Generative RL 进展

Language-model agents increasingly tackle long-horizon tasks whose interaction histories exceed the model's active context. Recent work has begun to use reinforcement learning to make memory control part of the policy, often relying on predefined memory tools within domain-specific training environments of relatively short horizons. This setup ties learned memory behavior to environment-specific interfaces that lie outside the base model's pre-training and must be learned from scratch, so even after post-training, agents struggle to use memory in long-horizon tasks. To address these limitations, we introduce Coding Agent Memory Gym (CAMG), a suite of long-horizon agentic-RL environments spanning Shop, Coding, DeepResearch, and AutoResearch. Alongside each environment's native task interface, CAMG provides executable shell access and an episode-persistent workspace, enabling agents to create, revise, search, and reuse files as memory throughout an episode. We also introduce CAMG-RL, which trains a single policy jointly across all four environments with fully asynchronous PPO, learning this file-based memory behavior directly from downstream task reward, and we train CAMG-RL-4B and CAMG-RL-9B from Qwen3.5 models of matching size. On SWE-bench Verified and MLE-bench Lite, CAMG-RL-4B and CAMG-RL-9B are competitive with Qwen3.5-35B-A3B and Qwen3.5-122B-A10B, respectively.

Power sampling has emerged as a training-free approach to LLM reasoning, eliciting capabilities comparable to reinforcement learning by sharpening the model distribution over complete responses. Despite this success, power sampling remains underexplored in large vision-language models (LVLMs). We transfer Power-SMC to LVLM decoding by defining a sequence-power target conditioned on both the image and the prompt. This direct transfer provides a strong training-free baseline, but leaves two aspects of finite-particle multimodal inference unaddressed. At the particle level, global resampling can collapse genealogies, while particle-based power sampling does not diversify trajectories through distinct visual cues in multimodal decoding, limiting exploration under a finite particle budget. At the answer level, sequence-level sharpening makes distinct reasoning trajectories compete even when they support the same answer. We introduce ReSight-SMC, a verifier-free two-stage power sampler for LVLM inference. Its first stage uses ancestry-isolated SMC islands to preserve independent trajectory families and routes a bounded set of prefix-conditioned visual scouts to prefix-relevant image regions while discouraging redundant overlap. Each scout temporarily increases attention to the image tokens and emphasizes its routed region. Exact importance correction preserves the base LVLM sequence-power target. The second stage aggregates terminal importance mass by canonical answer, powers the answer marginal, and samples an answer together with a supporting trajectory. Across four LVLM backbones and five benchmarks, ReSight-SMC achieves stronger aggregate performance than Power-SMC over both the reasoning and perception benchmark groups. Without post-training, it remains competitive in aggregate with backbone-matched models trained using reinforcement learning.

Tool-based image editing (image retouching) is commonly formulated with autoregressive multimodal large language models (MLLMs) that sequentially generate reasoning, tool selections, and parameter values. In this work, we present a novel approach to tool-based image editing by framing the task as a flow matching problem. We introduce FlowTool, a framework that directly models the distribution of high-quality tool parameters conditioned on the input image and user instruction using conditional rectified flow. FlowTool combines a vision-language model backbone for multimodal understanding with a Diffusion Transformer parameter generator that transforms Gaussian noise into an editing plan. We train FlowTool with a two-stage supervised flow-matching curriculum, followed by reward-based post-training. Across MMArt-Bench, FlowTool-Eval, ArtEdit-Bench, and MIT-Adobe5K, FlowTool achieves significantly stronger reference-based performance than specialized MLLM editing agents and proprietary MLLMs, while remaining competitive with proprietary models under reference-free evaluation. Moreover, FlowTool significantly improves inference efficiency, reducing latency by at least $50\times$ while requiring nearly $2\times$ less memory than the compared baselines. These results demonstrate that tool-based image editing can be effectively modeled as conditional generation over structured continuous editing parameters, without autoregressive reasoning.

Reinforcement learning (RL) has become an effective post-training paradigm for long-horizon large language model (LLM) agents. However, we find that the resulting policies can be sensitive to various policy perturbations, such as hidden-state noise, pruning, and quantization. In this work, we study how to improve perturbation robustness during policy optimization. We first introduce the notion of a perturbation robust policy and analyze conditions under which perturbed policy updates preserve stable monotonic improvement. Based on this analysis, we introduce Stable Perturbation-Robust Policy Optimization (SPrPO), which applies adaptive and sensitivity-aware perturbations during RL training. We evaluate SPrPO on ALFWorld and WebShop and conduct systematic experiments across multiple perturbation types and scales, showing improved perturbation robustness while maintaining stable policy optimization.

