Generative RL 进展

A central question in LLM reasoning is whether reinforcement learning (RL) instills genuinely new capabilities or merely reshapes how existing knowledge is expressed during inference. Building on the distribution-sharpening hypothesis, which holds that RL reallocates probability mass toward high-reward trajectories already latent in base models, we ask: can we unlock those latent paths without costly RL fine-tuning? We present Decision-Flow Sampling (DF-Sample), a training-free, data-free inference-time framework that constructs a hierarchical reasoning tree, scores terminal nodes for quality, and back-propagates utilities to inform each intermediate branching decision. Unlike conventional sampling strategies that make purely local step-wise choices, DF-Sample performs explicit global trajectory evaluation before committing to a path, recovering high-quality but low-probability reasoning chains that standard decoding overlooks. On GPQA, DF-Sample achieves 45.6% accuracy, surpassing power sampling (38.9%) and GRPO (39.9%), showing that a training-free method can outperform a trained one. Across three models and four benchmarks, DF-Sample consistently outperforms baselines, indicating substantial latent reasoning potential in pretrained base models.

Multimodal AI, powered by Large Language Models (LLMs) and Vision-Language Models (VLMs), is transforming assistive technologies by enabling simultaneous processing of visual and textual data. This advancement holds significant promise for over 43 million visually impaired and neuro-divergent individuals worldwide who face persistent challenges in navigating indoor and outdoor environments due to limited spatial awareness and insufficient environmental cues. Existing navigation aids often lack comprehensive 3D scene understanding, relying on constrained route-based strategies that hinder user autonomy. In this paper, we introduce a novel end-to-end framework that integrates LLMs, VLMs and digital twin technologies to deliver a spatially cognitive navigation support for visually impaired and neuro-divergent users. Our system captures video input via standard mobile phone cameras, and employs SLAM3R to generate dense 3D point clouds from monocular RGB sequences in real-time. Our custom post-processing algorithm ensures accurate point cloud alignment across multiple viewpoints without requiring predefined reference points. This enhances the capabilities of SpatialLM to produce structured 3D representations, including architectural elements and oriented object bounding boxes. The enriched spatial data is then processed by a locally deployed LLM, which interprets 3D contexts to generate detailed scene descriptions and precise distance measurements between users and surrounding objects. We evaluated our approach across diverse video scenarios featuring various perspectives, looped walking views and captured in multiple environments. The evaluation results demonstrate consistent accuracy in 3D scene interpretation and object localisation, underscoring the potential of our system as a transformative assistive navigation solution that combines advanced visual perception with spatial reasoning

Diffusion Language Models (DLMs) offer promising parallel generation capabilities but lag behind autoregressive models in complex reasoning and tool-use tasks. While Reinforcement Learning (RL) has recently been applied to enhance DLMs, standard RL approaches suffer from an exploration bottleneck. To address this, we inject reasoning priors from a stronger teacher model to guide RL exploration. In this paper, we introduce CanvasAnneal, a curriculum-guided diffusion RL framework. During the initial RL phase, we warm-start exploration by injecting teacher-generated reasoning traces into the initial diffusion canvas. As training progresses, we gradually remove this guidance and require the model to generate more of the reasoning trajectory independently. Across mathematical reasoning and tool-use benchmarks, CanvasAnneal improves over standard diffu-GRPO on MATH500, Countdown, and Tau2 and substantially accelerates reward improvement on several tasks, while gains are task-dependent. Our results suggest that structured training-time guidance can alleviate exploration bottlenecks in diffusion RL and speed up convergence on harder tasks.

While Reinforcement Learning from Human Feedback (RLHF) is the standard paradigm for aligning large language models with human preferences, its effectiveness in pluralistic settings has been called into question. Notably, recent work by Gölz et al. (2025) demonstrated that the distortion — defined as the multiplicative gap between the average user utility of the RLHF policy and the optimal average utility — can scale exponentially with the Bradley-Terry temperature parameter $β$ when users have heterogeneous preferences. In this work, we present a fine-grained analysis of the distortion of RLHF with reward clipping and demonstrate that such exponential degradation is not a fundamental property of the algorithm but rather a consequence of distribution mismatch between the distribution generating preference data ($μ$) and the KL reference policy ($π_{\mathrm{ref}}$). To this end, we establish tight upper and lower bounds on the distortion of RLHF across multiple regimes of the KL regularization strength. We show that in a representative regime, under the Bradley-Terry model, the distortion is $\tildeΘ(βB + β)$, where $B$ is an upper bound on the log density ratio between $μ$ and $π_{\mathrm{ref}}$. In particular, when there is no distribution mismatch (i.e., $μ= π_{\mathrm{ref}}$), RLHF achieves the optimal distortion of $O(β)$ up to a constant. Our results suggest that, to reasonably maximize average utility with RLHF, it is preferable to use on-policy sampled preference data or to fine-tune before RLHF on data from a source close to $μ$.

