Generative RL 进展

LLM preference alignment aims to optimize models toward human preferences across diverse user instructions. Reinforcement learning has become a major post-training approach for this goal, but existing proxy rewards are often outcome-level, mainly evaluating the final response while providing limited guidance for the reasoning trajectory. This can make credit assignment coarse when multiple responses receive similar final scores, leaving trajectory-level preferences under-specified. To address this limitation, we propose Thinking Checklist Reward (TCR), a process-oriented reward for RL-based preference alignment. TCR converts preference pairs into sample-specific thinking checklists and uses them to evaluate whether the generated reasoning trace addresses the preference-implied considerations. To reduce overlap with outcome-level supervision, TCR further introduces an exponential moving average (EMA) residual formulation to isolate a complementary thinking surplus beyond what is predictable from the outcome reward. Experiments on five models from three model families show that TCR consistently improves alignment performance across diverse benchmarks, with ablations further validating the importance of EMA-based residual formulation and sample-specific checklist supervision.

Thermal-to-visible face translation presents fundamental challenges including geometric discontinuities, semantic attribute mismatches, and identity degradation. We propose MTVDiff, a novel multimodal latent diffusion framework that synergistically integrates depth and textual information to address these limitations while preserving identity characteristics. The MTVDiff framework presents three core technical contributions: (1) a Dual-Branch Cross-Attention Fusion (DBCAF) module for multi-scale thermal-depth feature extraction and fusion; (2) a Gated Text-to-Visual Feature Alignment mechanism for semantically-guided generation; and (3) Spatial Feature Transformations (SFT) for adaptive multimodal prior integration. Extensive experiments on the MCXFace and SpeakingFaces datasets demonstrate that our multimodal approach significantly outperforms existing GAN-based and diffusion-based approaches across multiple metrics, achieving substantial improvements in both image quality and face verification performance, with FID reductions of up to 48.3% and Rank-1 accuracy improvements of up to 8.9%. Our work provides a robust solution for face recognition systems operating under varying illumination conditions and advances the state-of-the-art in cross-spectral facial image translation through effective multimodal integration.

In RLHF pipelines, reward scoring blocks policy updates. Slow scoring bottlenecks the entire loop, since no update runs until every rollout gets a score. And yet most setups just default to PyTorch eager mode or torch.compile, no one checks if that's actually fastest. Scoring itself is small. Rollout generation eats far more of a typical RLHF step. But scoring and generation fight over the same CPU and GPU resources, so a faster scoring engine doesn't shrink step time on its own. It mainly frees up capacity generation can use instead. We built a native C++ inference engine on ONNX Runtime. First step: confirm correctness. Output matched the PyTorch reference to 5.7 x 10^-6 on CPU and 4.2 x 10^-3 on GPU, close enough to trust. Then we tested it against PyTorch eager mode, torch.compile, and FastAPI, on both CPU and GPU. CPU was decisive. Our engine beat every baseline, confidence intervals didn't even overlap. GPU gave a different view: we beat PyTorch and FastAPI, but torch.compile came out ahead. Further testing traced the speedup to ONNX Runtime itself, not C++ as a language. And batching strategy mattered more than either the language or the runtime choice, more than we expected. The results are from repeated, independent runs, since single runs just aren't reliable enough to trust.

With the rapid advancement of large language models (LLMs), modern systems not only possess strong foundational capabilities and extensive knowledge, but can also solve complex problems via long, multi-step reasoning. However, as reasoning traces become longer, LLMs may produce a substantial amount of hallucinated content during the reasoning process, which is often difficult to detect. In this work, we conduct a fine-grained analysis of hallucinations arising in LLM reasoning and find that the reasoning traces are particularly prone to Context-Sensitive Factual Hallucinations: cases where the model actually has the relevant knowledge, yet makes factual errors due to contextual interference during reasoning. To address this issue, we propose Step-level Self-Consistency Group Relative Policy Optimization (SSC-GRPO), which assigns step-level rewards to reasoning traces by computing self-consistency scores of individual steps across multiple rollouts. Compared with prior methods, SSC-GRPO achieves state-of-the-art performance on both mathematical reasoning benchmarks and hallucination leaderboards. Our results offer a new perspective for detecting and mitigating hallucinations in the reasoning process of large language models.

