RL for Language Models

We study reference-free post-training for multilingual machine translation with open large language models. Starting from the supervised-finetuned MiLMMT-46-v0.1 models, we apply Group Relative Policy Optimization (GRPO) with a reward that averages two reference-free quality estimation models and is gated by language identification. We then linearly interpolate the supervised fine-tuning (SFT) and reinforcement learning (RL) model checkpoints to obtain MiLMMT-46-v1.0. Across 46 languages, the resulting models consistently improve translation quality over their SFT counterparts, outperform strong recent open baselines, including Seed-X, HY-MT2, and TranslateGemma, and achieve leading reference-free scores against evaluated proprietary systems such as Google Translate, Gemini 3 Pro, and GPT-5. We further investigate on-policy distillation (OPD) and find that it reaches, but does not surpass, the quality frontier achieved by RL with checkpoint interpolation. We release the models and code to facilitate future research.

Agent skills represent a standardized format for packaging procedural knowledge and domain expertise, serving within agent harness systems as an essential mechanism to continually constrain a language model's behavior space for repeatable, high-quality task execution. However, because strong closed-source models entail high inference costs, current popular agent harnesses, such as Codex and OpenClaw, remain prohibitively expensive when deploying these skills to accomplish real-world tasks. The rapid capability enhancement of open-source models deployable on consumer-grade GPUs presents a compelling opportunity to drastically reduce these costs by leveraging skill-based behavioral constraints. Nevertheless, automatically generating effective skills tailored specifically for such compact models remains a significant practical challenge. To address this, we propose SKILLER, a natural-language-driven reinforcement learning framework designed to automatically generate executor-specific skills for small models, which employs a strong model as the actor and critic, treats the small-model agent system as the environment, and propagates all reinforcement learning signals entirely via natural language. Extensive experimental evaluations across five relevant benchmarks using Qwen3.5-9B and Qwen3.5-4B demonstrate that SKILLER outperforms three open-source and one closed-source skill generation or evolution methods, achieving absolute gains ranging from 4.3 to 20.4 percentage points for the 9B model and 1.8 to 13.3 points for the 4B model, while remarkably matching the performance of strong closed-source models on single-skill tasks in SkillsBench. The project is available at https://github.com/DANG-ai/SKILLER.

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

Due to the diversity of real-world time series, no single forecasting model consistently dominates across all samples. Ensemble learning addresses this by combining complementary model strengths, yet existing methods rely on fixed rules or black-box models based solely on numerical inputs, failing to leverage LLM reasoning for interpretable weighting decisions. We propose REATS, which leverages LLM reasoning capabilities as an intelligent ensemble router that jointly processes textual temporal pattern descriptions and numerical features to produce interpretable, sample-adaptive ensemble weights through chain-of-thought reasoning. To enable effective LLM-based ensembling, we study its key design choices and propose: (i) a structured input pipeline that transforms raw time series into hybrid textual--numerical representations with fixed token cost, enabling rule-based chain-of-thought construction without API dependency, augmented with retrieved similar-sample priors; (ii) a diverse multi-row weight supervision scheme coupled with a token-efficient percentage-table format that reduces numerical complexity and mitigates LLM hallucinations; and (iii) a two-stage fine-tuning framework combining SFT with GRPO, where a reciprocal reward mapping transforms the continuous unbounded MSE gap into bounded signals with amplified near-oracle sensitivity, addressing the uniform sensitivity and outlier-dominated advantage compression inherent in naive reward designs for regression-based GRPO. Experiments on eight benchmarks demonstrate that REATS outperforms competitive ensemble baselines while providing natural language explanations and demonstrating strong transfer learning and out-of-domain generalization to unseen candidate models.

