnet 4.6 achieving balanced improvements in both. Our findings reveal both the promise and limitations of inference-time personalization for privacy preference modeling and position CIDER as a resource for advancing personalized privacy alignment.

Direct Preference Optimization (DPO) aggregates token-level log-probability ratios via uniform summation, implicitly treating all tokens as contributing equally to the preference signal. However, the contribution of individual tokens to the preference signal varies. We introduce token credit, which modulates each token's KL regularization based on its contribution to the preference outcome. We derive that effective token credit is proportional to the magnitude of each token's implicit reward, and observe that this quantity evolves substantially during training. This implies that static token credit becomes increasingly misaligned as training progresses. In this work, we propose Se-DPO (Self-Evolving Token Credit for DPO), a live mechanism that derives token credit from the model's own evolving internal signals during DPO training. Since the reward signal varies in reliability across positions, Se-DPO calibrates token credit based on both the strength and the confidence of each token's contribution. Se-DPO requires no external models, adding only a lightweight calibration network with minimal computational overhead. Experiments show that Se-DPO improves over DPO by up to 9.8 points on AlpacaEval~2 and 12.2 points on Arena-Hard.

Efficient text-to-image generation requires both reinforcement-learning (RL)-based reward alignment and few-step distillation, yet these procedures are typically performed sequentially, increasing training cost and risking the loss of reward gains during compression. We instead take an RL-native perspective: diffusion RL already generates reward-scored finite-step trajectories, whose intermediate states provide a natural source of distillation supervision rather than a disposable byproduct of sampling. Based on this insight, we propose REST (Reward-Enhanced Scored-Trajectory Distillation), a single-stage RL-distillation co-training framework that attaches a decoupled student to an arbitrary RL teacher. The student learns segment-wise from the teacher's evolving rollout trajectories while leaving the original teacher optimization unchanged. To prevent uniform imitation from preserving undesirable low-reward behaviors, we further introduce Advantage-Modulated Distillation (AMD), which transforms rollout advantages into signed weights over a base distillation loss. AMD strengthens supervision from preferred trajectories and mildly repels the student from low-reward ones. The resulting framework is lightweight and plug-and-play, requires no extra image rollouts, no separate distillation dataset, and no adversarial training. Experiments on compositional generation, visual text rendering, and human-preference alignment show that REST enables few-step CFG-free inference that matches or surpasses its 40-step RL teacher, with an overall additional training cost below 25% over pure RL. REST improves DrawBench PickScore over RTDMD by 0.82 while requiring only one-fifth of the training iterations.

Video coding is advancing into the low and ultra-low bitrate regime, driven by end-to-end codecs that replace the hand-crafted pipeline with jointly optimized neural networks and generative codecs that exploit the priors of video generation models. Yet the dominant metrics, LPIPS and DISTS, measure feature and texture similarity rather than content fidelity: a reconstruction that hallucinates a wrong face or blurs text into convincing strokes can still score well, even when a human rejects it instantly. To address this, we propose CodecArena, the first vision-language framework for video coding quality assessment, casting codec evaluation as source-conditioned comparative reasoning between a reference and its reconstructions. We optimize CodecArena with Facet-GRPO, a visual reinforcement learning scheme that aligns pairwise codec preferences while grounding the verdict in five fidelity facets: identity, objects, text, texture, and temporal consistency. Its facet-anchored reward uses automatically derived facet directions as weak anchors, rather than human per-facet labels, to prevent any single sub-score from dominating the holistic preference and to yield interpretable fine-grained quality judgments. To support training and evaluation in this underexplored regime, we construct two complementary resources: CodecArena-1K, a fully automatic preference dataset of 1,500 comparison groups built from traditional, neural, and generative codec reconstructions with fused vision-language and objective supervision; and CodecArena-Bench, a human-ranked benchmark with source-disjoint videos for fair out-of-domain evaluation. Extensive experiments demonstrate that CodecArena achieves state-of-the-art agreement with human judgments on source-disjoint content across diverse codecs and bitrates, surpassing perceptual metrics and prior vision-language evaluators.

Text-to-image diffusion models have achieved remarkable success in generating high-quality images from a given text prompt. Subject-driven generation aims to synthesize customized images to mimic the appearance of subjects in given reference images within different visual contexts specified by the text prompts. The central challenge here is that, when the reference image changes, the diffusion model cannot efficiently adapt to different visual contexts while consistently maintaining the subject identity. Existing methods either train the model with a large domain-specific dataset or fine-tune the model using the reference image for hundreds of iterations before actual image generation. In this work, we explore a new approach, called In-Loop Model Adaptation (IMA), which adapts the core diffusion model at each generation step during the actual process of image generation, without being trained on the reference image before the generation process. To this end, we establish a DDIM inversion chain that maps the reference image to a sequence of latent, as well as a text-to-image generation chain which generates the image from the text prompt only. We then introduce a masked latent consistency loss and a noise regularization loss to characterize the latent-noise difference between the diffusion model and these two chains at each generation step. This coupled latent-noise loss is used to guide the in-loop model adaptation to preserve the subject identity specified by the reference image while maintaining accurate alignment with the text prompt, resulting in high-fidelity text-to-image generation. Our extensive experiments demonstrate that our proposed IMA method significantly improves the performance of subject-driven text-to-image generation.