Language models trained with reinforcement learning may learn to optimize the grader's judgment rather than the intended objective. This "reward-seeking" is difficult to measure because a model that pursues the grader's judgment and one that pursues the intended objective behave identically whenever the grader rewards the intended behavior. We measure reward-seeking using Contrastive Synthetic Document Finetuning to change a model's beliefs about what the grader rewards, putting those beliefs in conflict with what users or developers want, and measuring the rate at which the model adopts each party's preferred behavior. Applied to intermediate checkpoints of a capabilities-focused OpenAI o3 RL run, without safety training, we find that these checkpoints often side with grader preferences over those of users or developers on coding and alignment tasks. This tendency to side with the grader trends upward throughout RL training. For example, in an environment that forces a choice between keeping a promise to a supervisor and breaking it to complete the task, a late capabilities-focused o3 checkpoint breaks the promise 87% of the time when SDF documents say the grader rewards task completion, versus 9% when they say it rewards honesty (a choice its chain-of-thought often makes explicit). An earlier checkpoint is far less sensitive (40% vs. 24%). Our method also generalizes to reward-hacking models. A model organism trained to reward-hack (gpt-oss-120b) is more than twice as sensitive to grader preferences as the unmodified model, with the mean behavioral shift in favor of the grader rising from 33% to 86%. These results indicate that RL can increase reward-seeking over the course of training, producing models that may act against their developers' intentions when they believe that doing so leads to higher reward.

Multi-turn medical consultation agents must decide what to ask, adapt to patient responses, and determine when the collected evidence is sufficient. However, coupled evaluation conflates the quality of the policy-elicited history with policy-specific terminal diagnosis generation: strong generation can compensate for a thin history, while weaker generation can obscure a rich one. We introduce MedDDC-Eval, a diagnosis-decoupled testbed that treats elicited history as the comparison object and holds the history-to-diagnosis mapping constant through a shared frozen reader. Across two held-out sources, a grounded interface and an auditable diagnosis-trajectory-efficiency (D/T/E) harness measure diagnostic usefulness, information acquisition, and efficiency. Directional semantic coverage followed by deterministic one-to-one assignment yields coherent precision-recall counts for open-ended items, with at most one credited match per prediction or reference. Holding histories fixed, changing only the diagnostic reader shifts diagnosis F1 by 2.2-19.0 points and reverses 18% and 36% of pairwise policy orderings on the Record and Dialogue splits. We further apply standard Group Relative Policy Optimization (GRPO) over interactive multi-turn rollouts to post-train Qwen3-32B using diagnosis-result and trajectory feedback. On the 100-case Record and 70-case Dialogue splits, the trained policy improves over its initialization by 9.7 and 4.6 total-score points; removing either primary signal lowers held-out joint performance. These results show that MedDDC-Eval supports controlled attribution, interpretable elicited-history measurement, and evaluation-guided evidence-acquisition policy development.

Zero-shot video moment retrieval aims to overcome the limitations of traditional approaches that require large-scale datasets annotated with text and its relevant temporal spans. Despite advances in pre-trained vision-language models and multimodal large language models, existing ZMR methods still heavily depend on query-to-video content similarity, making them vulnerable to modality and language-style gaps. These gaps lead to unreliable span proposals and unstable moment retrieval results. To address this issue, we propose Self-Similarity-based Moment Proposal and Scoring that instead exploits intrinsic relationships within videos, enabling robust span generation and scoring. By deriving self-similarity only from the video content, we circumvent the noisy and mismatched patterns of query-frame or query-caption similarities, thereby mitigating both modality and language-style gaps. Furthermore, we introduce a query-aware MLLM-based reasoning stage to further sharpen alignment between text and video. Extensive experiments demonstrate that Self-SiMS achieves state-of-the-art performance across ZMR benchmarks.

Reinforcement learning with verifiable rewards (RLVR) improves reasoning in large language models. Yet, typical RLVR approaches fail on difficult problems: when a model cannot generate any correct solutions, it receives zero learning signal. Providing privileged guidance during training, such as solution prefixes, can help overcome this learning cliff by steering the model towards {correct solutions with non-zero reward}. {We call these rollouts off-context: they are generated from a training prompt that contains privileged guidance, while the target objective is defined by the original prompt without that guidance.} {We introduce} Off-Context GRPO (OC-GRPO), a minimally modified variant of GRPO that uses guided rollouts but applies an importance-corrected objective to steer the update back toward the original unguided objective, avoiding the mismatch that destabilizes uncorrected guided training. Empirically, our algorithm achieves a 3.9% absolute improvement (13.8% relative gain) over vanilla GRPO on average across standard mathematical reasoning benchmarks with negligible additional cost.