On unlabeled test data, reinforcement learning lacks a ground-truth reward; test-time RL methods derive one from the model's own roll-outs, rewarding those that match the majority vote over $N$ sampled answers. That vote discards a correct answer whenever it is a minority and scores every majority-matching roll-out identically. We replace it with \emph{CoRE} (Consensus Rewards via Equilibrium): the $N$ roll-outs form a graph whose edges combine answer agreement, reasoning similarity, and generation confidence, and replicator dynamics extract its dominant set, yielding a refined pseudo-label, a graded per-roll-out reward, and a per-question cohesiveness gate. CoRE strictly generalizes voting: majority voting is recovered as a special case; a block-value analysis gives a sharp threshold for when consensus recovers a correct minority against a larger wrong plurality; and confidence calibration provably lowers that threshold multiplicatively. Across seven backbones and five benchmarks (42 model--benchmark cells, three seeds each), \emph{CoRE} improves the untrained base by $+21.7$ points on average versus $+20.4$ for majority-vote TTRL, wins wherever agreement is contestable with margins over the vote of up to $+7.5$ points, and reaches the voting baseline's plateau accuracy in $54$--$70$% fewer steps. Consensus, not counting: treating the roll-out group as a graph rather than a ballot box turns a brittle vote into a calibrated, graded, self-supervised reward at no extra roll-out cost.

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

Recent advances in reward modeling show a paradigm shift from discriminative reward models to generative reward models. However, despite their strong capabilities in response ranking, generative reward models have not realized their potential in reinforcement learning (RL). Our analysis reveals that this limitation arises from a mismatch between the comparative nature of generative reward modeling and the scalar scoring paradigm adopted by existing RL algorithms. To bridge this gap, we propose a Ranking-based Reward Construction (RRC) approach, which enables generative reward models to provide more effective RL learning signals by deriving rewards from relative preference rankings. RRC introduces two complementary strategies: self-competitive ranking, which exploits comparisons among sampled responses, and anchor-guided ranking, which enables scalable ranking-based reward construction with a small set of reference responses. Experiments across open-ended chat and reasoning benchmarks demonstrate that RRC substantially improves RL training with generative reward models, achieving consistent gains over existing reward construction approaches. Our code can be found at https://github.com/wangclnlp/RRC.

Critic-free group-based reinforcement learning has become a scalable approach for post-training large language models. However, most existing methods allocate the same number of rollouts to every task and trajectory state, even though some rollouts provide much more useful learning signals than others. Recent work has started to treat rollout generation as an adaptive decision, but two important limitations remain. First, intervention strategies are often based on fixed heuristics and therefore cannot adjust as the policy changes during training. Second, these methods usually decide only how many rollouts to generate, without explicitly controlling where and how to intervene. To address these limitations, we propose Recoverability-Aware Intervention Learning (RAIL), a training-time framework that learns how to generate rollouts based on the improvement produced by each intervention. RAIL models intervention selection as an online contextual-bandit problem and trains a recoverability controller using intervention traces collected through a shadow-to-live procedure. This allows the controller to keep learning while the underlying policy evolves. We evaluate RAIL in terms of effectiveness, adaptivity, expressiveness, and efficiency. Across multiple settings, RAIL consistently improves performance under limited rollout budgets. These results show that recoverability-aware intervention provides a principled way to generate more informative and less redundant rollouts, leading to stronger learning signals during post-training.

We tackle the challenging yet underexplored task of Generalized Referring Expression Comprehension (GREC), which requires a model to localize the object described by a textual expression when it exists (positive sample) and to refuse output when it does not (negative sample). Although Multimodal Large Language Models (MLLMs) excel at localizing existing objects, they often fail to reject nonexistent ones due to the absence of negative samples during training, producing hallucinated bounding boxes. Existing post-training approaches such as supervised fine-tuning (SFT) and reinforcement learning (RL) enhance refusal behavior but usually degrade localization accuracy on positive samples, undermining the model's core competence. To address this, we propose Refusal-Calibrated Group Relative Policy Optimization (RC-GRPO), a calibrated RL strategy that strengthens the refusal ability of MLLMs while preserving localization performance. It enforces "None" outputs in rollouts for valid advantage estimation on negative samples and applies a penalty to prevent over-refusal on positives, achieving a balanced trade-off between accuracy and reliability. A second-stage reasoning reinforcement further consolidates causal understanding and interpretability. Experiments on three GREC benchmarks demonstrate that RC-GRPO attains superior localization accuracy while maintaining strong refusal capability.