External natural-language skills provide large language model (LLM) agents with reusable and editable guidance for solving complex tasks. Yet their effectiveness depends not only on skill quality, but also on whether the policy can translate the provided guidance into appropriate actions. However, methods specifically designed to improve this skill-utilization ability remain largely underexplored. In practice, skill-based agents are commonly trained with reinforcement learning objectives centered on task-level rewards, which offer limited supervision and struggle to capture subtle differences in how effectively the policy uses the provided skills. We propose BCSD (Bidirectional Context Self-Distillation), a framework that combines self-distillation with reinforcement learning to train LLM agents to use external skills more effectively. Unlike prior self-distillation methods that rely on a single privileged context, BCSD evaluates each trajectory from two complementary skill-context views. The augmented view introduces higher-level Meta-Skill guidance, while the reduced view prunes general guidance to highlight task-specific skills. Their complementary token-level signals are combined to rescale the RL advantage. Experiments on ALFWorld and WebShop demonstrate that BCSD achieves the strongest overall performance across model scales, enabling agents to utilize external skills more effectively. Ablation studies further verify the complementary contributions of the augmented and reduced context views. Code will be released to ensure full reproducibility.

Recent advances in visual generative models have enabled high-quality image and video generation, but evaluating these models often demands sampling hundreds or thousands of images or videos, which is computationally expensive. Existing evaluation methods also rely on rigid pipelines that overlook specific user needs and provide numerical results without clear explanations. Mimicking how humans quickly form impressions of a model's capabilities from only a few samples, we propose the Evaluation Agent framework, which employs human-like strategies for efficient, dynamic, multi-round evaluations, offering detailed, user-tailored analyses. Given a natural-language evaluation request, the agent decomposes it into sub-aspects, generates targeted prompts, samples images or videos from the evaluated model, invokes suitable evaluation tools, and iteratively updates its plan from the observed evidence, covering both predefined benchmark dimensions and open-ended user concerns. The framework is thus efficient, promptable, explainable, and scalable across models and tools. Experiments show that Evaluation Agent reduces evaluation time to 10% of traditional methods while delivering comparable results. We further introduce Open Evaluation Agent (Open-EA) by constructing EA-CoT-10K, a corpus of history-conditioned step-level instruction-tuning records derived from multi-round evaluation rollouts, and training EA-3B from Qwen2.5-3B-Instruct as a local planning backbone that preserves the structured reasoning, tool invocation, and summary protocol of the API-based agent while reducing dependence on proprietary backbones. Experiments validate the API-based agent on established T2I/T2V benchmarks and open-ended queries, and evaluate Open-EA on four in-domain and three out-of-domain T2V generator families, showing partial cross-family transfer of the learned policy.

Reinforcement Learning with Verifiable Rewards (RLVR) makes Multimodal Large Language Models more accurate, but the gains are brittle: simply paraphrasing a question or changing the prompt template can degrade them, which challenges reliable deployment in high-stakes scenarios like medical VQA. We trace this to two issues of the standard RL objective. First, the binary verifier conflates format with content, so the reward signal cannot tell a wrong answer apart from a misformatted one. Second, the training distribution covers only a thin slice of the real-world prompts that the model might meet at deployment, so policies that perform well on the training distribution can behave differently under unseen prompts during test. Both failures call for a robust post-training method that helps the policy cover a broader distribution of semantically equivalent prompts, and we identify two measures that help achieve this objective: separating format from semantics in the reward, and applying policy invariance across perturbed prompts with equivalent semantics. We therefore propose Prompt-Invariant RLVR (PIRL), consisting of a dynamic trinary reward and a consistency regularizer based on an embedding-space adversary. Under stress testing, PIRL's average accuracy on benchmarks drops by only $\le 1%$, where GRPO drops ~3%. On dynamic evaluation, PIRL also achieves the smallest performance drop.

Autonomous coding agents powered by large language models can now generate code patches directly from bug reports, but a fundamental gap remains: once a patch is produced, no mechanism independently verifies whether it truly resolves the reported problem. Prior work has sought to address this through iterative self-refinement and inference-time scaling, but these approaches either review the patch under the same interpretation that produced it or broaden candidate generation without verifying individual patches, and neither provides an explicit verification signal for assessing patch correctness. We propose RETRACE, a training-free post-generation verification framework that derives such a signal through bidirectional reconstruction and reconciliation. When a coding agent generates a candidate patch for an issue, RETRACE performs forward reconstruction to build an explicit repair rationale from the issue and the agent's trajectory; backward reconstruction then independently infers, from the patch and its trajectory alone and without access to the original issue, a description of the problem the patch appears to address, and compares this reconstruction against the original issue to produce an alignment verdict; a reconciliation stage then checks the consistency between the forward rationale and the patch, diagnoses the source of any misalignment, and either submits the patch or produces targeted revision guidance. Evaluated on SWE-bench Verified with two backbones (GPT-5-mini and MiniMax-2.5), RETRACE lifts Pass@1 by 7.0% and 3.6% respectively on the mini-SWE-agent scaffold, and delivers comparable gains on OpenHands without modification. Ablation experiments show that both the forward and backward stages contribute to the overall improvement and that adding reconciliation yields further gains.