Large Language Model alignment typically relies on learned proxy reward models, which significantly increase the memory footprint during training and are notoriously prone to instability and reward hacking. While offline methods like Direct Preference Optimization (DPO) bypass the reward model, they lose the ability to perform online exploration. If no optimization constraints are applied, this can lead to format collapse in bounded, continuous spaces. To resolve this, we propose Direct Preference Density Alignment: An alternative framework that removes the need for a learned proxy reward model while strictly preserving the benefits of online reinforcement learning. We leverage large-scale user data (approximately 90,000 samples) to construct non-parametric preference density maps, establishing an empirical reward surface. In addition to removing the reward model, Direct Preference Density Alignment enables the combination of the online structural grounding of Group Relative Policy Optimization (GRPO) with the targeted offline refinement of DPO. We show that this GRPO+DPO combination achieves the highest performance, and in a blind audio equalization listening test, enables a 1.5B-parameter model to achieve perceptual parity with a carefully prompt-engineered GPT-4o mini baseline, using only a fraction of the inference compute.

Video tokenizers have emerged as a cornerstone of modern video modeling, underpinning progress in compression, reconstruction and generation by mapping high-dimensional visual signals into compact latent spaces. However, despite this progress, current tokenization paradigms largely remain within the 2D visual domain, treating videos as image sequences rather than observations of an underlying dynamic 3D world. Consequently, the learned tokens inherit this observation-centric bias, limiting their capacity to compactly represent real-world 4D scenes. To mitigate this issue, we propose VideoTok4D, a novel 4D-aware video tokenizer for compact world representation. Specifically, our approach comprises three key designs: 1) a spatiotemporal disentanglement strategy that factorizes videos into static and dynamic tokens for holistic world modeling; 2) a track-aware dynamic attention mechanism that aggregates trajectory-aligned cues to promote cross-view motion consistency; and 3) Co4DGen, a diffusion prior learned over the resulting VideoTok4D token space for efficient 4D scene generation. Extensive experiments have demonstrated that our proposed method achieves state-of-the-art performance while requiring up to 4 orders of magnitude less storage than dense 4D representations. Moreover, the compact token space substantially shortens diffusion sequences, enabling efficient generation.

Instruction-guided image editing has a training-time blind spot. Generative editors are never required to semantically verify whether their outputs actually satisfy the instruction. Supervision stops at reconstruction and input textual-level conditioning. This produces incomplete edits, spatial spillover, and poor localization. We present IABEdit, a model-agnostic framework that embeds differentiable semantic verification into training. A frozen vision-language model extracts spatially-aware descriptors from the ground-truth edit. A trainable aligner then reproduces them from the generated output. The residual between the two becomes a gradient that teaches the generator both what to edit and where, with no inference-time VLM cost. IABEdit is compatible with diverse backbones, including U-Net (Stable Diffusion) and MMDiT (FLUX), without altering their inference pipelines. On MagicBrush, it improves structural fidelity by +3.49 DINO-I over the best diffusion baseline and +1.26 over the best overall baseline, while remaining competitive on instruction alignment. It also achieves state-of-the-art instruction adherence performance on RealEdit and EMU Edit benchmarks based on embedding-based metrics. Most consequentially, on the D-LORD surveillance benchmark, it surpasses the proprietary Gemini agent by +5.13 DINO-P under heavy occlusion, where preserving identity is hardest. This shows that gradient-aligned VLM distillation holds up under real-world-like surveillance and occlusion conditions. Human and GPT-4o evaluations confirm perceptually precise, well-localized edits.