Generating a rigorous paper introduction with large language models (LLMs) remains challenging, since it requires coordinating background, gap identification, method and contribution within a coherent narrative. Existing solutions externalize this process as multi-stage prompts or agent workflows which are expensive and vulnerable to cross-stage drift. We propose StructPO, a struct-aware policy learning framework that internalizes the entire multi-stage writing workflow into a single-pass policy controlled by explicit stage tokens. StructPO introduces struct-aware credit assignment to decouple local stage quality from global coherence and refinement-guided optimization to internalize revision behavior into the first-pass policy. Experiments show that StructPO improves semantic alignment, structural rationality and inference efficiency over workflow-based baselines, generalizes to out-of-domain settings, and remains competitive with GPT-5.1 in human evaluation when scaled to Qwen3-32B. These results show that internalizing academic writing workflows through fine-grained policy optimization offers a viable alternative to costly external orchestration.

Spoken emotional dialogue requires a model to understand a user's spoken input and generate a response that is both semantically appropriate and emotionally expressive. This is challenging because communicative intent may be stated explicitly in lexical content or conveyed more implicitly through paralinguistic cues, which can complement or diverge from the words themselves. However, two limitations constrain progress in this area: the scarcity of benchmarks that distinguish these intent expressions, and the lack of reinforcement learning objectives that jointly account for response quality and emotional expression. To address the lack of suitable benchmarks, we introduce ParaIntent, a Chinese benchmark comprising 14 intent categories with balanced explicit and implicit samples, together with a multidimensional evaluation protocol covering intent fulfillment, response quality, and emotional expression. For policy optimization, existing approaches either use a shared objective for text and speech or apply reinforcement learning to only one modality, leaving modality-specific learning signals entangled within policy optimization. Motivated by this, we propose Acoustic-Lexical Decoupled Policy Optimization (ALPO), which computes independent textual and acoustic advantages and routes them to the corresponding text and speech tokens within a unified rollout. Under identical reward functions and training budgets, ALPO improves over standard GRPO on most automatic metrics and achieves the best subjective results among the fine-tuned variants, with particularly clear gains in emotional expressiveness on both the synthetic and human-recorded test sets.

Search-augmented LM agents are typically trained with a binary exact-match reward, which throws away most of what a failed trajectory tells us about why it failed. We introduce HindSearch, a hindsight self-distillation procedure for GRPO: after each rollout, a frozen judge writes a short critique of every failed trajectory using the gold answer, and the critique supplies an auxiliary on-policy distillation signal on the student's search actions. On the standard seven-benchmark suite with Qwen2.5-3B-Instruct, HindSearch reaches 39.4% average EM, outperforming prior search-RL baselines. Removing the judge's access to the gold answer erases most of the gain, isolating hindsight as the source of the improvement.

Modern large language models (LLMs) are expected not just to answer correctly, but to adapt their behavior to different human values and use cases. As a result, multi-reward reinforcement learning (RL) has become an increasingly important problem for LLMs, where each reward captures a different aspect of desired behavior. However, optimizing with multiple rewards suffers from a more severe alignment tax issue, where different optimization objectives can trade off or even conflict with each other, leading to unstable and inefficient post-training. In this work, we propose PRISM, a new multi-reward RL framework built upon the idea of policy-space decomposition and composition. Instead of compositing different rewards, PRISM optimizes a set of standalone positive policies and a global negative policy. This alleviates the potential conflict during multi-reward policy optimization, while enabling controllability during inference by flexible policy composition. Experiments on scientific reasoning, tool-use reasoning, and helpfulness-safety alignment show that PRISM consistently outperforms existing multi-reward RL baselines, with extra controllability for inference-time preference control.

Supervised fine-tuning (SFT) can equip large language models (LLMs) with domain knowledge for high-performance computing (HPC) tasks such as data race detection and benchmark question answering. However, knowledge alone does not guarantee task-appropriate behavior: the same SFT model that correctly classifies 88.65% of C/C++ data race samples produces verbose, imprecise answers to factual queries, with 65.9% of MLPerf responses exceeding 40 characters. Reinforcement learning (RL) post-training addresses this gap by optimizing for task-specific rewards rather than token-level imitation. Yet HPC tasks exhibit extreme heterogeneity, with binary classification, factual QA, and semantic generation differing by 58x in answer length, spanning three distinct reward distributions, and showing widely varying SFT accuracy. This makes uniform-weight RL methods such as GRPO suboptimal. We propose HARGO, Heterogeneity-Aware Reward-Guided Optimization, which introduces per-response importance weighting via confidence-modulated advantage: computing a discrimination signal from group-level reward contrast and a confidence signal from reference model log-probabilities, then modulating the advantage before computing per-response weights, without requiring task-type labels. Across four HPC tasks and nine methods, HARGO achieves the best performance on all three primary metrics: WinRate 54.62%, Data Race F1 91.30%, and PLP Similarity 0.8558. Ablation confirms complementary contributions from both signals. HARGO establishes the best overall alignment quality among compared methods for heterogeneous HPC tasks.