Chain-of-thought (CoT) reasoning is widely used to improve both the performance and interpretability of large language models (LLMs), yet the generated reasoning may not faithfully support the final answer. We study this problem from a causal perspective, where a faithful CoT process should follow the chain $Z\rightarrow X\rightarrow Y$, with $Z$, $X$, and $Y$ denoting the instruction, reasoning chain, and final answer, respectively. In this process, the instruction should affect the answer only through the reasoning chain. However, conventional autoregressive LLMs condition answer generation on both the instruction and the CoT, which still allows a direct instruction-to-answer shortcut. To address this issue, we propose CASE, a framework that combines training-time causal alignment and inference-time structural enforcement. During training, CASE builds counterfactual-CoT, biased-instruction, and empty-instruction datasets, and applies selective-loss fine-tuning to strengthen CoT-to-answer dependence while suppressing instruction shortcuts. During inference, CASE masks direct attention from instruction tokens to answer tokens, preventing the model from bypassing the generated CoT. We provide an information-theoretic analysis showing how these components promote faithful chains. Experiments on three models and four benchmarks show that CASE achieves a 37% average per-setting relative improvement in overall CoT faithfulness over the strongest baselines, exhibits stronger cross-dataset faithfulness transfer, and maintains competitive average accuracy. Code is available at https://github.com/oddwang/CASE.

Large Language Models (LLMs) excel at multi-step reasoning, yet current parallel reasoning approaches often fail to distinguish the contributions of individual reasoning paths. Many paths may be redundant, misleading, or even detrimental, but outcome-level rewards assign uniform reward, leading to ambiguous learning signals and unstable training. We propose Parallel Shapley, a reinforcement learning framework that attributes fine-grained, path-level contributions in multi-path reasoning. Treating each path as a player in a cooperative game, we leverage Shapley values to quantify marginal contributions, using a generative reward model to evaluate path utilities and Monte Carlo sampling for efficient approximation. Experiments on mathematical reasoning benchmarks show that Parallel Shapley outperforms existing baselines while providing more stable and interpretable training. Our framework effectively "fishes out the free riders," assigning reward proportionally and improving multi-path reasoning in LLMs.

Leveraging large language models (LLMs) for molecular generation has shown remarkable potential in chemical and drug design. Current methods primarily rely on supervised training or fine-tuning with limited datasets, which are insufficient to capture complex molecular design objectives. While some approaches attempt to guide generation toward specific goals, they often lack direct optimization mechanisms, making it difficult to align generated molecules with desired properties. To tackle these challenges, we propose LLMol, a principled reinforcement learning framework that directly incorporates verifiable rewards for targeted molecule generation. The key insight is to formulate molecular design as a goal-conditioned sequence prediction task, where verifiable rewards serve as explicit supervision to drive generation toward desired objectives. LLMol follows a two-stage training paradigm combining supervised learning and reinforcement learning. In the first stage, large language models are supervised fine-tuned to capture chemical syntax and molecular distributions. In the second stage, we introduce Reinforcement Learning with Verifiable Rewards (RLVR), which directly integrates property-based reward signals to guide molecular generation toward task-specific objectives. To address the high variance and instability common in discrete sequence optimization, we adopt Group Relative Policy Optimization (GRPO), a stable on-policy algorithm that smooths reward signals and improves training robustness. This framework enables LLMol to effectively handle a range of molecular design tasks, including single-property targeting (e.g., penalized logP, QED) and structure-constrained optimization. Experimental results demonstrate that LLMol consistently outperforms existing methods, achieving higher success rates and improved efficiency across diverse molecular benchmarks.

Large language models (LLMs) have driven rapid progress in electronic design automation (EDA), yet their application to radio-frequency (RF) circuit design remains limited by the scarcity of domain-specific datasets and standardized benchmarks. We present RF-Agent, which addresses this gap through textbook-driven knowledge distillation. A multi-agent Question-Thinking-Solution-Answer (QTSA) pipeline converts a subsection-level corpus from seven canonical RF textbooks into the first-of-its-kind RF-domain reasoning dataset (over 11,000 samples) with a dedicated multiple-choice benchmark. On this benchmark we study two adaptation strategies: supervised fine-tuning (SFT) and three retrieval-augmented generation (RAG) configurations (semantic, keyword, hybrid). Across multiple LLM families, domain-specific SFT significantly improves RF reasoning, especially for small and medium-sized models; among RAG configurations, semantic retrieval performs best, indicating embedding-based context alignment suits RF reasoning better than naive fusion. The dataset and benchmark provide a reusable foundation for future work on LLM-aided RF circuit design.