Reinforcement learning (RL) post-training often uses distinct GPU kernels for rollout and policy update. In synchronous PPO and GRPO, numerical disagreement can perturb ratios between current token probabilities and those assigned during rollout. Recomputing rollout log-probabilities with the policy-update backend avoids this discrepancy but adds a forward pass. Bitwise-consistent unified kernels permit reuse when the policy snapshot and probability processing match the objective. Their optimization must preserve agreement across distinct execution regimes. We present KernelBraid, an agentic framework starting from a hand-tuned, bitwise-consistent implementation. Its optimization intermediate representation (IR) organizes source-code search by linking implementations and modifications to numerical requirements, workload measurements, and derivation history. The agent coordinates changes and retains verified intermediates for further exploration; promotion requires passing correctness checks and improving aggregate latency within per-workload limits. Across 12 end-to-end training configurations on H20, KernelBraid achieves 1.10x average throughput relative to AReaL with log-probability recomputation, and the mean training-reward ratio rounds to 1.00x. Isolated-layer profiling yields 1.40x average speedup in summed phase time across 15 model-GPU pairs. Operator-level evaluation covers correctness and performance for 10 operators on A100, H20, and H200, all passing the prescribed bitwise checks. Unified-attention search achieves 2.52x speedup in summed workload latency over the starting implementation using 7M LLM tokens; ablations assess the contributions of retained evidence and branch exploration to search efficiency and attained performance. Our code is open-sourced at https://github.com/areal-project/AReaL-TIK.

Scaling the horizon of agentic LLMs is bottlenecked by the need to fit ever longer context traces in GPU memory. Context compaction has been the most popular mechanism to alleviate this issue, keeping GPU memory constant for a given trace. Unfortunately, most compaction strategies rely on prefilling the LLM context many times over, hindering training throughput. To alleviate this bottleneck and enable efficient trainable compaction, we propose KV-streams, a plug-and-play strategy compatible with any compaction strategy that substantially increases throughput while showing no evidence of hindering performance. KV-streams enable scalable compaction by streaming the KV cache forward rather than flushing it after each compaction. We show that KV-streams enable three different compaction strategies, achieving a 2.6 to 5x wall-clock speedup in training. Beyond efficiency, we find that the streamed KV cache can act as a recurrent state, carrying forward information that has long since disappeared from the context. Specifically, in a controlled setting we show that, contrary to prior work, RL alone is all that is needed for this behavior to emerge. Overall, we show KV-streams to be an efficient and lightweight plug-and-play addition to any post-training pipeline.

Continuous diffusion generates complete reasoning solutions through iterative refinement in latent space. We introduce the Continuous Embedding Diffusion Reasoner (CEDR), an ELF-based training and inference recipe. Our experiments show that accurate decoding alone does not ensure strong reasoning performance. We therefore learn compact representations from multiple layers of a strong autoregressive teacher. Their decomposition also enables asynchronous denoising at different rates. We show that prompt encodings need only preserve the information required for the correct text-conditional score, rather than exactly match teacher features, and use a staged curriculum to learn a compact prompt encoder that replaces the teacher Transformer at inference. We adapt DiffusionNFT to learned self-conditioning guidance and incorporate gold-solution endpoints to supplement sparse rewards. Our supervised models outperform reported results from recent continuous-diffusion baselines at comparable backbone scales on mathematical reasoning and HumanEval code generation. With a 638M-parameter denoising backbone and learned prompt conditioning, post-NFT CEDR-L achieves 63.74% pass@1 on GSM8K and 24.6% on MATH500 at 64 denoising steps, and 32.85% on HumanEval and 30.18% on HumanEval+ at 128 denoising steps. Code will be available at: https://github.com/chengxiang/CEDR.

Preference-based alignment methods such as Direct Preference Optimization (DPO) use pairwise preferences labeled by human annotators to fine-tune language models. However, annotators carry systematic biases toward some attributes: a name that signals a gender or an ethnicity, a persona, a language variety, a formatting convention, or length. If not properly addressed, these systematic biases can be absorbed and amplified during alignment. Existing methods address length bias or annotator disagreement, but fail to eliminate biases toward arbitrary attributes. To address this limitation, we propose Bias-Adjusted DPO (BA-DPO), a generalization of DPO that adds one bias parameter per annotator toward responses carrying a declared attribute. We prove that the objective is convex in the bias parameters and that the votes identify each annotator's bias up to a shared constant. The remaining constant is what fixes the aligned model's attribute rate: by default the rate of the reference model, or a target rate, which we use to bring a biased policy to statistical parity. On a corpus with planted biases, DPO drives the attribute from a balanced start to probability 0.96 and BA-DPO removes 81 to 95% of that shift; on MultiPref with real annotators it removes about half of DPO's lengthening. Both hold at 0.5B with full fine-tuning and at 8B with LoRA, at no higher KL than DPO and no loss in judged quality.

