Generative RL 进展

Concept unlearning in text-to-image diffusion models is critical for safe and practical deployment: with rising privacy concerns, copyright disputes, trademark constraints, and safety regulations, deployed systems must be able to suppress unwanted concepts after training. Existing methods often remove the target concept effectively, but practical unlearning also requires an equally fundamental property: the unlearned model should retain quality, diversity, and semantic coverage on benign generation. The gold standard is a retain-only model trained from scratch without the unwanted data. However, common erasure objectives do not specify which post-unlearning distribution should approximate this reference, leaving retention as an implicit consequence of the update rule. We propose TILDE, TILt-based Distributional Erasure, which formulates concept unlearning as a distributional alignment problem: the desired target is the minimum-deviation conditional distribution from the pretrained model under a forgetting constraint. This energy-tilted, anchor-free target suppresses concept-expressing images while preserving benign relative mass for each prompt. We instantiate this principle with residual $\nabla$-GFlowNet training, which learns the score correction induced by the forget energy relative to the pretrained diffusion model. Across objects, artistic styles, and characters, TILDE achieves strong forgetting while improving retention and distributional fidelity over prior baselines.

Video face restoration (VFR) aims to recover high-quality and temporally consistent facial details from severely degraded video sequences; however, existing methods still struggle to balance spatial fidelity and temporal coherence under complex degradations. To address this, we propose FADRA, a frequency-aware diffusion framework with iterative residual adaptation specifically tailored for robust VFR. We first leverage the strong temporal consistency of a pre-trained text-to-video diffusion model and introduce lightweight LoRA adapters together with a Low-Quality (LQ) Pixel-Alignment Feature Fusion module to efficiently adapt the frozen generative prior to the VFR task. To further adapt the frozen diffusion backbone to the downstream VFR task beyond LoRA-based adaptation, we introduce a Repeated Residual Adaptation Head (RRAH) for step-wise residual refinement after the diffusion backbone. To make this refinement explicitly guided by the degraded observation, RRAH further takes the LQ latent together with the current velocity prediction as input, allowing the model to repeatedly revisit LQ cues and predict residual updates at each flow-matching step. This LQ-guided repeated residual adaptation helps recover fine facial details while preserving the inherent temporal priors of the pre-trained model. Furthermore, to ensure the structural integrity of perceptually important details, we introduce a Frequency-Aware Loss that provides explicit supervision across multiple spectral bands, emphasizing visually sensitive frequency components that are crucial for perceptual quality and prone to temporal jittering. Extensive experiments demonstrate that FADRA recovers better facial structures and produces more temporally consistent videos than state-of-the-art methods, leading to clear gains in both quantitative metrics and visual perception.

Text-to-image diffusion models have achieved high visual fidelity and broad adoption, but remain vulnerable to safety violations when adversaries exploit them to synthesize illicit content. Existing alignment paradigms, from input sanitization to structural feature pruning, are largely organized around unsafe concepts explicitly exposed during filtering, editing, or localization. This leaves a blind spot for visual synonym attacks (VSA), a jailbreak where benign-looking prompts elicit prohibited imagery through implicit visual associations. As a result, current defenses face a safety-utility dilemma: they may either under-mitigate VSA threats or over-suppress visually similar benign concepts. The core challenge is that VSA hides the unsafe target at the textual surface while revealing it through generation-time visual-semantic convergence. In this work, we therefore shift from static suppression of pre-specified unsafe concepts to dynamic tracing of how unsafe semantics emerge during generation. Our mechanistic analysis shows that VSA and explicit unsafe prompts converge through sparse semantic-injecting attention heads, which serve as inference-time bottlenecks for prohibited visual semantics. Based on this insight, we propose AEGIS (Adaptive Evasion Guard via Identification and Steering), an inference-time defense that applies similarity-aware repulsion only at the identified vulnerable heads. Evaluated against 16 baselines, AEGIS improves both safety and utility. On SD 1.4, it reduces ASR to $\mathbf{0.00}/\mathbf{0.03}$ for in-domain violence/nudity VSA and achieves ASRs $\le \mathbf{0.09}$ on out-of-domain explicit and adversarial attacks. It preserves benign fidelity, avoids suppressing hard-negative concepts, and transfers to SD 2.1 and FLUX.1 after re-identifying the critical heads for each backbone.