Recent advances in Generative AI have substantially improved financial sentiment analysis through post-trained financial large language models (LLMs). However, existing approaches remain confined to a market-agnostic, supervised learning paradigm that relies on limited, static and human-annotated datasets, and thus are incapable of adapting to evolving market conditions. To address this limitation, we introduce FinSMART, the first market-aligned reinforcement learning framework for financial sentiment analysis, which directly optimizes sentiment signals using realized market outcomes. To deal with the noisy, non-stationary, and multifactorial nature of financial markets, FinSMART incorporates a signal extraction pipeline that combines market-aware data filtering with a discrete asymmetric trading reward, enabling stable reinforcement learning from economically meaningful market feedback. Experimental results demonstrate that FinSMART significantly outperforms existing state-of-the-art methods in profitability, risk-adjusted performance, and sentiment signal quality, improving cumulative trading returns by 220% over the strongest baseline. Uniquely, the FinSMART framework naturally supports market-aware retraining, at any point in time, by replacing costly manual annotation with newly observed financial articles and their realized market outcomes. Such a retraining strategy enables the model to continuously adapt to changing market dynamics, resulting in consistent performance gains over its static counterpart. These findings demonstrate the practical applicability of market-aligned reinforcement learning and highlight its potential as a next-generation paradigm for developing adaptive financial LLMs.

The alignment of Small Language Models (SLMs) in the 70--500M parameter range using reinforcement learning is often considered unstable, though the underlying failure mechanisms have not been systematically investigated. In the State-of-the-Art (SOTA) research, fifteen (model, corpus) configurations were trained using Proximal Policy Optimization (PPO). The experiments included Pythia-70M, 160M, 410M and SmolLM2-135M, 360M on the TinyStories, CNN/DailyMail, and Wikitext-103 corpora. Three reproducible failure modes were identified in small-scale language models: silent LoRA parameter freezing in standard PEFT/TRL pipelines, numerical overflow in importance ratios when using bfloat16, and catastrophic policy collapse due to reward-model error. These issues were addressed using a merge-and-reinitialize adapter technique, float32 precision during PPO updates, and a three-layer safety mechanism comprising reward whitening, importance-ratio guarding, and weight rollback. In this paper, a capacity-headroom hypothesis is proposed, which states that PPO performance at the SLM scale depends on both a fluent supervised model ($\text{PPL}<20$) and a discriminative reward signal, rather than on the number of model parameters. The proposed system converged stably in all experiments and improved preference win rate over the SFT baseline in configurations with a fluent prior and an informative reward signal. Furthermore, it outperformed instruction-tuned baselines while requiring significantly less training data. All checkpoints, preference datasets, and training scripts are publicly released$^§$.

Language model alignment aims to make model behavior reliably reflect desirable properties such as helpfulness, safety, and instruction following. Current approaches typically use supervised fine-tuning on demonstrations or reinforcement learning with rewards derived from verifiers or human feedback. These paradigms leave an important question underexplored: can demonstrations alone yield an implicit reward that can be inspected, reused, and optimized on-policy to align AI? Motivated by inverse reinforcement learning, we introduce Projected Alignment Reward Estimated from Demonstrations (PARED). PARED recovers the implicit reward underlying expert demonstrations as an explicit function over a small set of response-level features, learned by a lightweight discriminator that separates demonstrations from the policy's own samples in this feature space. Unlike a standard reward model, PARED requires no task-specific preference annotations: demonstrations provide the task-specific supervision, which can be augmented with AI feedback as additional dimensions of supervision. Through experiments involving inference-time reranking and adversarial on-policy RL, we show that the recovered reward improves a base policy without a supervised loss and yields further gains when optimized after standard supervised fine-tuning. Additionally, we demonstrate that PARED can be used for contextual alignment, in which a single policy can be tailored to the preferences of different audiences.