Multi-modal image generation, particularly subject-driven customization, has garnered growing attention in recent years. Despite the rapid advancement of generative models, their evaluation remains largely lagging. Existing methods, whether embedding-based or Multi-modal Large Language Model (MLLM)-based, evaluate alignment with each modal condition in isolation, which contradicts the simultaneous condition alignment objective of multi-modal image generation, leading to poor consistency with human judgments. To address this challenge, we propose UFO, the first unified framework for omni-condition alignment simultaneous evaluation. Specifically, UFO introduces a novel Atomized Chain-of-Evaluation paradigm, i.e., it first decomposes omni-condition alignment into a sequential chain of fine-grained, disentangled Atomic Evaluation Units (AEUs), categorizes them into distinct modality-relevance classes, and then employs general or dedicated functional calls for accurate verification of different AEU types. Experimental results demonstrate that UFO achieves the highest correlation with human evaluation preferences, delivering an average improvement of 15.25%. Furthermore, we present UFO-Bench, a dedicated benchmark designed to holistically evaluate the performance of existing customization models under the diverse mutual interactions of textual and visual conditions.

Training diffusion models with multiple rewards requires distinguishing user preference from reward informativeness. User preference determines how much each reward should contribute to the overall objective; reward informativeness determines when its feedback is useful during denoising. Some rewards can meaningfully evaluate a sample as soon as global structure emerges, but others become informative only when the sample is nearly clean. To address both questions jointly, we propose ReCAST (Reward Credit ASsignment across T}imesteps), the first method, to our knowledge, for per-reward, timestep-dependent credit assignment in diffusion reward fine-tuning. ReCAST separates user preferences from temporal allocation through a reward-by-timestep weight matrix $W$, whose row sums match the user-specified reward budgets $λ$, while its column sums are equal, assigning the same total weight to each denoising step. Under these marginal constraints, ReCAST allocates weight according to each reward's informativeness, quantified by its Rényi discriminability gain at each step. These gains telescope to the total discriminability between the reward-induced positive policy and the current policy, providing a basis for temporal credit assignment. We evaluate ReCAST by training SD3.5-Medium under two distinct four-reward settings, each across five reward budgets $λ$. ReCAST improves the training rewards in one setting and matches them in the other, improves every held-out judge in both, and is preferred by an independent LLM-as-a-Judge. Together, these results show that ReCAST yields improvements that generalize beyond the training rewards and support its core principle: assigning each reward greater weight at the denoising timesteps where its feedback is most informative.

Reinforcement learning is now the standard way to train large language model agents on long-horizon tasks, where dozens of interdependent actions precede a single sparse reward. Critic-free, group-relative methods such as GRPO suit this regime, but they broadcast one trajectory-level scalar to every step and cannot say which decision drove the outcome. GiGPO recovers a step-level signal by grouping time steps that share an anchor state, yet it merges the step- and episode-level estimates under one fixed weight, spending the same resolution on a pivotal branching decision as on a routine, near-deterministic transition. We argue that the right resolution is state-dependent, and propose GACA, a critic-free estimator whose granularity follows an uncertainty-based criticality proxy. GACA scores every step by the negative log-likelihood its own rollout already records, then blends the two advantages with a per-step weight that grows with that score, so the gradient places more weight on the fine-grained signal at above-average NLL and on the episode-level signal below it. We derive an exact risk decomposition for the implemented mixture and show that sufficiently small modulation improves on fixed mixing under positive directional alignment. A separate conditional result bounds local action-value variation using expected NLL, while an error-projection analysis characterizes when mixing adds value beyond scalar uncertainty reweighting. On ALFWorld and WebShop, GACA improves task success over GRPO and GiGPO at both 1.5B and 7B scales.

We post-train Qwen3.8-27B for Korean response style — verbosity, list and markdown usage, discourse structure and register — and measure two behaviours the objective never targets: abstention on ambiguous social questions in KoBBQ, where the benchmark-correct answer is UNKNOWN, and unprompted disclosure in securities guidance. Both move, and the changes are expressed primarily through the model's emission policy: how often it answers and how much it says. Matched target-form controls show that answer propensity depends on the training target, not the prompt set or recipe alone. Holding prompts, recipe, data volume and serving fixed and changing only the target text, three style seeds give positive answer-rate point estimates (mean +0.82 pp) and three neutral seeds negative ones (mean -1.53 pp); the observed seed ranges do not overlap and the means differ by 2.34 pp. A length-matched arm lies between them, and a fourth arm that stays short while preserving hedging is unstable across seeds, so which feature of the form is responsible is unresolved. For absolute stereotyped exposure the decomposition into an answer-propensity term and a conditional-composition term is an algebraic identity, not a finding; its empirical content is where the movement went. Across the trained checkpoints the changes are dominated by answer propensity while the composition term stays small, and because that term is evaluated on treatment-dependent answered subsets we do not read it as evidence about latent preference. Two measurement results follow. A between-arm contrast in conditional stereotyped share does not identify a change in conditional content preference when answer status is treatment-dependent. And agreement between two rule detectors for the same construct runs from 0.44 to 0.99 depending on which checkpoint produced the text — observable without any reference labels.