High-integrity systems, such as autonomous vehicle fleets and large-scale energy infrastructures, rely on structured assurance cases to justify safety claims. To remain valid under evolving operational conditions, such cases must be examined against potential challenges, known as defeaters. While large language models (LLMs) can support the scalable generation of candidate defeaters, assessing their quality remains largely manual and subjective process. This paper presents an automated approach for supporting the assessment of LLM-generated defeaters using natural language processing techniques. The method combines structural features from assurance case graphs with semantic embeddings and meta-classifiers trained on expert-assessed defeater annotations. We evaluate the approach through two case studies in the automotive and energy domains. The results show substantial human reviewer dissensus, with Cohen's kappa values below 0.442, highlighting the difficulty of consistent manual assessment. Against this background, the proposed classifiers achieve an average F1-score of 0.84 in validation and show improved alignment with individual expert ratings. The findings suggest that automated assessment can help reduce subjective variance and provide scalable decision support for assurance case review, while leaving final judgment to domain experts.

Preserving and revitalising traditional textiles such as Ulos, a cultural heritage of the Batak ethnic group in North Sumatra, Indonesia, requires balancing fidelity to tradition with innovative approaches that meet contemporary design demands. Traditional Ulos weaving faces two key limitations: a narrow range of motifs and a time-intensive design process. This study presents a generative AI framework that fine-tunes two pretrained latent diffusion models: Protogen v3.4 and Stable Diffusion v1.4, on a curated, annotated dataset of high-resolution Ulos motifs to generate culturally consistent yet novel designs. Model performance is evaluated quantitatively using Frechet Inception Distance (FID), Inception Score (IS), and qualitatively through assessments by traditional weavers and members of the public. Protogen v3.4 consistently outperforms Stable Diffusion v1.4, achieving substantially lower FID (~10.5x) and higher IS (2.0x), indicating superior visual fidelity, diversity, and closer alignment with the real Ulos motif distribution. We further examine the effects of strength and guidance scale on generation quality across both models. Lower strength values consistently yield higher fidelity (lower FID), while higher strength values increase generative diversity at the cost of realism, revealing a clear fidelity-diversity tradeoff for both models. Across all tested configurations, a guidance scale of 5-9 provides the most effective balance between fidelity and diversity, stabilising FID, KID, and IS, and is recommended as the operating range for high-quality, diverse Ulos motif generation. These findings demonstrate that carefully fine-tuned generative AI can support the creative renewal of intangible cultural heritage while preserving its stylistic and symbolic integrity.

Reinforcement Learning is commonly used to train large language models using environmental feedback. In applied settings, the environment usually provides sparse or delayed feedback. This makes it difficult for the model to pinpoint which actions in its reasoning led to success or failure. So, learning effectively from these signals is hard because the model must determine how each failure should inform meaningful behavioral corrections in subsequent iterations. We introduce a training framework, Self-Review Reinforcement Learning, that embeds an explicit self-review step into each RL episode. When a first-pass response fails, the model generates a self-review to identify what went wrong, which conditions an improved second attempt. Unlike inference-time reflection approaches, such as Reflexion, the framework optimizes self-review with policy gradients and internalizes improvements into the base policy via selective distillation, ensuring they persist across future episodes. A cross-episode memory keeps successful self-reviews for reuse when encountering similar tasks in future episodes during training. We evaluate SRRL against a standard RLVR baseline using the GRPO optimizer across two language models, Qwen 3-4B and OLMo-3- 7B, on GSM8K benchmark. SRRL consistently outperforms the RLVR in final reward performance and achieves greater learning efficiency by successfully transforming feedback into behavioral improvement.

Reinforcement learning (RL) post-training for large language models (LLMs) follows a efficient paradigm of "rollout then update", which inevitably results in off-policy training data. To resolve this, Importance sampling (IS) is proposed, while the token-level ratios compound over long sequences, causing severe variance exploded. A natural idea is "transferring" these off-policy token into on-policy token, so that the importance scores for correction are unnecessary. Following this idea, we propose Selective Importance Sampling (SIS), which is inspired by rejection sampling. Concretely, SIS implements by viewing off-policy model as proposal distribution, and implement a token-level rejection test: accepted tokens are viewed as on-policy, so that receive unit importance score, while rejected tokens retain the standard IS correction. Our proposed SIS is theoretically proved reducing the gap between token-level and sequence-level off-policy gradient estimators. The SIS acts as a plug-in that only modifies the importance ratio in the policy loss, adding negligible wall-clock overhead, and can be combine with a vast vary of RL post-training algorithms. Experiments on dense and MoE LLMs across math and agent benchmarks show that SIS consistently improves all objectives, while providing substantially stronger robustness under off-policy data.