Group Relative Policy Optimization (GRPO) has shown strong effectiveness in reinforcement learning from verifiable feedback, where sampled rollouts can be compared within a group using task-provided correctness signals. However, extending group-relative optimization beyond verifiable settings is challenging because success in many tasks is not captured by a single correctness criterion. We propose Reference-Relative Policy Optimization (RRPO), which generalizes GRPO by replacing direct correctness-based advantage construction with reference-relative contrastive comparisons. RRPO first uses stratified conditional rollouts to construct positive and negative anchor sets, and then trains a metric projection head with a set-contrastive objective to compare candidate rollouts against these anchors. The resulting alignment scores directly define contrastive advantages: during policy optimization, the projection head is frozen, and the scores are centered within each rollout group in a standard group-relative objective. We evaluate RRPO using anchor-based contrastive advantages throughout policy optimization, without relying on task ground-truth verifiers. Across verifiable reasoning, open-ended generation, and post-SFT settings, RRPO remains competitive with verifier-based optimization, improves over weakly supervised baselines, and provides additional gains after supervised fine-tuning.

Structural fidelity is essential to scientific methodology diagrams. To communicate research logic, these diagrams must faithfully render components, directional relations, and textual annotations. Since a single error, such as a reversed arrow or an unreadable equation, can invalidate the entire figure, structural fidelity is inherently conjunctive: correctness on one axis cannot compensate for failure on another. Current open-source models fail to satisfy this criterion. Supervised fine-tuning (SFT) learns plausible layouts but cannot reliably ensure structural correctness, while scalar reward-based post-training obscures which structural dimension has failed. To address this, we introduce SciForma, a framework for the structure faithful generation of scientific methodology diagrams. Specifically, SciForma decomposes diagram quality into three structural axes: Component, Arrow, and Text, guided by a structural inventory. Built on this foundation, we curate SciFormaData-700K for structured training and SciFormaBench-2K for logic-verified evaluation. To close the gap left by SFT, we develop Multi-Dimensional Conjunctive Preference Optimization (M-DPO), which enforces simultaneous correctness across all axes and adaptively routes gradients to the most deficient dimension in post-training. The same structural inventory also enables iterative editing at inference time to correct residual errors. This combination allows SciForma-9B to exceed all open-source baselines and GPT-Image-1.5 on both SciFormaBench-2K and AIBench, bringing open scientific diagram generation close to proprietary-level structural fidelity. Our code and data will be available at: https://github.com/microsoft/SciForma.

Large Language Models (LLMs) often struggle to navigate value conflicts when trained with the compressed scalar rewards of Reinforcement Learning from Human Feedback (RLHF). To address this challenge, we investigate how chain-of-thought (CoT) reasoning can help improve performance in this domain. Geometrically, we show that CoT correlates with further smoothing the model's loss landscape in its sharpest direction, helping resolve the optimization instability of traditional scalar rewards. We also demonstrate via relevant downstream benchmarks that value conflict-focused CoT may generalize to different kinds of moral reasoning, demonstrating that this CoT has the potential to be an effective mechanism for better moral reasoning. To capitalize on this potential, we create a new value conflict-focused CoT design that further smooths the sharpest direction of the loss landscape and increases moral reasoning performance. This finding shows that explicitly modifying and improving the design of reasoning dynamics offers a promising avenue for improving model performance on user requests with complex value conflicts, advancing pluralistic alignment in LLMs.