Neurosymbolic systems rely on mathematical solvers to guarantee reasoning correctness, yet solvers are fundamentally blind to whether a formal translation maintains strict reference-equivalence to a designated formalization. We formalize this vulnerability as Verdict-Preserving-Unfaithfulness (VPU): a failure mode where an incorrect encoding executes successfully and matches the expected verdict. We theoretically prove that structural, verdict-only verification heuristics are mathematically bounded to chance-level detection on these deceptively valid traces. To resolve this, we introduce Generative Verification (GenV), which distills an offline Z3-equivalence oracle into a reference-free, continuous reference-equivalence score by repurposing the language model's native vocabulary space. Mechanistic analysis via decision-projected logit lenses and sparse autoencoders shows this generative readout natively extracts precise spatial error coordinates without explicit localization training. Empirically, our oracle-mined verifier (GenV+HN) achieves 0.961 AUROC in reference-equivalence verification, generalizes zero-shot across unseen translators and divergent formal styles, and yields an 11.3-point downstream accuracy gain in agentic test-time compute allocation.

In recent years, artificial intelligence has made extraordinary progress thanks to large-scale models capable of generalization and the generation of complex outputs. However, transferring this potential into embodied agents reveals a significant limitation: the most advanced systems rely on pre-existing datasets and human feedback strategies that are powerful but insufficient in dynamic or unknown contexts. To adapt, an agent must acquire knowledge through direct interaction with its environment. One strategy to address this challenge involves introducing higher-level mechanisms, such as intrinsic motivations, which leverage curiosity and competence, to guide exploration and learning in complex environments. While this flexibility expands autonomy, it complicates the task of ensuring agents remain aligned with human goals. Alignment, already a challenge for artificial systems in general, becomes even more complex in unstructured and dynamic contexts where predefined rules prove insufficient. To be effective and adaptable, norms must be rooted in experience through an epistemological process that starting from simple, situated principles allows for the gradual construction of more complex rules through experience, autonomous learning, and cooperation with other moral agents. Similarly to children learning social norms by exploring their environment and participating in collective practices, artificial agents must also be educated toward alignment. Following Dennett, the status of a moral agent is not innate but is attributed gradually based on the ability to responsibly manage increasing degrees of freedom. From this perspective, the regulatory sandboxes can be viewed as pedagogical environments for AI: dynamic spaces where alignment develops as a formative process, progressively shaping autonomous behaviors through interaction and cooperation in scenarios of increasing complexity.

Large language model serving costs scale directly with output sequence length, yet standard preference alignment often inflates response verbosity without improving utility. We study whether the parameterization of post-training updates affects generation length: low-rank subspaces alter sequence length without modifying the alignment loss. We present LOCUS, a method that selects a task-aware low-rank adaptation subspace to minimize output-token cost subject to a utility constraint. Within this subspace, post-training retains the native preference objective with a frozen backbone. Across Anthropic HH-RLHF dialogue preferences, we evaluate two $\sim$3B decoder backbones, Pythia-2.8B and Qwen2.5-3B, against protocol-matched full-parameter DPO and DrDPO branches and the released SamPO checkpoint. LOCUS reduces continuation length by up to 39.84% on Pythia-2.8B and by 14.87--17.58% on Qwen2.5-3B while updating only 0.24--0.28% of model parameters, with no material change in the internal preference diagnostic.