Large Language Models (LLMs) can produce detailed answers to complex queries, but these answers are typically presented as dense linear text, which makes fine-grained inspection, navigation, and return visits difficult. We present ChatImage, a system that converts long-form LLM answers into interactive visual images. Given a textual answer, ChatImage first normalizes its content into structured visual modules, plans a visual layout, and renders a coherent image. It then applies a second grounding pass to the rendered image with vision grounding models such as LocateAnything and MiMo-Vision, with optional SAM-style mask refinement, to identify the visible regions that should support interaction. From these grounded regions, ChatImage overlays transparent clickable hotspots on the image. Each hotspot opens a detail panel and a region-scoped follow-up thread, allowing the user to inspect and query a specific part of the answer without re-reading the full response. Instead of treating planned coordinates as the final interaction geometry, ChatImage uses them as priors and grounds the interaction targets after rendering, which improves consistency between visual content and clickable regions. We release a reference implementation and introduce a 30-question benchmark covering infographic, map, and scene-based answer formats. Evaluation with configured external models reports interaction-loop completion, a strict visual-alignment gate, and a SAM-based mask-completeness diagnostic.

Text-to-image (T2I) diffusion models have achieved striking progress but still struggle to synthesize rare concepts involving unusual attribute-object pairings, often resulting in concept omission or semantic drift where a dominant entity overwhelms the generation. Tracing these failures to a lack of compositional balance during the denoising trajectory, we propose RADIANCE, a training-free framework that treats inference as a closed-loop feedback process. RADIANCE augments pretrained backbones with three modular components: (1) a Compositional Similarity Monitor (CSM) that tracks the emergence of objects and attributes in intermediate latents via CLIP-based feedback; (2) a Bidirectional Scale Controller (BSC) that applies a reactive "restoring force" using positive and negative IP-Adapter scales to rebalance biased trajectories; and (3) a Feedback Guidance Scheduler (FGS) that coordinates these updates across timesteps without additional training. We further extend the framework to multi-object prompts via Delayed Adapter Activation (DAA) and Layer-wise Alternating Guidance (LAG) to prevent premature concept fusion. By overlapping monitoring and denoising through pipelined execution, RADIANCE maintains competitive latency while significantly enhancing the per-sample success rate and effective throughput. Experiments on RareBench and T2I-CompBench demonstrate that RADIANCE consistently enhances compositional alignment and perceptual quality over state-of-the-art baselines.

Reinforcement learning holds significant potential for training large language models (LLMs) to handle multi-turn interactive tasks. However, in long-horizon, multi-turn tasks characterized by sparse outcome rewards, directly training with outcome rewards often results in slow convergence due to the sparsity of signals and the lack of fine-grained feedback. Furthermore, the model may fail to learn successful trajectories that are not sampled during training, thereby limiting its performance. Conversely, while employing customized dense process rewards provides richer signals and accelerates convergence, these surrogate rewards may exhibit potential misalignment with the ground-truth outcome rewards. This inconsistency can bias the training direction and ultimately degrade the model's final performance. In this work, we propose Reward-Swap Policy Optimization (RSPO), a method designed to leverage the rich information from dense process rewards to facilitate training with outcome rewards. By utilizing a reward-swap mechanism, RSPO ensures the diversity of sampled trajectories while guaranteeing consistency between the optimization objective and the true outcome rewards, thereby elevating the performance ceiling of the model. We conduct extensive experiments on two challenging agent benchmarks, WebShop and ALFWorld. By applying our method to various reinforcement learning algorithms, including GRPO, PPO, and GiGPO, we demonstrate that RSPO achieves consistent performance improvements across different baselines and benchmarks.

In this paper, we introduce Claim-Level Rubric Rewards (CuRe), a structured reward framework designed to address the reward-design bottleneck in reinforcement learning for dense video captioning. Existing reward designs generally fall into two categories: holistic response-level judgment across heterogeneous criteria, or alignment-based evaluation against reference captions. However, both paradigms suffer from fundamental limitations. Holistic rewards struggle to ensure factual accuracy and are prone to stylistic reward hacking, while reference-based rewards overly rely on rigid textual alignment, failing to preserve the completeness and diversity inherent to open-ended generation tasks. To address these challenges, CuRe reformulates reward modeling as fine-grained claim-level verification. Specifically, CuRe decomposes captions into category-aware atomic claims through a structured rubric, converting holistic evaluation into simpler and more reliable claim-level verification.