Reinforcement Learning (RL) training for Large Language Models (LLMs) often suffers from instability due to the discrepancy between training and inference. This training-inference discrepancy stems from two primary factors: an architectural separation between training and inference engines, and the use of low-precision quantization in inference versus higher-precision computation in training. To address training instability issues caused by high training-inference discrepancy, we present the principles and methods for its adaptive control. We propose Adaptive Control Reinforcement Learning (ACRL), which adaptively maintains the training-inference discrepancy within a reasonable range to ensure stable RL training. Beyond stabilization, ACRL inherently increases policy entropy, thereby enhancing exploration and improving accuracy. The experimental results show that when the inference engine utilizes FP8 quantization, ACRL consistently maintains the training-inference discrepancy within a reasonable range and stabilizes RL training. Furthermore, ACRL not only matches the accuracy of the BF16 baseline but also outperforms importance sampling (IS) fixes.

Large language models show strong promise for information extraction (IE), but existing reflection-based correction methods are often misaligned with structured extraction outputs. Free-form self-reflection can flag an error, yet it rarely identifies whether the failure is a missing span, wrong label, boundary mismatch, invalid relation type, or reversed argument order. We introduce LA-RL (Label-Aware Reflective Reinforcement Learning), an outcome-supervised framework that guides IE self-correction with task-grounded diagnostic labels. A single backbone first predicts an extraction, diagnoses task-specific error labels, and then revises its output conditioned on the diagnosis. Training starts from diagnostic data labeled by an annotation model for cold-start supervised fine-tuning and proceeds through two GRPO stages that reward final extraction quality, format validity, and first-pass correctness, without a process reward model. Experiments on named entity recognition, relation extraction, and event extraction show consistent same-backbone gains over SFT, including 6.83 average F1 on SciER relation extraction, about 20 F1 on out-of-distribution relation extraction, and 14.80 trigger F1 plus 17.50 argument F1 on DuEE1.0. Ablations show that reflection structure is task-sensitive: stronger constraints benefit relation extraction, whereas named entity recognition needs less restrictive correction under domain shift.

Recent work has shown that large language models (LLMs) can iteratively improve their outputs by incorporating generated samples and their corresponding evaluation scores as in-context examples. Despite these empirical findings, the theoretical foundations underlying this phenomenon remain poorly understood. In this paper, we show that score-conditioned In-Context Learning (ICL) admits a structural correspondence to policy gradient optimization. We first provide a constructive proof that self-attention mechanisms can implement reward-weighted aggregation analogous to the REINFORCE algorithm under specific weight matrix configurations, and discuss the relationship between this construction and the behavior of pretrained transformers. The correspondence is directional in hidden-state space and holds exactly only under the stated simplifying conditions; we quantify its strength empirically. Within our simplified hidden-state model, we furthermore derive an exact upper bound on the distribution shift induced by a bounded attention update, yielding a trust-region-like analogy to KL-constrained policy optimization. We validate our theory through extensive experiments across multiple LLMs, demonstrating that LLMs effectively utilize score information to shift output distributions toward high-scoring exemplars, and that attention weights exhibit a strong correlation with example scores.

Model merging plays a crucial role in consolidating multiple specialized models into a single, unified model, especially in the era of large language models (LLMs). Recent research has primarily focused on developing strategies to enhance merging performance with the trained models, while the impact of training paradigms, such as supervised fine-tuning (SFT) and reinforcement learning (RL), on the effectiveness of model merging remains underexplored. In this study, we systematically explore the merging behavior of RL-trained LLMs compared to those trained with traditional SFT. Through comprehensive evaluations across five representative tasks, we find that RL significantly reduces task conflicts and results in less performance degradation after merging, making RL-trained models particularly well-suited for this process. To unearth the reasons behind the superior suitability of RL for model merging, we conduct extensive empirical experiments and theoretical analyses. Our findings highlight three key factors: (1) On-policy training data in RL control the gradient updates in a smaller magnitude, reducing the risk of overwriting existing knowledge for other tasks in the model. (2) The RL optimization objective, which favors ``\textit{enough is as good as a feast}", progressively reduces the magnitude and the number of conflict parameter updates as the model converges. (3) Joint optimization of positive and negative examples in RL steers the model towards an unbiased task-specific parameter subspace, ensuring robust performance while further preventing parameter conflicts.