Spatial reasoning depends not only on metric properties such as distance, angle, and shape, but also on topological relations that remain invariant under continuous deformation. Cognitive science identifies these relations as foundational to spatial understanding, yet foundation-model evaluations largely focus on metric or viewpoint-dependent relations. We introduce MindTopo, a benchmark of topological intuition across five properties grounded in cognitive science and formal topology: continuity, separation, order, enclosure, and knots. MindTopo evaluates each property at two cognitive levels. Reasoning asks a model to identify topological relations or infer how they change. Planning instantiates a foundation model as a closed-loop agent whose policy selects environment actions. MindTopo contains 11,030 instances across 13 procedurally generated task types with controllable difficulty. We benchmark 14 MLLMs and study agent configurations augmented with image and video generation, including 3 video generative models in planning settings. Every MLLM performs better on reasoning than on planning, and the best-performing model remains far below observed human performance. On Qwen3-VL-2B-Instruct, supervised fine-tuning and reinforcement learning improve reasoning more than planning. Generated observations retain local cues and reach plausible endpoints, but audited rollouts do not reliably follow environment dynamics or preserve topology across transitions. Our website is at https://mind-topo.github.io/

While Multimodal Large Language Models (MLLMs) have achieved remarkable progress across diverse tasks, they suffer from hallucinations where generated outputs contradict or misrepresent input semantics. Existing research typically addresses hallucination detection within a single modality or task type, limiting generalizability. We introduce OmniHallu, a unified hallucination detection framework spanning both comprehension and generation tasks across image, video, and audio modalities. We contribute OmniHallu-Bench, a 10,000-sample benchmark with claim-level human annotations covering six cross-modal tasks: image-to-text (I2T), video-to-text (V2T), audio-to-text (A2T), text-to-image (T2I), text-to-video (T2V), and text-to-audio (T2A). Our multi-agent architecture decomposes model outputs into atomic claims, verifies them through modality-specific experts, and aggregates evidence via structured reasoning. We further propose a preference-optimized trainable verifier that approximates the multi-agent decision boundary, reducing expert calls by 66% with minimal performance loss. Extensive experiments reveal a consistent modality-dependent performance gradient and provide fine-grained insights into cross-modal hallucination patterns.

High-dimensional dense text embeddings and large language models face real obstacles in financial-disclosure analysis: context-window limits, hallucination risk, high computational cost, and the arbitrary rotation of vector spaces across independently trained models. We present a training-free, alignment-free framework for corporate intelligence built on deterministic sparse seed vectors. Hashing word strings into a fixed high-dimensional basis places all documents and all temporal epochs in a common coordinate system by construction, removing any need for training or alignment. Accumulating these seed vectors across sentence contexts yields corpus-specific semantic signatures that compose linearly, supporting sub-second document comparison, issuer fingerprinting, tracking of how an issuer's vocabulary shifts between filings, and thematic sentence extraction, all on ordinary CPU hardware. Demonstrating the approach on a multi-year corpus of SEC filings (10-K, 10-Q, 8-K), we show how material corporate events, among them Boeing's 737 MAX crisis, Intel's supply-chain disruptions, and Bunge's acquisition of Viterra, emerge as distinct, interpretable semantic profiles, each traceable to the exact source sentences that produced it, with no domain-specific training and no LLM inference.

Test-time reinforcement learning can adapt vision-language models (VLMs) to unlabeled target data, but its effectiveness is fundamentally limited by the reliability of self-generated learning signals. To assess the reliability of consensus-based learning signals, we analyze VLM test-time reinforcement learning across diverse VQA datasets and model sizes, revealing two limitations. First, gains from consensus-based test-time training largely come from answer normalization rather than content correction. Second, many initial VLM responses are incorrect due to the model's limited ability to jointly use the image and the question; consensus rewards derived from these outputs may preserve the resulting grounding errors rather than correct them. Motivated by these, we propose TTIQ, a test-time reinforcement learning framework that harnesses image-question dependence for better vlm adaptation. TTIQ teacher-forces each sampled response under the original image-question pair and its image- and question-ablated variants, using the resulting token-level likelihood changes to estimate dependence on each input. It combines image and question dependence with calibrated confidence to construct a response-level reward that favors jointly grounded responses, and uses the token-level signals to assign greater positive policy credit to tokens supported by both inputs. This design favors responses that are jointly grounded in the image and the question and sufficiently confident, rather than merely popular. Experiments across eight VQA datasets and multiple VLM sizes show that TTIQ achieves the best average performance at every model scale. It further generalizes across VLM families, while models trained on one dataset improve performance on unseen datasets without further training.