Pairwise human comparisons are a primary interface through which modern AI systems learn human preferences. RLHF and related alignment pipelines typically model such comparisons with Bradley--Terry log-odds, where choice probabilities are governed by latent reward differences. This paper examines what this assumption misses through a reduced-form model motivated by rational inattention, in which each label is generated by a low-capacity evaluation channel. The model separates two forms of ambiguity that standard reward modeling tends to conflate: a comparison may be difficult because the two candidates are genuinely close in value, or because the relevant distinction is hard to detect under limited attention. We show that limited attention can fundamentally distort what pairwise comparisons reveal. In particular, passive comparison data cannot generally distinguish reward, attention, and default tendencies, and heterogeneous attention can make standard Bradley--Terry reward modeling recover misleading rankings. Our analysis shows that learning is governed not by the raw number of labels, but by the amount of attended information each label carries. A case study on human votes over language-model pairs from Chatbot Arena exhibits the predicted signature, a cyclic component of the comparison data that exceeds sampling noise and that no scalar reward can represent; a second case study on perceptual comparisons shows that response times and gaze carry gap information that the labels do not. This perspective suggests that human feedback should be treated not as direct revealed preference, but as an attention-limited measurement process: a weak preference signal may reflect hidden evaluation difficulty rather than genuine indifference.

Reinforcement Learning (RL) is the dominant paradigm for training Large Language Model (LLM) agents on long-horizon tasks. However, sparse and delayed rewards often lead to trajectory neglect, in which agents lose focus on the task goal and interaction history at intermediate steps. Prior work has explored step-level supervision using Shannon-entropy-based uncertainty signals, which conflate inherent state complexity with agent confidence and therefore provide unreliable estimates of decision reliability. To address this issue, we propose normalized entropy, which measures confidence deviations relative to an agent's average behavior under a given state, thereby strengthening the association between low-quality actions and trajectory neglect. Building on this insight, we introduce Selective Trajectory-Aware Policy Optimization (STAPO), a hierarchical group-based RL framework. STAPO leverages normalized entropy to locate outlier steps associated with trajectory neglect and optimizes them via a joint mechanism of trajectory-aware reward and trajectory-independent penalty, enhancing trajectory awareness while preserving training stability. Extensive experiments on ALFWorld, WebShop, and Search-Augmented QA demonstrate that STAPO achieves state-of-the-art performance while substantially alleviating trajectory neglect, validating its effectiveness and robustness for agentic tasks.

We study how correctness is assembled inside aligned language models, not only whether the final answer is right. Using layer-wise difference-in-differences (DiD) trajectories over polarity-controlled minimal pairs, we identify the wrong-dip: in mid layers (25-90% depth), internal preference transiently commits to the incorrect answer and is rescued only by late-layer correction. We verify this causally with patchscope-style activation transplantation across 17 models, three families, and 64x scale (0.5B-32B). Four findings follow. (1) Alignment amplification of the causal wrong-dip is recipe-specific and emergent: it emerges at 3B in Qwen2.5, remains high, and peaks at 32B (paired t up to 9.7), reverses in Llama-3-8B (t=-2.31), and sits between for Mistral-7B. (2) The dip predicts real compression failures: high-dip items are 3-7x more likely to flip under late-layer low-rank compression, block dropping, or structured pruning, while quantization flips are dip-blind, a double dissociation confirmed by late-layer ablation. (3) The dip is trainable: a LoRA fine-tune with a mid-layer wrong-margin penalty matches output-only SFT accuracy while cutting the causal dip by 67-70% and improving compression robustness; output-only SFT worsens the causal dip by up to 2.8x at perfect surface accuracy. (4) With controlled readouts, the phenomenon survives natural-language I/O: dip stratification of structural-damage failures is significant on naturalistic vignettes, and free-form fragility separates into a dip-auditable late-rescue layer and a dip-blind interface layer. Together, output-level correctness can hide a late-rescue process that governs compression risk, post-training quality, and evaluation distortion.

Physics reasoning fails structurally in small language models: an error at any step propagates forward, corrupting every inference that follows. Limited domain knowledge, hallucination under multi-step derivation, and distributional sensitivity compound this failure. We propose a step-level reward framework that identifies the first reasoning error, generates targeted structured feedback, and trains the model to revise its solution via policy gradient with KL regularization, without exposing it to ground truth solutions as generation targets. Unlike annotation-dependent step-level methods, no preference data construction is required and the external verifier operates exclusively at training time. Across five physics benchmarks, our framework delivers accuracy gains of 17-20% over CoT prompting and 10-16% over the strongest baseline, reduces calculation errors from 56.9% to 23.5%, and reduces miscomprehension errors from 22.3% to 12.0% in the best observed cases. Conceptual errors reduce from 89.7% to 68.7%, yet persist as the hardest failure mode across all conditions.