PPO and the GRPO baseline studied here use clipped surrogate objectives whose favorable-direction saturation introduces an abrupt change in the scalar objective's derivative. We ask whether Output Reset (OR), a smooth one-sided saturation rule, offers a useful alternative for large language model post-training. PPO-OR and GRPO-OR replace the clipped policy term with an OR squared-margin loss in rollout-relative token log-ratio space; the advantage sign determines the update direction, and a token contributes zero direct OR residual after crossing the favorable margin. We compare PPO-clip with PPO-OR under generalized advantage estimation (GAE), and GRPO with GRPO-OR under group-relative advantages, using Llama-3.2-1B-Instruct on Anthropic hh-rlhf with one shared reward model and three seeds per method. Under GAE, PPO-OR has a mean final training-time reward-model score $0.305$ higher than PPO-clip, with a larger observed across-seed spread. Under group-relative advantages, GRPO-OR does not have a higher mean score, but shows a smaller observed spread, a near-zero terminal OR residual, and a declining overshoot fraction, while the matched GRPO clipped-objective trace remains variable. Both group-relative methods exhibit substantially larger rollout-to-current log-ratio displacement than the GAE methods, and OR does not consistently reduce it. Thus, OR changes optimization behavior in both matched comparisons, but the observed reward effect differs between them. At $G=2$, the GRPO-OR diagnostics do not translate into a reward-score gain. Whether larger groups change this outcome remains open. The reported scores are training-time reward-model measurements, not held-out human-preference performance.

Human-object centric video personalization (HOCVP) is a core task within subject-driven video generation. However, existing methods suffer from two key limitations. First, most approaches focusing on inter-subject personalization still struggle to strike a balance between high subject fidelity and accurate interaction patterns between humans and diverse objects, especially when objects represent abstract concepts such as logos. Second, while intra-subject references (e.g., OCR maps, multi-view inputs) are expected to enhance subject fidelity, most existing works lack mechanisms to understand such latent correspondence. To address both challenges, we propose HOMIE, an HOCVP framework that tackles both inter- and intra-subject input settings in a unified manner. Compared to previous approaches, HOMIE proposes a better MLLM integration strategy to extract knowledge of reference-level relationships without compromising the controllability of text encoders or incurring costly re-alignment. Specifically, we introduce global multimodal guidance within self-attention to better align MLLM-derived semantic features with VAE tokens. Furthermore, we propose modality-reference embedding to differentiate tokens from MLLM features and VAE tokens and associate intra-subject reference image tokens. Extensive experiments validate that our method achieves state-of-the-art performance across various HOCVP tasks. Project Page: https://yiyangcai.github.io/homie-page.github.io/

Zero-shot text-guided editing of real-world music recordings requires balancing semantic modification with faithful preservation of the original musical structure. Although recent diffusion transformers trained with rectified flow have achieved remarkable success in text-to-music generation, extending them to edit existing recordings remains challenging because editing requires accurate deterministic inversion, reliable structural preservation, and numerically stable integration throughout the inversion and generation processes. We present FlowSonic, a zero-shot music editing framework built upon a pretrained diffusion transformer trained with rectified flow. FlowSonic first deterministically inverts a real-world recording into the latent space and preserves its musical structure during editing by reusing cross-attention representations extracted during inversion. To improve the numerical reliability of inversion-based editing, we introduce a high-order ODE solver and systematically investigate how different numerical integration schemes influence trajectory stability, structural preservation, and semantic controllability. Comprehensive experiments on timbre-transfer and genre-modification tasks demonstrate that FlowSonic consistently outperforms existing music editing methods across semantic alignment, harmonic preservation, structural consistency, and perceptual audio quality. We further provide geometric and empirical analyses showing how the proposed numerical integration strategy improves latent trajectory stability and leads to more reliable music editing.

The prevailing inference framework for diffusion models formulates generation fundamentally as a problem of numerical integration. This perspective casts the model as an exact estimator, neglecting the inherent statistical uncertainty of the denoising process. In this work, we propose Forward-Process Aligned Diffusion prediction (DiFA), a training-free framework that reframes inference-time data prediction refinement as a sequential state estimation problem. Rather than reusing past outputs solely for numerical integration, DiFA treats iterative data predictions along the reverse trajectory as correlated observations to build a forward-aligned temporal consensus. Inspired by Kalman filtering, this consensus aggregates historical predictions according to structural consistency and noise-level compatibility. To counteract the over-smoothing tendency of temporal consensus, we introduce a deviation guidance mechanism to adaptively preserve residual details. Empirically, DiFA yields significant improvements on CIFAR-10 and ImageNet across the evaluated metrics, including FID, IS, and FD-DINOv2, demonstrating that aligning inference with the forward statistical structure substantially improves generative fidelity.