Social intelligence enables agents to read social context, infer intent, and adapt over sustained dialogue. As language models become autonomous collaborators, it is central to building effective and trustworthy human-AI interaction. Existing reinforcement learning methods optimize single-turn utterances and sparse outcome rewards, producing short-sighted policies that struggle to manage goal-relationship tensions across multi-turn interactions. We propose SocialRL, a multi-turn reinforcement learning framework addressing both challenges. First, we apply multi-turn reinforcement learning using PPO that propagates delayed outcome rewards back to each turn, enabling long-horizon planning. Second, we design six process reward dimensions capturing the goal-relationship trade-off, including goal advancement, relational attunement, contextual coherence, etc. A reward model dynamically generates fine-grained scoring criteria for each dimension, while a stage-aware weight schedule prioritizes relationship-building in early turns, goal advancement mid-way, and balanced closure late. Across multiple social-dialogue benchmarks, SocialRL improves Goal Achievement by an average of 9.2 percentage points over the corresponding Base models. These results demonstrate the effectiveness of SocialRL across synthetic and real social scenes, as well as standard and challenging social scenarios.

Reinforcement learning over a frozen reasoner has become a common recipe for teaching a policy which external tools to invoke. We show that this recipe becomes structurally mismatched in specialist scientific settings where the complete tool-subset space is enumerable. There, a small set of recurring computational capabilities covers the domain, so the space of tool subsets is combinatorial yet small enough to enumerate, and GRPO still estimates an action expectation from a handful of sampled rollouts. Worse, the approximation degrades as training succeeds: as the policy concentrates on preferred subsets it resamples them, sampled rewards collide, and the group-normalized advantage vanishes. On genomic reasoning the fraction of questions yielding no reward signal rises from 0.2% under a uniform reference policy to 20.8% after GRPO training. As a remedy, we introduce FGPO (Full-Group Policy Optimization), which (1) scores every tool subset and optimizes the exact action expectation, so each update sees the complete action space, and (2) precomputes the reward of each question--subset pair into an exhaustive table, removing frozen-reasoner calls from the training loop entirely. Across five frozen reasoners and three genomic benchmarks, FGPO outperforms GRPO in all 15 settings by 6.75 points on average and up to 14.20, while a standard on-demand GRPO schedule would require 2.4 times as many frozen-reasoner reward evaluations and, on GenomeQA, FGPO cuts invoked tools per question from 2.36 to 1.40.

Asynchronous reinforcement learning has become the standard way to scale training for large language models (LLM), but the resulting policy lag biases the critic toward the stale behavior policy. Existing work on asynchronous LLM training corrects the actor and leaves this bias unaddressed, while the off-policy value correction of classical RL does not carry over to long-horizon agentic tasks, since a short correction horizon leaves the regression target free of the reward and a long one lets the product of importance ratios drift exponentially with the trajectory length. We propose BRACE, an anchored Bellman-residual correction for stale value models. BRACE bounds the correction horizon to a prefix of policy tokens and anchors a constant-weight Monte-Carlo tail beyond it, which separates policy correction from reward propagation. BRACE delivers a $9.8%$ relative improvement in mean@1 on BrowseComp-Plus over the strongest baseline, runs $2.46\times$ faster per step than synchronous training, and remains stable $50$ updates off-policy.

Data augmentation is fundamental to training modern deep vision and multimodal models. While individual methods, such as RandAug, CutMix, Mixup, RandErase, and DropPath, offer strong regularization effects, their combined use has saturated in performance due to overlapping functionalities, and aggressive pixel-level manipulations may disrupt delicate cross-modal alignment. This saturation motivates the search for a new augmentation axis within the embedding space rather than the input space. We introduce Aether, a simple plug-in method that applies diffusion-style random perturbations in the embedding space via controlled alpha-mixing, specifically designed to provide isotropic regularization that remains semantically consistent. Inspired by feature-space perturbations in language models and image degradation in generative pretraining, Aether induces mild yet effective perturbations that smooth the representations without compromising the fine-grained structural information required for strong vision-language encoders. Across diverse architectures and across multiple recognition tasks, Aether delivers consistent gains over the advanced recipe combining CutMix, Mixup, DropPath, and RandAug---a level of improvement rarely observed with modern augmentation alternatives. Notably, Aether demonstrates superior effectiveness in multi-modal alignment, succeeding where traditional pixel-space augmentations fail by providing a stable, isotropic regularization signal that respects the integrity of the high-dimensional feature space.