Large language models can generate plausible quantum code, but it is unclear whether they can reliably target the specific software development kit (SDK) version requested by the user. We study this problem as API drift and introduce quantum-api-drift, a benchmark for measuring version fidelity, defined here as execution success on the requested SDK version, cross-version compatibility, failure modes, and documentation-guided repair in LLM-generated quantum SDK code. We instantiate the benchmark with Qiskit, a representative quantum SDK that underwent substantial interface changes across v0.43, v1.3, and v2.0. We evaluate 17 models on 50 tasks with 3 samples per prompt, yielding 450 generated samples and 1,350 executions per model. Sixteen models are tested in a matched REST API setting with a 1024-token output cap, while GPT-5.4 (Codex CLI) is reported separately as a reference configuration. Across the 16 matched REST models, diagonal Pass@1 ranges from 0.02 to 0.85. Claude Opus 4.7 is strongest on v0.43 and v2.0, while Grok 4.20 is strongest on v1.3 at 0.513. Error profiles differ systematically by model strength: weaker models fail mainly with broken imports, while stronger models more often reach deprecation-level failures. Documentation-guided repair succeeds for 0.19 to 0.59 of repair attempts overall and is consistently much more effective for migration to v2.0 than to v1.3. The benchmark artifacts are publicly available at https://github.com/arasyi/quantum-api-drift. These results show that version alignment is a distinct evaluation axis for quantum code generation and that API drift remains only partly recoverable even with migration guidance.

Diffusion large language models (dLLMs) generate text by iteratively denoising a masked sequence, offering a parallel alternative to autoregressive models, but eliciting strong reasoning through post-training remains difficult: supervised fine-tuning is off-policy and suffers from exposure bias, while reinforcement learning gives only sparse, sequence-level rewards and is hard to apply without tractable sequence likelihoods. On-policy self-distillation (OPSD) offers a promising alternative, using one model as both student and teacher to provide dense, token-level, on-policy supervision, but its effectiveness hinges on giving the teacher privileged information (PI) - typically an instance-specific ground-truth reference unavailable at inference - so the student ends up distilling a weak PI-free consensus policy that yields little improvement on dLLM reasoning. We introduce dOPSD, which instead derives the teacher's privilege directly from the student's own denoising trajectory, evaluating masked positions using later, more-decoded steps of that same trajectory rather than an external label, so the teacher's advantage emerges from the model's own decoding process; on Dream and LLaDA, dOPSD improves both in-domain math reasoning and out-of-domain code generation, outperforming supervised and on-policy baselines.

Learning-based stereo matching models struggle in underwater environments due to scarce in-domain data and the difficulty of extracting discriminative correspondences from degraded imagery. In this work, we present $AquaStereo$, a perception-enhanced framework with a data simulation pipeline and a self-distillation strategy that jointly address data scarcity and feature degradation in underwater stereo matching. First, a depth-conditioned diffusion pipeline renders underwater stereo pairs while preserving binocular geometry, with a lightweight left-right consistency module ensuring geometric alignment. Training on this synthetic corpus effectively narrows the terrestrial-underwater gap and improves zero-shot robustness. Second, a frozen binocular teacher trained on clean terrestrial pairs guides a student exposed to rendered underwater pairs with perturbations. A stage-weighted sequence loss is performed to align the student's disparities with the teacher's geometry, while a clean-branch supervision with shared pseudo targets prevents scale drift. To further enhance feature stability under turbidity and low texture, we introduce learnable perception frames, a perception-enhanced feature formulation that constructs robust matching descriptors by fusing temporal cues from two auxiliary views encoded by a video backbone with semantic features extracted by a strong image encoder. Extensive experiments demonstrate that $AquaStereo$ substantially improves robustness and zero-shot generalization in challenging underwater scenarios. The code is available at https://github.com/qz-wei/AquaStereo.

Subject-driven and multi-element video generation are central to controllable video synthesis, but existing methods still struggle to preserve identity consistency and model complex relationships among multiple subjects. In this paper, we propose Aura, a unified framework for high-fidelity and identity-consistent video generation. To better capture scene dynamics and subject interactions, we introduce AI director-level captions that provide dense and structured descriptions of video content. We further leverage a vision-language model (VLM) with learnable queries to extract multimodal semantic features from textual and visual references, covering both global semantics and fine-grained visual cues. To bridge the representational gap between the VLM and the Diffusion Transformer (DiT), we design a two-stage alignment strategy that progressively maps VLM features into the DiT feature space. For visual conditioning, we adopt token concatenation to inject reference information directly into the generation process. To distinguish heterogeneous subject types and reduce common copy-paste artifacts, we develop a subject-aware RoPE-Shift mechanism. To further differentiate reference images of different categories, we introduce subject-aware learnable tokens. In addition, we introduce Memory Tokens to balance the training signal across examples with different numbers of reference subjects. During inference, Progressive-APG (Adaptive Prompt Guidance) further alleviates oversaturation and improves semantic alignment with user prompts. Finally, we build a high-quality video-subject image dataset through a dedicated data construction pipeline. Extensive experiments show that our method achieves state-of-the-art performance on both single-subject generation and more challenging multi-element scenarios.