Multimodal large language models (MLLMs) have achieved impressive performance in multimodal emotion recognition (MER) tasks and lifted MER to a new level that is complex emotion understanding with advanced video understanding abilities and natural language description. However, existing MLLM-based methods often use a fixed prompt to perceive the emotions, ignoring the dynamicity and complexity of the emotion source in the multimodal inputs. To address these issues, we propose a novel Reinforcement Learning-based Dynamic Agent Specialization framework (\textbf{EmoAgent-R1}) to optimize the emotion recognition, reasoning, and generalization abilities of an MLLM with dynamic agent specialization based on reinforcement learning. Specifically, we first adopt a cold start strategy to endow an MLLM with preliminary emotion recognition, reasoning, and agent routing ability by training with synthetic answer-conditioned chain-of-thought data and agent routing data. Then, we further train the MLLM with reinforcement learning to perceive emotions in a two-step agentic workflow with agent selection and agent specialization. To effectively train EmoAgent-R1, we propose a novel Progressive Group-Relative Policy Optimization (P-GRPO) to combine group-based relative advantages with a PMI-inspired progressive token-level modulation to transform sparse rewards into fine-grained learning signals, mitigating the coarse-grained uniform credit assignment issue in GRPO. Extensive experiments on MER benchmarks demonstrate the superiority of our EmoAgent-R1 in stronger emotion reasoning performance and improved optimization stability.

Retrieval-augmented large language models frequently face contexts that interleave useful evidence with misleading statements or instruction-like content. Blanket refusal discards valid evidence, whereas uncritical adoption yields incorrect or unsafe answers. The ability to selectively adopt relevant information while rejecting deceptive or harmful content is therefore critical for reliable deployment in real-world retrieval settings. We introduce SelectBench, a controlled benchmark and training set for selective evidence adoption, and post-train Qwen3.5-4B directly with DAPO using either deterministic rule rewards or a frozen semantic judge. On the corrected 325-example SelectBench-v2 test set, strict success rises from 22.46% for the original checkpoint to 25.54% with DAPO-Rule and 26.46% with DAPO-DeepSeek. Both trained policies reduce forbidden-content adoption and produce shorter, more focused responses, yet prompt-injection following does not improve. The paired gains are modest and fail to survive Holm correction, suggesting that stronger reward shaping or additional training iterations may be needed for more robust gains. DAPO-DeepSeek exhibits no material degradation on MMLU or clean HotpotQA, indicating that the post-training procedure preserves general capabilities. These results demonstrate a directional improvement in selective evidence use, while identifying injection resistance and statistical robustness as important remaining challenges for future work.

Post-training with evaluator feedback on policy-induced samples serves as a major mechanism for improving large language models. As policies improve, these sampled responses become close in quality. These close candidates create a bottleneck for policy optimization: collapsed relative evaluator score gaps yield weak or misleading policy supervision. We theoretically characterize why these gaps matter through a probability allocation view, showing that the directional gain of shifting probability mass from one response to another is exactly the evaluator score gap between them. This identifies relative score gaps as the policy optimization signals that guide updates. Motivated by this view, we propose DynamicRubric, a response-set-conditioned evaluator--policy co-evolution framework that generates weighted binary rubric items for each candidate set and aggregates the resulting judgments into response-level scores. In our experiments with 8B backbones, DynamicRubric improves evaluator performance and provides stronger policy supervision than baselines using a 70B reward model or a 235B static rubric generator. DynamicRubric-optimized policies also show gains on verifiable reasoning and coding tasks. A DynamicRubric-optimized model is fully deployed in WeChat Search's AI answering scenario, where it serves all online traffic across tens of millions of requests per day and improves key online metrics. These results suggest a principle for evaluator-guided post-training: evaluators should evolve with the policies they supervise.