The reasoning trajectory of a Large Language Model (LLM) is often treated as a verbalized description of its internal reasoning. However, such trajectories can be unfaithful: a model may rely on shortcuts to reach an answer and then post-rationalize the decision with a seemingly coherent chain of thought. Detecting this shortcut reasoning is challenging because existing monitors and verifiers mainly inspect textual traces or final outcomes, rather than how the model's belief in its answer develops during generation. We introduce ConfLens, a framework that tracks the evolution of confidence in the final answer throughout reasoning. Across three shortcut reasoning settings, we observe a common pattern of premature confidence, where shortcut samples become highly confident in the final answer at early reasoning stages. Existing confidence estimation methods, however, show limited generalizability, reliability, or efficiency for detecting this behavior. We therefore propose the Distributional Answer Commitment Score (DACS), a distributional confidence estimator that measures the entropy of the model's probability distribution over answer commitment at each reasoning step. DACS captures how concentrated the model's answer belief is without requiring ground-truth answers or task-specific verifiers. We further convert ConfLens detection results into interpretable signals for reward models to reduce their preference for shortcut reasoning. Experiments on mathematical and code reasoning tasks show that ConfLens with DACS improves shortcut reasoning detection by over 4.3% F1 compared with strong baselines and reduces the mismatch between faithfulness and correctness in reward model preferences.

Latent visual reasoning (LVR) enables multimodal large language models (MLLMs) to perform intermediate computation in continuous latent tokens rather than expressing every reasoning step in words. However, unlike textual CoT, latent reasoning is not directly observable, making it difficult to supervise what latent tokens learn. In this work, we first conduct a thorough analysis of latent-token behavior and identify a latent evidence-credit gap: latent tokens respond only weakly to image perturbations that alter the correct answer. We hypothesize that this issue stems from the lack of explicit supervision during GRPO training. These findings suggest that a final-answer reward provides too little guidance on what visual evidence to preserve or how credit should be assigned across latent tokens. To bridge this gap, we propose ReaLVR, which brings visual-evidence supervision to the model's own free-running latent trajectories. ReaLVR contrasts correct and model-generated wrong answers to determine where stronger supervision is needed, and relevant and mismatched visual evidence to specify what to preserve. Across three model families, ReaLVR consistently outperforms evaluated LVR baselines, achieving the highest five-task average of 63.7% on Qwen2.5-VL-7B. Crucially, we are the first to scale visual reasoning in latent space, showing that our framework continues to deliver robust improvements at frontier model scales up to 235B. Further analyses show more question-sensitive latent-token positions, stronger alignment with relevant visual regions, and greater fixed-context dependence on the most attended latent tokens.

Reinforcement learning with verifiable rewards (RLVR) is often limited by insufficient exploration: difficult problems can yield uniformly incorrect rollout groups and therefore little learning signal. We show that such failures need not reflect missing capability. Instead, finite sampling often concentrates on a problem-specific dominant reasoning strategy while leaving alternative strategies already supported by the model unexplored. Moreover, the accessibility of these strategies evolves during RL: some are internalized into autonomous behavior, while others become difficult to elicit before being absorbed. Motivated by these observations, we introduce Problem--Strategy Rollout Allocation (PSRA), which treats unguided and strategy-conditioned prompts as competing exploration arms and uses Bayesian sequential allocation to direct a fixed rollout budget toward arms most likely to yield informative, non-saturated groups. A preservation objective keeps useful strategy-conditioned routes accessible while successful guided behaviors are transferred to the unguided policy. Across Qwen2.5 models from 1.5B to 7B and two RL training corpora, PSRA consistently improves reasoning performance, reduces dead saturation, strengthens out-of-distribution transfer, and maintains larger gains under increased inference budgets.