Automated fact-checking remains a challenge for Large Language Models (LLMs) due to "query brittleness" in traditional retrieval systems. We propose DeLIVeR (Decomposed Learning for Information-grounded Veracity Recognition), a framework that treats evidence retrieval as a reinforced strategic exploration task. DeLIVeR utilizes a Planner LLM to decompose complex claims into targeted question sets, which are used to traverse structured Knowledge Graphs (KGs) for high-precision evidence. We optimize the Planner's policy using Group Relative Policy Optimization (GRPO) with a reward system prioritizing structural diversity and verdict accuracy. Our evaluation on LIAR, FEVER, and PolitiFact shows that DeLIVeR significantly outperforms state-of-the-art baselines. Using Qwen2.5-7B, our framework achieved peak F1-scores of 83.73, 84.57, and 79.70 respectively, representing a 10-15% improvement over HippoRAG2. By shifting to a reinforced question-planning strategy, DeLIVeR effectively bridges multi-hop reasoning gaps and provides an auditable, transparent path for verifiable misinformation detection.

Human-like private chat requires more than fluent response generation: a system must preserve persona, relationship, memory, bounded knowledge, medium-specific timing, and a coherent multi-turn arc. We present AnthroDial, a closed-loop framework that formulates anthropomorphic dialogue as a joint problem of system architecture, executable evaluation, and diagnostic alignment. It combines (1) a role-conditioned scheduled dialogue runtime with persona and scenario cards, long-term memory, virtual time, and single-draft message decisions; (2) an executable benchmark with an L0 validity gate, five per-turn dimensions, and five dialogue-level dimensions; and (3) a post-training pipeline that filters 16,436 scheduled-decision examples for SFT and applies GRPO with a cognitive-diagnostic, ZPD-aware reward. The reward maintains Kalman-filtered capability estimates for each behavioral dimension, upweights dimensions with larger capability deficits, and uses rollout scores as task-level ZPD matches to focus optimization on learnable weak skills. On a benchmark with 55 personas, 50 scenarios, 50 persona-scenario bindings, and 100 role-conditioned cases per model, we evaluate 16 systems spanning frontier baselines, open models, thinking/no-think variants, and SFT/RL ablations. The strongest non-trained baseline reaches 32.00% strict ACC, while Qwen3.6-27B-SFT+RL reaches 39.00% strict ACC and a 98.5 overall score. In the 9B no-think setting, SFT and RL improve strict ACC from 0.00% to 13.00% and 18.37%. These results show that anthropomorphic dialogue benefits when generation, evaluation, and reward shaping share the same behavioral dimensions.

Auto-bidding plays an essential role in online advertising, automatically adjusting bids for advertisers to optimize their commercial goals. The emerging AI-Generated Bidding (AIGB) paradigm widely adopts generative modeling to optimize bidding strategies, yet suffers from the limited mode coverage of offline datasets and inadequate task-state understanding, hindering effective exploration of optimal strategies. Large Language Models (LLMs), with prior world knowledge and reasoning capabilities, offer a promising approach to overcome these limitations. However, directly applying LLMs to auto-bidding tasks faces inherent challenges in limited numerical precision, hallucinations, and inference latency. To address these limitations, we propose AIGB-R1, a hierarchical self-evolving auto-bidding framework aiming to enhance AI-Generated Bidding via LLMs' Reasoning capabilities, comprising a high-level Planner module for macro-level strategy planning and a low-level Executor module for fine-grained decision-making. Building upon this, we design an experience-driven self-evolving loop, enabling autonomous strategy exploration and optimization from accumulated experience. We adopt a two-stage pipeline of offline pre-training and post-training alignment, and build an interactive bidding simulation environment for strategy rollout. Furthermore, we propose Decoupled Group Relative Policy Optimization (D-GRPO) to achieve end-to-end optimization via advantage decoupling. Experimental results on a large-scale public dataset demonstrate the effectiveness of AIGB-R1.