Frontier gains in language-model reasoning come from reinforcement learning on reasoning traces and are concentrated in domains with a cheap, sound verifier. We argue the field's binding constraint is the verification gap: no scalable, incorruptible reward for reasoning outside formal domains. We make four contributions. (1) Theory: in a joint-Gaussian model of best-of-N selection, verifier-gold correlation rho is the exact exchange rate between test-time compute and capability, and an unsound verifier pays a polynomial penalty N^(1/rho^2); a margin-free copula form predicts realized soundness of real LLM judges to 4% median error. (2) Demonstration: in program-synthesis testbeds with executable ground truth, including a pre-registered scaled replication, unsound verifiers lose Soundness-under-Pressure as optimization grows (0.94 to 0.32 at N=4096) while a sound verifier improves monotonically; reality-anchored settlement beats a frozen verifier under i.i.d. and adversarial pressure, driving the hacking gap from ~0.27 to ~0; soundness scales log-linearly with settled labels, with on-policy settlement ~10x more label-efficient than random labeling. With real LLM judges and unit-test execution as gold, a weak judge loses soundness under best-of-N (p<0.001), a stronger judge is more robust, and selection alone manufactures +0.53 hacking gaps from honest samples. Under real GRPO training, a frozen reward model traces the full overoptimization curve (executed reward collapses 90%) while the same model refit on a 10% settlement stream preserves 6x the executed reward. (3) Paradigm: proof-carrying cognition, where reasoning steps are typed probabilistic claims priced by a self-built world model trained only on held-out reality and settled by proper scoring rules. (4) Benchmark: we specify Soundness-under-Pressure as the headline metric for a reality-settled reasoning benchmark.

Preference alignment for flow and diffusion models now spans online reinforcement learning and offline preference optimization, but the relation between these methods remains unclear. In particular, existing forward-process alignment methods require fresh samples from the current model, while offline methods based on fixed preference pairs rely primarily on positive-only fine-tuning or DPO-style likelihood-ratio surrogates. We organize these approaches through a divergence-based framework and introduce FlowCPO, an offline forward-KL objective that uses both preferred and dispreferred samples without online rollouts. For linear interpolation, we show under explicit regularity conditions that the forward-KL objective is bounded by a contrastive flow matching loss, yielding a tractable surrogate on fixed data. We further show that this loss is nonnegative, whereas the signed regression loss of simplified FlowDPO can be unbounded below. In the in-domain setting, FlowCPO achieves higher mean GenEval and OCR scores than the evaluated baselines, reaching 0.84 and 0.87 versus 0.81 and 0.74 for FlowDPO at CFG 3.0. In the out-of-domain setting, the results are mixed, with the best GenEval result but lower reward scores than RFT on several metrics.

Personalized language models aim to adapt responses to individual users, whose preferences are often latent and revealed gradually through interaction. Existing training-free methods rely on stored histories or retrieved memories, but they often struggle to reconcile long- term preferences with short-term topic-specific needs. To address this issue, we propose HyperTrace, a training-free framework that formulates online personalization as latent preference tracing. HyperTrace maintains interpretable natural-language hypotheses over short-term intent and long-term preferences, and updates them through an SMC-style reweight process using an LLM-based surrogate choice model. By updating these hypotheses across turns and sessions, HyperTrace enables personalization without parameter updates. Experiments on PRISM and PersonaMem-v2 show that HyperTrace improves response alignment, preference prediction, and profile consistency over strong online baselines, demonstrating the effectiveness of tracing latent user preferences for robust personalization. Code and scripts are available in the repository: https://github.com/jiseshen/HyperTrace.

Multi-modal large language models (MLLMs) have demonstrated significant potential in image quality assessment (IQA) by bridging visual perception with descriptive evaluations. However, existing approaches mainly focus on holistic quality prediction, often functioning as black boxes that provide limited insight into where distortions occur and how they affect perceived quality, hindering fine-grained analysis of localized and heterogeneous degradations. We propose GS-IQA, a framework that reformulates IQA as a progressive Where--What--How diagnosis, emulating the human perceptual process from an initial glance to closer scrutiny. Since a severity judgment is meaningful only for a correctly localized and recognized region, we realize this progression through a two-stage reinforcement learning paradigm that respects such dependencies: the glance stage uses a perception-gated reward to establish where degradations lie and what they are, activating severity feedback only once both are correct, while the scrutiny stage introduces online reward-conditioned degradation generation to synthesize hard examples targeted at the model's perceptual bottlenecks, sharpening its discrimination of subtle severity variations. To enable systematic evaluation, we construct Diag-Bench, a region-level IQA benchmark of about 25K curated samples spanning 12 distortion types and five ordinal severity levels. Extensive experiments show that GS-IQA consistently surpasses state-of-the-art methods in distortion localization, recognition, and severity estimation, and that its diagnostic representations transfer effectively to conventional global quality prediction across diverse external benchmarks. Code and data will be released.