Test-time scaling and post-training have improved LLM performance in coding and mathematical reasoning, but their effectiveness for individual stance prediction remains unclear. We study this question by predicting a person's stance in a new discussion from their history. We evaluate widely used test-time scaling strategies and post-training methods, such as supervised fine-tuning and reinforcement learning, and identify four failure modes across generation, selection, and learning: (1) incorrect consensus, where repeated samples agree on the wrong stance; (2) selection failure, where generation covers the observed stance but selection misses it; (3) response overfitting, where supervised fine-tuning improves imitation but harms prediction; and (4) early plateau, where reinforcement learning shows modest initial gains followed by limited further improvement. We expose these failures using STANCE-BENCH, which contains 2499 prediction tasks from 500 Hacker News users. Guided by this analysis, we explore a simple approach that combines direct scores for all candidate stances with explicit assessments of support from the individual's history. On the 781-task test set, this approach achieves 21.83 discussion-specific Macro F1 with Qwen3-8B, compared with 19.27 for direct scoring. Our results motivate evaluating candidate generation, final selection, and person-specific evidence use separately. Our data is available at https://github.com/stance-bench/Stance-Bench.

Outcome-based reinforcement learning can produce models with similar task performance but very different ways of communicating about their mistakes. We study failure disclosure: whether a model admits that an attempted solution failed rather than staying silent or presenting it as successful. Across repeated outcome-only GRPO training runs, failure disclosure varies far more than task accuracy. The pattern extends to a second reasoning task and stabilized PPO, persists at 7B, and also appears in an instruction-conditioned 32B setting. We also find that small floating-point and sampling differences during training can redirect reporting behavior even when the task objective and earlier training history are held fixed. Additional tests show that failure disclosure is not a single decision: Checking the answer, entering a report, and completing the admission can separate, and the weak point depends on the task and response format. Further, experiments with neutral controls show more broadly that behaviors left weakly constrained by training are especially likely to vary across runs, of which failure disclosure is an example. We can reduce variability in failure disclosure by discouraging the model from drifting from its starting policy on failed, well-formed responses. This makes reporting substantially more consistent, though its effect on task performance depends on the setting. Stable task accuracy therefore does not guarantee stable safety-relevant behavior: Researchers should measure these behaviors directly across runs and design training methods that keep them reliable.

Online safe reinforcement learning (RL) seeks policies that maximize reward while satisfying safety constraints. A popular line of research in safe RL relaxes safety to a soft expected-cost constraint and solves the resulting Constrained Markov Decision Process via primal-dual Lagrangian updates that only enforce safety on average. To address this limitation, hard, state-wise constraints are introduced and often imposed through Hamilton-Jacobi (HJ) reachability. Yet such constraints require solving different objectives in the feasible and infeasible regions: reward maximization in the former, recovery toward the feasible regions in the latter. The resulting target action distributions are inherently multimodal, and this structure poses a fundamental challenge for the Gaussian or deterministic actors used in existing HJ-based safe RL, which often collapse onto suboptimal modes. Diffusion policies provide the expressiveness needed to represent such distributions, and recent work on Q-score matching offers a route to training them for online RL by score regression — but has been applied only to reward maximization. We propose Safe Score Matching (SSM), an off-policy actor-critic method that adapts Q-score matching to hard-constrained safe RL by gating a two-branch score target with HJ reachability: inside the feasible set, the denoising process degenerates to Q-score matching on actions classified as viable by the HJ critic; outside, a recovery branch biases denoising toward regions with lower worst-case violation. On quadrotor and fixed-wing trajectory-tracking and stabilize-and-avoid benchmarks, SSM attains the best or near-best task performance with low false-safe rates, whereas the primal-dual baseline admits more unsafe behavior and reachability-based baselines tend to be more conservative; on Safety-Gymnasium velocity tasks, SSM attains the lowest cost with competitive reward.

Universal multimodal retrieval requires compact embeddings that preserve task-relevant semantic information across diverse modalities. Prior works have incorporated latent reasoning into multimodal embedding learning to refine this information before embedding extraction. However, most existing approaches remain confined to deterministic latent paths, without exploring alternative trajectories to discover better embeddings. Thus, we propose VaME (Variational Multimodal Embeddings), a framework that models latent reasoning as a learnable distribution over trajectories. Specifically, we first introduce Variational Latent Reasoning (VLR) to enable autoregressive exploration in latent space, guided by answer reconstruction through a lightweight decoder. Meanwhile, we augment the original embedding-token readout with a latent-fused embedding to facilitate exploration during subsequent reinforcement learning. Finally, we optimize latent reasoning over stochastic variational trajectories through reinforcement learning, using Semantic Decoding Reward (SDR) to favor semantically meaningful trajectories with interpretable decoded outcomes. On the 78-task MMEB-V2 benchmark, spanning image, video, and visual-document retrieval, VaME outperforms most explicit CoT-based models and all latent-reasoning baselines. VaME also demonstrates robust performance on reasoning-intensive benchmarks such as MRMR, with substantial gains after reinforcement learning. Importantly, VaME achieves these gains with at least a 4.25x inference speedup over the deterministic latent autoregressive baselines. The code will be made publicly available.