Recent agentic approaches to LLM-based kernel generation have achieved impressive results on CUDA. For emerging AI accelerators such as AWS Trainium and Inferentia, automated kernel generation and optimization remain largely unaddressed. Writing kernels for these chips via the Neuron Kernel Interface (NKI) is particularly challenging: developers must navigate a multi-engine architecture, tile-based programming, and explicit data movement across multi-level memory hierarchy. Moreover, no publicly-available training data, benchmarks, or tool-augmented agents exist for this domain. We introduce NKI-Agent, the first system combining domain-specific supervised fine-tuning (SFT) with a compile-verify-fix agent loop for NKI kernel generation. We adapt the existing CUDA-Agent framework to Neuron hardware, curate 6,000 NKI kernel generation tasks for training, and construct NKIBench, a 250-task benchmark across three difficulty levels. Evaluated on real Trn1 hardware, NKI-Agent with Claude Opus 4.8 and a rank-aware system prompt achieves a 77.3% pass rate on the 150-task NKIBench. We show that tool use is critical: Opus 4.8 scores 6% in single-shot mode without agent tools. On a 60-task subset, we show that an SFT-trained Qwen3-Coder-30B-A3B achieves 25.0% pass rate at 1/100th the cost, outperforming Claude Sonnet 4 (15.0%). We also report that Group Relative Policy Optimization (GRPO) with binary compilation reward fails to improve over SFT, providing guidance on reward design for RL-based kernel generation.

Autoregressive (AR) text-to-speech (TTS) models generate discrete speech tokens sequentially, which makes inference slow and can degrade robustness by propagating local errors and hallucinations. This limitation stems from their left-to-right AR commitment: each token must be determined before future speech-token context is available. However, such ordering is not an inherent requirement for TTS, as the full input text is available before synthesis. In this paper, we introduce DELTA-TTS, a lightweight LoRA-based adaptation framework that converts a pretrained AR TTS model into a discrete diffusion language model (dLLM) for confidence-ordered speech-token decoding. To better capture the local structure of speech, DELTA-TTS incorporates a convolution module that injects local acoustic context, together with a $1/t$-weighted training objective and a time-shifted inference schedule that defer low-confidence positions to later steps. Trained on only $585$ hours of LibriTTS, DELTA-TTS achieves a $1.75%$ WER on Seed-TTS test-en, outperforming its AR backbone while generating tokens $3.3\times$ faster. Further analysis shows that DELTA-TTS produces sharper text--speech alignment, increases overall decoding confidence, and mitigates hallucinations observed in AR generation.

Language models can reach the right normative verdict for the wrong reason. We introduce NormWorlds-CF, a solver-verified environment for counterfactual normative reasoning in executable rule worlds. Its deterministic solver produces final answers, proof and falsification certificates, argument statuses, support sets, and paired-world change labels, enabling supervision and evaluation without LLM judges. The benchmark contains staged SFT diagnostics and a compact paired-world task with 270 root families and 1080 canonical-to-variant pairs. The SFT diagnostics show that final-answer supervision is an unsafe proxy: answer-only SFT reaches perfect accuracy on answer tasks but scores zero on falsification, while proof-plus-falsification training with targeted replay reaches strong all-task accuracy. For the structured-change task, we introduce metamorphic-relation GRPO (MR-GRPO), a class-conditioned reward for GRPO that gives partial credit for relation families and solver-visible change fields. In matched 1.7B continuation experiments, MR-GRPO improves held-out relation accuracy and relation-family correctness, and reduces wrong-family error, compared to sparse and answer-only GRPO. In Qwen3-4B three-seed validation, answer-only reward improves answer-change fields but weakens relation-family structure, sparse reward preserves coarse relation labels best, and MR-GRPO delivers the strongest balanced performance across answer-change, support-change, status-change, and soft root-level metamorphic-relation metrics. These results show that verified counterfactual structure can shape post-training beyond final answers, while exact full change-record generation, invariant subtype recognition, and out-of-distribution (OOD) transfer remain open problems.