Reinforcement learning with verifiable rewards (RLVR) has been established as a viable paradigm for the post-training of Large Language Models (LLMs), including downstream tasks, such as Neural Machine Translation (NMT). With the latest research indicating that RLVR could be the preferred training method for translating legal documents due to the induced reasoning capabilities, it raises the question whether it is really attributed to the reasoning or more generally to the training paradigm. We investigate the importance of including the model's reasoning trace in the generated responses during both training and inference by systematically omitting it from one of the phases. Our experiments show that including the reasoning, specifically during inference, has a positive effect on the overall translation quality. Furthermore, we recognise that the reasoning leads to an increase in output tokens, hence we study the cost-quality tradeoff between the increased computational demands and the improved translation quality.

Neural machine translation (NMT) in the legal domain is a linguistically and conceptually demanding task, primarily due to the complexity of legal language and the high level of precision it requires. The recent emergence of reasoning-capable language models opens new possibilities for tackling such challenges. They add to a set of other previously proposed techniques to enhance the translation quality, which includes supervised fine-tuning and reinforcement learning. In this work, we perform a comparison between these various approaches. More particularly, we evaluate small language models such as Qwen3.5 4B, Qwen3.5 9B, and Gemma 3 12B enhanced with various re-training paradigms and compare their performances against frontier reasoning models. We focus on the Swiss legal system, which -- with its unique multilingual statutes -- offers a particularly challenging testbed for reasoning-augmented models. Our results show that the quality of small ``base'' models can be greatly enhanced, and that reinforcement learning with verifiable rewards can be applied to NMT in the legal domain and surpasses the translation quality of supervised fine-tuning. The performance of enhanced small models is close to the one of state-of-the-art reasoning models yet remains inferior. We also note that re-training paradigms yield diminishing returns as model size increase. The code and models are publicly available at https://github.com/aixiuxiuxiu/Legal-MT-SFT-RL.

Large-scale online reinforcement learning (RL) is the predominant means of eliciting advanced abilities including long-term reasoning and agentic tool use in large language models (LLMs). However, continuing to scale it across vast task domains of interest remains challenging in both computational infrastructure and cost, especially when considering RL as merely a one-off learning stage. Recently, a widely used technique for distilling knowledge across various domains and training stages, multi-teacher on-policy distillation (MOPD), helps to decouple the RL stage, saving costs, while maintaining generality across vast domains. Nonetheless, similar to online RL, MOPD requires coupled inference and backward passes, which continues to limit its scalability and computational efficiency. To address these challenges, we propose REGEN: Replay-recycling for Expert-to-Generalist Distillation with Offline RL. Instead of distilling from multiple teacher models, REGEN trains a generalist by simply recycling the replay memory -- the free by-product of the teachers' specialized RL training -- and employing offline RL algorithms. REGEN completely decouples the rollout sampling from the backward training process and thus greatly reduces the training cost. Across mathematical reasoning, code generation, and instruction following, REGEN matches the accuracy of MOPD at substantially lower cost. It potentially turns online RL into a data synthesis process instead of a one-off learning stage, and can be extended to large-scale post-training without requiring heavy computational load. Code is available at https://github.com/yunjie-sysu/REGEN.

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.

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 \texttt{Llama-3.2-1B-Instruct} on Anthropic \texttt{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.

Entropy control has become an effective tool in reinforcement learning (RL) of large language models (LLMs), helping balance exploration-exploitation trade-off during alignment process. Such RL paradigm is often conducted on mixtures of heterogeneous tasks, which induce distinct entropy regimes under the same policy, making global or token-level entropy regulation insufficient to corresponding heterogeneous needs of exploration. This heterogeneity further makes GRPO-style normalized advantages induce an entropy-dependent bias, making advantage signals across prompt groups statistically non-comparable. To address this issue, we propose Group Entropy-Controlled Policy Optimization (GEPO), a lightweight extension to GRPO that uses group entropy, estimated from existing grouped samples to perform entropy-conditioned asymmetric advantage shaping. GEPO attenuates positive advantages in low-entropy groups to reduce over-exploitation, and negative advantages in high-entropy groups to preserve exploration, with adaptive thresholds derived from historical entropy statistics. Extensive experiments on two base models across thirteen benchmarks spanning mathematics, physics, science, code generation, and instruction following show that GEPO consistently outperforms GRPO and recent entropy-controlled methods, delivering balanced cross-task improvements while preserving task-specific exploration levels throughout training.