Test-time reinforcement learning enables vision-language models (VLMs) to adapt using unlabeled inputs. However, repeated sampling under fixed visual conditions can reinforce shared perceptual errors, while sequence-level rewards fail to isolate visual perception the foundational bottleneck that anchors multimodal reasoning risking the degradation of pre-trained reasoning capabilities. We propose TTRSD, a test-time reinforcement learning framework combining multi-view answer-level self-distillation with visual contrastive token selection. A shared policy aggregates teacher predictions across original, cropped, and downsampled views into an answer distribution. Student trajectories generated from the original image receive rewards based on the support for their final answers in this distribution. To allocate this feedback precisely toward perceptual bottlenecks, we compare the log-probabilities of the same sampled tokens under original and visually ablated inputs while holding their textual prefixes fixed, selecting visually sensitive positions for policy-gradient updates. TTRSD separates update direction, determined by group-relative advantages, from update position, determined by visual sensitivity, without requiring ground-truth labels, external verifiers, or a separate teacher. With only 20 unlabeled adaptation samples, TTRSD improves performance across seven benchmarks and three VLMs, raising InternVL3-2B's MMMU accuracy from 35.79% to 49.32%(+13.53%), demonstrating cross-dataset generalization while preserving inherent reasoning integrity.

Efficient exploration often remains a central bottleneck in reinforcement learning with verifiable rewards (RLVR). Although temperature control and test-time scaling strategies can increase rollout diversity of large language models (LLMs), they either expand the sample budget at rollout time or leave the benefit of exploration unquantified. To this end, we propose Temperature-Grouped Reinforcement Learning (TGRL), which turns temperature-induced diversity into an explicit training signal. For each prompt, TGRL partitions its rollout group into low- and high-temperature subsets, estimates exploration gain through their reward contrast, and allocates this group-level signal as token-level credit using Jensen--Shannon (JS) divergence between the corresponding temperature-scaled next-token distributions induced by the same logits. Notably, TGRL reaches equivalent accuracy up to 36% faster than strong RLVR baselines without expanding the rollout budget. Across 11 benchmarks from diverse domains, TGRL broadly improves over strong RLVR baselines: it improves the six-benchmark math average by 1.6% at 32B, raises CodeForces rating by 196.7 points and LiveCodeBench Pass@16 by 4.4%, and improves ALFWorld/WebShop success rates by 6.3%/4.9%. Comprehensive ablations and wall-clock analysis confirm the efficacy of all proposed components. Code is available at https://github.com/1229095296/TGRL/tree/main.

In-context learning (ICL) is crucial for boosting the inference performance of large language models (LLMs). However, the effectiveness of ICL in LLMs is greatly influenced by the choice of demonstration sets. Exhaustive searches over these sets are combinatorial, and existing selectors often rely on relevance or likelihood proxies to implicitly assess ICL quality. Making repeated queries to the target LLM with these strategies can incur substantial costs. This work simplifies selection by framing it as a constrained local search problem and presents local demonstration editing (LDE). Starting with an initially retrieved set of demonstrations, LDE employs a single structured edit to explore its surrounding neighborhood while balancing performance gains with search costs. Technically, LDE is reduced to a policy search problem, for which we train a small LLM, referred to as Jev-LDE. This model as the System-1 modifies the retrieved demonstration set by performing actions such as Keep, Delete, or Replace elements, all within a framework of reinforcement learning with verifiable rewards. At test time, Jev-LDE executes a single edit of the retrieved demonstration set, followed by one inference from the target LLM, avoiding the need for iterative context scoring or subset searches. Across standard classification benchmarks, various target LLMs with Jev-LDE as the plug-and-play module consistently improve ICL performance, and Jev-LDE shows transferability to held-out benchmarks and models without retraining. These findings indicate that the LDE approach offers an efficient and adaptable method for harnessing the ICL capabilities of target LLMs.