Prior studies have demonstrated that diffusion classifiers achieve robust zero-shot classification performance. However, their effectiveness is strongly tied to the pretraining data distribution: they perform well in majority, high-density regions of the data manifold, but are significantly less accurate in minority, low-density regions. Although prior works on minority sampling have focused on generating more minority-like images, what minority sampling fundamentally enables beyond generation remains underexplored. In this paper, we reveal a direct relationship between minority sampling in generation and the perception capability of diffusion classifiers. Specifically, we show that enhancing minority sampling broadens the coverage of underrepresented regions on the data manifold, thereby improving diffusion-based recognition. To exploit this connection, we propose Self-Improving Diffusion Classifiers with Minority Preference Optimization (MiPO), which fine-tunes a pretrained diffusion model using minority preference rewards. Using only arbitrary caption data, MiPO generates candidate samples, rewards those that better cover minority regions, and optimizes the model with LoRA and Group Relative Policy Optimization, without additional image data, external foundation models, or external reward models. This enables stable, prompt-adaptive minority sampling and translates low-density generative coverage into improved zero-shot diffusion classification. To sum up, we show that diffusion classifier perception is biased toward majority regions, demonstrate that this bias can be alleviated through minority preference optimization, and evaluate MiPO on five standard datasets.

Achieving simultaneous preference alignment and distillation acceleration in video diffusion models remains an open challenge. Existing methods optimize the two objectives over mismatched representation spaces, where improving one objective often compromises the other. To overcome this, we propose Reward Lightning, a unified framework that aligns and accelerates a video diffusion model within a single shared representation. Its central principle is homology: both objectives are evaluated on identical latent features, which mitigates the gradient conflicts that arise when they are optimized over disjoint representations. As a foundational component, we first introduce a latent reward model (LRM) that scores videos directly in the latent space, without decoding back to the pixel space. Building on the LRM, homologous preference distillation (HPD) reuses this shared backbone to perform adversarial distillation and preference alignment jointly, yielding few-step generators that remain faithful and well aligned. Extensive experiments demonstrate that the LRM surpasses pixel-level and latent-level reward baselines by $11.0%$ and $14.7%$ in preference accuracy, and that Reward Lightning generates high-fidelity videos in merely $1$ to $4$ steps, improving the average VBench score by $2.1%$ while leading in text alignment, motion quality, and visual quality. Project page: https://reward-lightning.github.io.

Diffusion models have become a dominant paradigm for conditional image generation, yet existing approaches generally follow two directions: task-specific designs that can improve performance but limit generalization, and training-free loss guidance that compresses rich conditions into scalar objectives and applies stepwise guidance, leading to information bottlenecks and error accumulation along the sampling trajectory. Given the urgent need for an effective unified framework across diverse conditional image generation tasks, we propose Data Injection and Contrastive Trajectory Refinement (DICT), a training-free inference method that enhances conditional image generation without introducing task-dependent architectures. DICT introduces Data Injection, where noise-perturbed conditional signals are integrated into early denoising stages; by performing guided denoising on these injected signals, DICT adaptively selects and distills task-salient information from the raw condition, effectively preserving spatial richness and ensuring precise condition-to-generation alignment. Furthermore, DICT applies Contrastive Trajectory Refinement across adjacent denoising states, enabling pairwise comparisons that progressively improve sample quality. These designs keep inference simple while improving cross-task transfer under a unified diffusion formulation. Extensive experiments on conditional image generation tasks (e.g., style transfer, image super-resolution, and image deblurring) show consistent gains in fidelity and perceptual quality over representative task-specific and loss-guided baselines.

The growing demand for image-to-video creation on mobile devices has increasingly focused on cinematic motion effects like bullet time, dolly zoom, slow motion, etc. While Diffusion Transformers (DiTs) exhibit strong performance in video generation, their large parameter sizes and multi-step iterative denoising processes lead to substantial computational overhead, making efficient generation on mobile devices challenging. We propose CineMobile to bridge the gap. In particular, CineMobile adopts a three-fold optimization strategy: (1) leveraging a distillation-guided pruning approach to derive a compact yet efficient model that retains the essential video generation capabilities required for cinematic effects; (2) optimizing the compressed model into a 4-step generator via a combination of diffusion distillation and reinforcement learning; (3) employing a hybrid post-training quantization strategy to compress the model footprint to under 1 GB. Experimental results show that compared to the teacher model with the Wan 2.1 architecture, CineMobile achieves a 40x speedup in generation while maintaining comparable visual quality. Specifically, CineMobile generates 49-frame 480p videos with a per-step denoising latency of 0.6s on an NVIDIA H200 GPU and 20s on the MediaTek Dimensity 8400 Ultimate 5G platform, with a peak memory usage of 1.8 GB, demonstrating its practical applicability for mobile-based image-to-video creation.