While many video-game environments (VGEs) have played crucial roles in advancing reinforcement learning (RL), developing novel VGEs or modifying existing ones to support new features, has been a laborious process requiring extensive hand-coding. Here we present PlayTrain, an RL framework that combines the abilities of large language models (LLMs) to robustly generate JavaScript (JS) games from a minimal human prompt, and an efficient pipeline that can run any JS game in a standard 'gym' environment. Not only are recent LLMs particularly good at writing JS code, but the JS format also allows users to easily play generated VGEs, while PlayTrain enables us to train RL agents on the exact same games. We demonstrate multiple use cases of PlayTrain, including cloning well-known Atari and ProcGen games in simple JS, where PlayTrain trains pixel-based agents end-to-end at over 1M agent-decisions per second on a single GPU node; and creating modified versions thereof (e.g., that support novel test sets, procedural generation logics, or game dynamics). Through PlayTrain, we reimagine RL VGE development: all we need is a single JS file, generated and modified through an LLM. We discuss promising future RL research directions that PlayTrain unlocks.

In reinforcement learning with verifiable rewards (RLVR) trained with group relative policy optimization (GRPO), the KL-free reward-advantage term studied here depends on within-group reward variation. If all rollouts in a group are correct or all are wrong, their group-relative advantages are identically zero; these zero-advantage silent groups provide no reward-advantage gradient, yet uniform sampling spends 39% of a run's rollouts on them. History-based prompt selection must first spend target-policy rollouts to estimate difficulty, creating a cold start with rollout waste; ThinkPrior instead uses an external anchor in one offline pass to construct a zero-rollout difficulty prior before the first target-policy rollout. The verifier-scored anchor pass rate supplies an external-anchor initialization for a Beta posterior; ThinkPrior selects by expected learnability and then updates from training outcomes, changing neither the loss nor the optimizer. On Qwen2.5-Math-7B across sixteen seeds, ThinkPrior more than halves early silent groups and cuts wasted rollouts through step 30 by nearly a fifth, while we detect no difference in final accuracy. On this 250-prompt pool the fixed-budget result is a reallocation rather than a net saving. The measured ThinkPrior+DAPO composition reduces generated rollouts by 10.6% while both arms retain the same 3840-rollout update budget. The prior requires no target-policy rollout before the first selection, but the posterior thereafter uses target-policy outcomes.

Real-world Multi-Objective Reinforcement Learning (MORL) often suffers from sparse rewards, reward conflicts, and late-stage reward tug-of-war, causing traditional linear scalarization to experience severe metric oscillations. To address optimization conflicts among multiple objectives in real-world deployment scenarios, we propose Multi-Marginal Preference Optimization (MMPO), a fine-grained framework that intervenes at the data, gradient, and constraint levels rather than relying on coarse-grained global scalarization. Specifically, MMPO performs exposure debiasing to mitigate sparse and biased rewards, applies priority-aware orthogonal projection to decouple conflicting gradients, and introduces self-prompted gradient constraints to prevent dominant objectives from overwhelming weaker ones. Experiments on real-world e-commerce datasets show that MMPO improves training stability and consistently achieves better performance across conflicting metrics. Moreover, it generalizes robustly to broader tasks such as ToolRL and code generation, demonstrating its effectiveness as a practical paradigm for multi-objective alignment.

Social interaction is central to children's language learning, but the effects of different forms of caregiver feedback are difficult to isolate in naturalistic data. We use child-like language models as controlled learners to test which forms of feedback support grammatical development. Small GPT-2-style models are pretrained on child-directed language from CHILDES, then fine-tuned with reinforcement learning using reward models trained to capture four feedback types: communicative feedback, structural alignment, semantic contingency, and affective feedback. Reward fine-tuning yields limited gains on minimal-pair evaluations, but clearer effects in free generation. Structural alignment produces the strongest improvements in grammaticality, providing a novel, plausible mechanistic account of how this feedback can support grammar learning. Communicative feedback yields more moderate gains. In contrast, semantic contingency and affective feedback do not improve grammaticality, although further analyses suggest that they may support other aspects of language learning beyond grammar. These results suggest that different forms of caregiver feedback make complementary contributions to language learning.