Prompt injection is a leading security risk for LLMs and LLM-based applications such as agents. State-of-the-art red-teaming methods for prompt injection leverage reinforcement learning (RL) to train an attacker LLM to generate effective injected prompts. However, when targeting frontier LLMs such as GPT-6-Luna, a major challenge is the cold-start problem: every attack attempt by the attacker LLM fails and thus receives zero reward, providing no signal for learning. In this work, we propose a curriculum learning-based method to address the cold-start problem. In particular, we propose to train the attacker LLM against a sequence of increasingly robust target LLMs, with each stage warm-starting from the attacker LLM obtained in the previous one. However, simply training against a weak target (e.g., GPT-4o-mini) may not sufficiently prepare the attacker LLM to obtain useful learning signals against a frontier LLM (e.g., GPT-5.6-Terra). Instead, we find that the design of the curriculum is critical: after each stage, the attacker LLM needs to partially succeed against the next target LLM such that it can learn from successful attempts to attack the new target. Our extensive evaluation shows that our method can effectively red-team frontier LLMs, achieving an attack success rate (ASR@10) of 93.8% and 45.0% against GPT-5.6-Luna and GPT-5.6-Terra on AgentDyn, whereas state-of-the-art RL methods such as RL-Hammer and PISmith achieve 0% ASR under the same setting. Moreover, we find that the attacker LLM transfers across targets, e.g., an attacker LLM trained to defeat one strong LLM (GPT-5.6-Terra) also succeeds against six other frontier LLMs (e.g., GPT-6-Luna) it was never trained on. Our code is available at https://github.com/albert-y1n/PIForge.

Reinforcement learning with verifiable rewards (RLVR) enhances reasoning in large language models (LLMs) through outcome-level feedback, yet recent approaches to finer-grained credit assignment often require auxiliary models, additional sampling, or privileged information. Although policy entropy provides a readily available signal, prioritizing uncertain positions under both reinforcement and penalization concentrates penalties where failed responses still retain alternatives for recovery, which can suppress opportunities for exploration. To address this, we introduce Entropic Advantage Policy Optimization (EAPO), an entropy-guided credit assignment method that treats success and failure asymmetrically. Specifically, motivated by the observation that success under uncertainty is less repeatable while confident failures tend to recur, EAPO couples normalized policy entropy with the sign of the response advantage to reinforce surprising success and correct repeated failure. It assigns stronger reinforcement to high-entropy decisions in successful responses and stronger penalties to low-entropy decisions in failed responses, while attenuating penalties at uncertain positions to preserve opportunities for recovery. By redistributing the response advantage across tokens, EAPO derives token-level credit directly from existing rollout signals without additional supervision. We validate EAPO on a range of reasoning tasks across both base and reasoning backbones, demonstrating that it achieves the best overall performance. We further show that EAPO promotes more effective exploration, broadening problem coverage and generating more diverse candidate answers.

World-action models can jointly predict future visual observations and robot actions. However, discrepancies may exist between their visual predictions and the consequences implied by generated actions. We observe that WAMs can often generate visually plausible task-completion outcomes before producing action sequences that reliably achieve them. Consequently, we treat the WAM-generated visual prediction as a goal-conditioned visual proposal rather than a directly executable plan. We use a frozen action-conditioned world model to predict action-conditioned consequences and construct feedback based on consistency between the two future predictions and alignment with the terminal goal. Leveraging this feedback, we employ Flow Policy Optimization (FPO) to optimize the action head of the WAM. This framework avoids online robot interaction and additional training of task-specific reward models. Across four real-world UR5 manipulation tasks, our method increases the mean success rate from 43.4% to 75.1%, compared with 61.4% for $π_{0.5}$. These results show that cross-model prediction discrepancy can provide useful feedback for improving robot policies under the evaluated manipulation tasks. Website: https://imagine-to-achieve.github.io/

Speech-to-speech translation (S2ST) in time-sensitive applications such as video dubbing requires not only semantic fidelity and speaker preservation, but also strict duration consistency to avoid audio-visual misalignment. However, existing S2ST systems largely generate target speech without explicit temporal planning, making duration control an unresolved challenge. We introduce DuraS2ST, a duration-aligned reasoning framework that enables a single speech language model to first generate an explicit chain-of-thought (CoT) for planning target wording and phonetic length, and then synthesize the corresponding speech tokens. To support this paradigm, we construct DuraSet-440K, a high-quality duration-aligned CoT corpus for supervised initialization. We further optimize the model with multi-modal multi-dimensional reinforcement learning, using a Duration Margin Reward to balance translation quality and duration consistency, and Modality-Aware Reward Attribution to assign rewards to appropriate token spans. Experiments on CVSS-T show that DuraS2ST achieves a strong balance between translation quality and duration consistency, outperforming competitive open-source and commercial baselines. Project page: https://github.com/Mia11939/DuraS2ST.