Integrating external tools with Large Language Models (LLMs) has emerged as a promising paradigm for accomplishing complex tasks. Since LLMs still struggle to effectively manage large tool collections, researchers have begun exploring retrieval-based methods to pre-select the most relevant options, addressing input length and latency constraints. However, existing retrievers are often misaligned with tool-calling LLMs due to their separate training processes. This paper presents PORTS, a novel odds ratio preference optimization method for training retrievers aimed at tool selection. Using a perplexity-inspired preference signal from a frozen LLM, our approach fine-tunes a retriever to find helpful tools by optimizing the correlation between the selection probabilities and the downstream performances while jointly enforcing a contrastive semantic loss between documentation strings. The versatility of PORTS and its ability to significantly improve tool selection accuracy are demonstrated through extensive experiments on six datasets, two encoder models, and three LLMs with diverse prior knowledge. With low computational demands, our alignment process facilitates generalization to new queries and tools, proving valuable for practical applications with evolving toolsets.

Post-training of frontier language models is conducted on curated task suites, and inevitably leaves a distribution shift between training and deployment environments. This exposes developers to generalization failures, which are relatively poorly understood. To better understand such generalization failures, we believe the community should construct clean demonstrations under simplified conditions. To facilitate this, we propose a simple and flexible way to construct language models which fail to generalize in controllable ways when subsequently trained with Reinforcement Learning (RL) on a given distribution of training tasks. Our construction uses Supervised Fine-Tuning on a dataset of a mixture of transcripts corresponding to a collection of 'conditional policies', which can each independently be assigned certain behaviors on each different task distribution, to obtain a model that is then well approximated as a 'mixture of conditional policies.' We observe that RL training then selects for policies that obtain the highest reward on the training distribution. This can produce striking behaviors: in a controlled setting with two distributions containing identical questions prepended with two different 'trigger strings', RL training on either distribution actively degrades performance on the other to zero, even though the underlying task is identical. We also use our construction to illustrate two novel ways in which generalization may fail in future language models, corresponding to distribution shifts of task coverage and temporal context respectively. While our construction is deliberately simple and may not closely resemble 'natural' generalization failures, the resulting 'model organisms' are of interest for alignment stress-testing and generalization science, and can be used as existence proofs that training success and generalization can come apart in structured ways.

LLMs have made substantial progress on automated code generation from natural-language descriptions of desired behavior (intent). Most existing methods improve generated programs through execution-guided code refinement: they generate a candidate solution, execute it, and patch the implementation using feedback, while leaving the underlying specification unchanged. This workflow implicitly assumes that the LLM's understanding of the intent is already correct and complete. In practice, however, intents are often ambiguous or underspecified. As a result, even a capable model may produce a correct implementation of the wrong intent, making specification mismatch a central bottleneck. This paper presents BeSpec, a behavioral model-based approach to specification alignment. BeSpec treats the task description as partial evidence about the intended behavior of the correct program. It first builds an explicit behavioral model, which are checkable properties that valid outputs must satisfy. BeSpec then generates candidate programs, executes them on probe inputs, and compares their observed behavior with the predicted behaviors. When observed behavior does not match the predicted behaviors, BeSpec either refines the specification or rejects the candidate program. We evaluate BeSpec with three LLMs on four benchmarks: CodeContests, xCodeEval, APPS, and the contamination-free LiveCodeBench. Against nine baselines, BeSpec achieves the highest Pass@1 and average pass rate across all settings, improving average Pass@1 over the strongest baseline by 8.1%--25.3% relative across the three LLMs. A failure analysis shows that after alignment, most remaining errors stem from algorithmic difficulty rather than misunderstood specifications, while ablation studies confirm that each major component of BeSpec contributes positively.

Reinforcement Learning with Verifiable Rewards (RLVR) has emerged as a promising paradigm for improving mathematical reasoning in language models. Yet most RLVR work rewards only the final answer (outcome-based rewards), leaving the impact of step-level process supervision (process rewards) underexplored especially for small models that lack the capacity to self-correct under sparse feedback. We systematically compare five reward conditions applied to Qwen2.5-0.5B fine-tuned with Group Relative Policy Optimization (GRPO) on GSM8K: a no-RL baseline, process-only, outcome-only, and three hybrid weightings ($λ\in \{0.9, 0.5, 0.1\}$ process weight). Process-only supervision achieves 63.73% test accuracy versus 53.75% for outcome-only, a nearly 10-percentage point gap while yielding reasoning traces with higher step validity and lower deviation from ground-truth chain length. Hybrid rewards generally correlate positively with process weight, with one notable anomaly: the low-process / high-outcome configuration ($λ=0.1$) underperforms pure outcome supervision, suggesting conflicting optimization signals. Error analysis using GPT-4o as a judge reveals distinct failure mode distributions: process models generate structurally inconsistent but arithmetically grounded traces, while outcome models produce concise but derivation-error-prone chains. Our results demonstrate that reward granularity is a first-order design decision for RLVR, with process-level supervision substantially improving both accuracy and trace fidelity in small language models.