Unified multimodal models (UMMs) aim to integrate visual understanding and generation within a single architecture, but architectural unification alone does not ensure semantic consistency. A model may describe the intended target correctly while generating an inconsistent edit. This exposes an understanding-generation alignment gap: linguistic and visual outputs live in different spaces, yet should be governed by the same target semantics. We study this gap in image editing, where an instruction defines a target state that can be both described and visually realized. Given a source image and an edit instruction, we compare a UMM's target caption with its edited image to test whether the two outputs converge on the same result. Our analysis shows that existing UMMs remain weakly aligned, especially for fine-grained entities, attributes, spatial relations, and local details, indicating that semantic unification is not achieved by architecture alone. To bridge this gap, we propose STBridge, a shared-target alignment framework that connects understanding and generation through a common target state. Here the target caption expresses the desired visual result, while the edited image realizes it visually, replacing separate task-specific paths with a shared information flow from target expression to target realization. STBridge follows an align-then-optimize strategy: supervised fine-tuning first establishes the shared-target channel, and sequential reinforcement learning further refines target-centered coordination. Across visual understanding, image generation, and image editing benchmarks, STBridge consistently improves over the initialization model. Alignment analysis confirms that STBridge narrows the gap between what the model describes and what it generates, demonstrating shared-target alignment as an effective post-training strategy for bridging understanding and generation in UMMs.

Value-aware AI systems require explicit computational representations of human values (groundings) and their aggregation into value systems in order to align their decisions with ours. As such representations are difficult to elicit, value learning seeks to infer them by observing human behaviour. This work addresses the lack of grounded value learning methods in generative AI: existing approaches typically replicate human preferences without awareness of the multidimensional structure of value alignment, or lack principled value system elicitation methods. To address these gaps, we adapt a previously validated value system learning method to the generative AI setting, which, based on pairwise prompt-response preference data, simultaneously learns: i) an implementation of a grounding for a set of values given by a multi-objective reward model, and ii) a value system representation in the form of a weighted linear scalarization of the previous grounding model. To ensure that the learned value systems are based on coherent value representations, our algorithm dynamically prioritizes the grounding learning process. We evaluate the method against baselines and a contemporary method on prompt-response preference datasets. Results show competitive performance and minimal trade-offs against the baselines, while improving explainability.

Text-to-video (T2V) generation models have achieved strong visual realism, but improving physical plausibility can come at the cost of semantic consistency with the input text. This tension arises because physical preference is typically determined by comparing dynamics between two videos, without accounting for whether either video faithfully depicts the scene specified by the prompt, making physical-semantic conflict a systematic tendency under this supervision paradigm. We formulate this challenge as a constrained preference optimization problem and propose Physical and Semantic Direct Preference Optimization (PSDPO), which modulates each preference pair's contribution based on the agreement between its physical and semantic signals. A gradient-level analysis shows that PSDPO bounds the semantic drift from conflicting pairs to a controllable residual, and further motivates a staged optimization protocol that provably reduces cumulative drift. The resulting method operates entirely within the standard DPO framework, requiring no auxiliary models or additional loss terms. Experiments show that PSDPO improves physical plausibility by up to $2\times$ over the baseline on VideoPhy-2, while maintaining strong semantic consistency on VBench, achieving a more reliable balance than existing preference-based methods.

Large vision-language models (LVLMs) have demonstrated remarkable capabilities in multimodal understanding. However, they remain prone to hallucinations, generating responses that are inconsistent with the visual evidence. Existing mitigation methods largely address language-prior bias or cross-modal imbalance, while progressive visual degradation across perception and memory remains underexplored. In this work, we propose Saliency-Driven Perceptual Realignment (SDPR), a training-free framework that mitigates the degradation of visual awareness throughout inference. Specifically, we first introduce saliency-driven attention redistribution to release attention hijacked by non-semantic sink tokens, thereby recovering critical visual evidence. Second, we identify spatial distortion in the KV cache and propose saliency-driven cache alignment to preserve query-relevant visual features during generation. Finally, we introduce prior-constrained contrastive decoding to penalize unfaithful predictions induced by dominant language priors. Our proposed SDPR is robust against hallucinations due to its holistic alignment of visual awareness across the entire generative trajectory. Extensive experiments across diverse LVLM architectures show that SDPR outperforms state-of-the-art methods on both hallucination and general-purpose benchmarks, requiring no additional training and incurring minimal runtime overhead. The code is available \href{https://github.com/PengSyuChen/SDPR}{\color{blue}{here}}.