Medical question answering spans diverse specialties and modalities, and individual medical large language models (LLMs) exhibit distinct strengths across tasks and domains. This heterogeneity suggests that combining specialists may enable broader coverage of medical questions than relying on any single model. However, existing LLM routing methods primarily seek to balance answer quality and inference cost, leaving open how to exploit differences in specialist competence to improve medical reasoning. In this paper, we introduce MedRouter, an agentic system that uses an embedding-based multi-label router to select and query specialist LLMs, then passes their responses to a generator to produce the final answer. We further propose SCALE (Specialist Competence-Aware Learning), a two-stage training framework that first trains the Router with specialist correctness supervision and then optimizes its selections through reinforcement learning. The second stage uses a Performance Gain Reward (PGR) that measures how specialist information affects the generator's answer correctness relative to answering without that information. Experiments on eight text-based and multimodal medical QA benchmarks show that MedRouter outperforms the strongest routing baseline by 8% in average accuracy. Our analysis of specialist outputs further reveals distinct strengths and complementary question-level coverage, motivating learned routing to combine these capabilities for more comprehensive medical reasoning.

The pursuit of recursive self-improvement (RSI) toward general intelligence is divided between macro-level language model scaling and the interaction-driven principles of "Era of Experience". Yet, any self-improving architecture ultimately rests upon its underlying optimization engine: if general intelligence requires learning from grounded interaction, the reinforcement learning (RL) update rule itself must be capable of cumulative adaptation. While algorithm self-discovery has produced Disco103 that surpassed PPO to achieve SOTA benchmark performance — its internal update machinery remains an uninspected black box. We present the first causal mechanistic audit of a self-discovered RL rule, structured directly around the five pillars of the Era of Experience: extended horizon, grounded reward scales, continuing streams, within-lifetime change, and exploration depth. By surgically pinning, freezing, and transplanting recurrent states while holding meta-parameters fixed, we test when learning history acts as an asset or a burden. Three findings organize the audit: (1) Recurrent history actively expands usable reward scales, sustaining a six-decade window versus three under zero-pinning. (2) Decoupling historical content from its maintenance reveals that the penalty of mismatched history stems from perpetual clamping; allowing imported state to evolve naturally attenuates this burden. (3) Under environmental change, controlling replay retention reverses the apparent adaptation advantage over DQN, demonstrating that external data turnover can confound internal plasticity. Validated through capability thresholds and ported to a second rule (OPEN), this work grounds macro-RSI ambitions in micro-level learning dynamics, establishing a foundational audit standard for next-generation, self-evolving RL algorithms.

The rapid proliferation of highly realistic AI-Generated Content (AIGC) necessitates robust and interpretable detection mechanisms. However, existing detectors are predominantly confined to single modalities and provide binary outputs without reasoning. While Multimodal Large Language Models (MLLMs) present a promising solution, their development is constrained by the scarcity of multimodal reasoning data and the reasoning-detection optimization dilemma, where explicit reasoning supervision can compromise detection accuracy. To this end, we introduce IVT-Set, a comprehensive dataset comprising over 152K diverse image, video, and text samples equipped with multi-granularity Chain-of-Thought (CoT) reasoning trajectories. Based on it, we propose IVT-Guard, a pioneering framework for unified and interpretable AIGC detection across image, video, and text modalities. Furthermore, to overcome the aforementioned optimization dilemma, we design a novel three-stage training paradigm: Artifact-Aware Pre-training, Artifact-to-Evidence Supervised Fine-Tuning via artifact-aware injection, and Evidence-Verdict Consistency Group Relative Policy Optimization. Extensive experiments demonstrate that IVT-Guard achieves state-of-the-art detection performance across in-domain, out-of-domain, and cross-dataset settings while delivering faithful reasoning. Code and data will be released.

Group-relative reinforcement learning (RL) relies on reward variation among sampled responses to estimate informative relative advantages. As language models become increasingly capable, existing training data can become reward-saturated: all sampled responses to the same problem might receive equally high rewards, where the group-relative learning signals vanish and leave previously useful data obsolete. In this work, we investigate whether useful learning signals can be recovered from such saturated data. We study interventions at four levels of group-policy RL pipelines---data, rollout, reward, and advantage---and conduct extensive RL training on saturated reasoning data only. While standard GRPO on saturated data would almost always yield near-0 advantages and near-noise signals, diverse interventions successfully recycle and repurpose such data: among the proposed strategies, interventions at rollout generation are consistently most effective: nudging the policy to generate ``high-quality'', incorrect solutions introduces rollouts with poor rewards into saturated groups as negative samples, which turns out to improve GRPO by 6.4% to 9.0% across Qwen3-1.7B and 4B. Other interventions such as increasing rollout temperature or adding auxiliary rewards can also restore non-zero advantages, but yield less consistent gains. Further analyses show that effective negative rollouts require informative negative trajectories, that the method remains effective alongside unsaturated data, and that it supports iterative recycling of newly saturated examples. While increasingly stronger LLMs would render more data as saturated, our results demonstrate that don't waste your saturated data: with the right strategies they can be recycled into useful RL training signals in an increasingly data-scarce world.

Reward models (RMs) are a key component of large language model post-training, providing reward signals for subsequent reinforcement learning. However, conventional discriminative RMs typically output only scalar scores, making it difficult to identify the response behaviors associated with their scoring decisions. Existing interpretation methods often rely on predefined high-level attributes and require repeated counterfactual interventions for each response pair to validate candidate explanations, lacking a closed-loop mechanism that uses RMs' feedback to train a reusable explainer. To address this, we propose RewardExplainer, a framework that obtains feedback from the target reward model through counterfactual rewriting and uses this feedback to further optimize the explainer. RewardExplainer generates open-ended, atomic, and intervenable natural-language scoring mechanisms, making explanations more concrete, readable, and actionable. It further converts counterfactual feedback into preference supervision, enabling the explainer to more faithfully capture the target RM's scoring preferences and sensitive behaviors than single-pass generation. Extensive experiments across multiple target RMs and explainer backbones show consistent improvements. Beyond interpretation, we use the generated mechanisms to identify potential bias patterns and construct targeted debiasing data for fine-tuning the reward model, improving robustness on reward-hacking benchmarks.

Reinforcement Learning from Human Feedback (RLHF) is a powerful technique for aligning large language models (LLMs) with human preference. However, it often suffers from the reward hacking issue, where policy optimization improves the proxy reward model while actually degrading performance with respect to the true human preference, due to the imperfection of the proxy. To address this, we propose Reward Model Boosting (RMB), a novel approach that enhances the robustness and reliability of the reward signal for RLHF. RMB first trains a set of reward models with a diversity-promoting regularizer. This encourages each model to learn complementary aspects of the reward landscape. Then, RMB learns a lightweight aggregator in the principle of boosting to aggregate the outputs of the diverse reward models into a more accurate and robust reward signal. Our extensive experiments demonstrate that RMB significantly improves reward accuracy on both in-distribution and out-of-distribution datasets, substantially mitigating the reward hacking issue and ultimately improving RLHF performance.

Training efficiency has become the central driver of recent progress in foundation models. To overcome the massive computational and data requirements of large-scale training, researchers increasingly adopt strategies such as selective data sampling, efficient pre-training, and simplified reinforcement learning pipelines. While these strategies drastically reduce overhead, they prompt a critical, yet neglected question: Is efficiency achieved at the expense of model robustness and security? To our knowledge, we present the first systematic cross-domain investigation of the efficiency-vulnerability trade-off. Across vision and language models, we show that efficiency-oriented training increases susceptibility to adversarial and privacy attacks. We characterize this vulnerability by analyzing the models' internal geometry and functional representations, demonstrating that the evaluated efficient variants consistently exhibit sharper loss geometry together with systematic changes in representational structure. We further extend our analysis to "zero RL training", finding that models trained using simplified RL recipes exhibit substantially greater susceptibility to catastrophic forgetting and more pronounced overconfidence than those trained through conventional alignment pipelines. Our findings suggest that training efficiency is rarely a "free lunch"; rather, the mechanisms that minimize computation can inadvertently compromise safety. We conclude by calling for a paradigm shift toward multi-objective training that jointly optimizes for performance, cost, and security.

Reasoning distillation from powerful teacher models to smaller students faces the Gap Curse: as teachers grow more sophisticated, their complex distributions increasingly diverge from what students can approximate, causing performance degradation. Existing mitigation strategies either filter out challenging examples through data selection or introduce weaker intermediate assistant models, inherently compromising supervision coverage or quality. We propose Teacher Alignment, which directly adapts the teacher toward the student's distribution without discarding data or degrading reasoning quality. However, naive alignment through standard knowledge distillation triggers catastrophic collapse of the teacher's reasoning capabilities. To address this, we reformulate teacher alignment as reinforcement learning and introduce TeacherGRPO, built on Group Relative Policy Optimization with two key innovations: (i) Curriculum Selective Alignment applies dual token- and distribution-level curricula to focus rewards on high-signal reasoning gaps while filtering noise from trivial tokens and uncertain tail distributions, and (ii) Importance-Adaptive Length Regularization selectively penalizes verbose redundancy while preserving pedagogically critical reasoning steps. The aligned teacher then distills knowledge to students via standard pipelines. Extensive experiments show TeacherGRPO significantly outperforms baselines across diverse reasoning benchmarks and distillation methods. Our code is available at https://github.com/LzyFischer/TeacherGRPO.