Current benchmarks for language models primarily evaluate execution on fully specified tasks. However, real user tasks are often ambiguous. Users arrive with incomplete, exploratory, or even inconsistent goals, requiring the assistant to first determine the intended task before carrying it out. We study this problem as task alignment: the ability to align with a user on their intended task. We introduce a general framework for converting specified tasks into underspecified interactions, formalized as a POMDP in which the model must infer a latent task from partial and evolving user intent. We validate our user simulator post hoc with a human user study. Across shopping, coding, and professional work settings, we find that while models often perform well once the task is specified, models still struggle with task alignment: current models act prematurely, interact ineffectively, and fail to resolve ambiguous requests. Models on average recover the user's intended task only 22-32% of the time under ambiguity. In a human study in the same setting, humans reach 48%, outperforming all evaluated models. We show that post-training with supervised fine-tuning and reinforcement learning improves task alignment, but models still lag behind humans in resolving uncertainty through interaction. Together, our results suggest that current models still lack key interaction abilities required for reliable agency.

Small-scale language models (SLMs) are attractive for retrieval-augmented generation (RAG) in resource-constrained settings, but their limited capacity makes them highly sensitive to noisy or spurious retrieved evidence. Existing preference-based methods such as RoseRAG select only the hardest single preference pair via hard argmin/argmax, discarding the remaining signal; others treat multiple pairs as independent binary comparisons, resulting in low data utilization. We propose RIMS, a three-stage preference optimization framework comprising (1) synthetic chain-of-thought preference data generation via rejection sampling using the target SLM itself without relying on proprietary models, (2) a differentiable soft aggregation mechanism that replaces hard selection with a smooth operator, preserving gradient signal from all preference pairs while retaining the discriminative structure of margin-aware selection, and (3) preference optimization with the smoothed objective applied to multiple alignment algorithms. We theoretically show that the smoothed approximation admits a controllable error bound and that smooth aggregation yields provably tighter gradient alignment to the oracle objective than hard selection. Experiments on four multi-hop question answering benchmarks show that our approach outperforms state-of-the-art baselines across multiple SLM backbones, achieving consistent gains in Exact Match and F1 under noisy retrieval conditions. Our implementation is available at https://github.com/tptrix29/RIMS.

Unified Multimodal Large Language Models (MLLMs) offer a promising paradigm for unifying visual understanding and generation, yet they still struggle to follow complex spatial instructions and logical constraints in controllable image generation. To address this gap, we present ATLAS, a unified framework that equips MLLMs with a human-like "Think, Plan, and Paint" paradigm. We adopt layout as the shared representation that connects the three stages, enabling the model to reason about spatial requirements, plan explicit object arrangements, and render the final image. We further improve plan-to-image fidelity with reinforcement-learning-based layout alignment. We instantiate ATLAS at 7B and 80B scales, achieving state-of-the-art performance among MLLMs on image generation benchmarks and an average 65.31% improvement over existing layout-based unified MLLMs. On spatially related tasks, ATLAS obtains an average 23.06% gain over the base models. Through the same layout interface, ATLAS also supports instruction-guided editing and multimodal grounding. We further introduce ATLAS-Reasoning, a benchmark for evaluating generation under complex spatial instructions.

Reinforcement learning (RL) has become central to improving large language models (LLMs) on complex reasoning tasks, yet RL post-training is largely studied in isolation from the pretraining that precedes it. As a result, two basic questions remain open: (1) how do pretraining choices (model size, data) shape the returns to RL compute, and (2) what does RL actually do to the model? These questions are difficult to study in the standard LLM setting: pretraining corpora are vast and uncontrolled, making it hard to attribute behaviors to pretraining versus RL, and systematic compute sweeps across both stages are prohibitively expensive. To address these challenges, we use chess as a controlled testbed for studying reasoning across the full pretraining-to-post-training pipeline. We follow the standard LLM training pipeline by pretraining language models from 5M to 1B parameters on human chess games, supervised fine-tuning on synthetic reasoning traces, and running RL on chess puzzles with verifiable rewards. Using this framework, we find that the post-RL performance at given RL compute level is well-predicted from the pretraining loss, and slope of the RL reward curves improves approximately linearly with the pretraining tokens. Beyond scaling, we find that RL does not simply sharpen the SFT policy: on easy puzzles it amplifies correct moves the SFT policy already preferred, while on hard puzzles it surfaces correct moves that were nearly absent under SFT. We further test whether our findings transfer beyond chess by training a 1B language model on math-domain text, where the same predictive pattern emerges: longer-pretrained checkpoints reach higher post-RL performance and improve faster under RL. In sum, we provide a quantitative account of the pretraining-to-RL interface and a controlled testbed for studying the science of reasoning across the full pretraining-to-post-training pipeline.