Modern large language models (LLMs) rely on reinforcement learning to build strong capabilities in individual domains, but integrating those capabilities into a single deployable model remains challenging. By routing each sample to the teacher whose domain matches it, existing approaches let a domain label decide which teacher provides supervision. However, domain expertise holds only on average: the matched teacher is not always correct on a given sample, while a teacher from another domain sometimes is. The reliable teacher therefore has to be identified per sample, not per domain. In this paper, we introduce Multi-Teacher Self-Distillation Policy Optimization (MT-SDPO), an on-policy distillation method that unifies several frozen teachers into one student model. MT-SDPO consists of three components: (1) self-anchors, where a rollout is supervised by a correct rollout from its own group; (2) answer-verified eligibility, where a teacher may supervise a sample only if its own answer passes a verifier; and (3) privileged distillation, which merges the anchor and all verified feedback into one context that an exponential moving average self-teacher reads and the student does not, thereby keeping one policy at deployment. Across five students from three model families, MT-SDPO lifts the weakest domain of Qwen3-8B by 14.79 points and narrows its domain gap by 74.7%, a better balance than serving one matched teacher per domain. Verified reliability, not domain membership, should decide who teaches. Code is available at https://github.com/hexixiang/MT-SDPO.
Meicong Zhang, Tiancheng Su, Jiahao Cheng +3cs.CL cs.AI
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.
Beining Wang, Weihang Su, Hongtao Tian +8cs.LG cs.AI
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.
On-policy self-distillation (OPSD) has emerged as a practical method for training large language models (LLMs) to reason, where a single model acts as both the teacher and the student with different levels of information access. However, recent studies have found that the teacher's dense token-level supervision, conditioned on privileged information, can lead to overfitting to in-domain patterns, suppress exploration, and hurt cross-domain generalization, while also introducing a more fundamental issue: *privileged information leakage*, where the student encodes answer-dependent shortcuts that are unavailable at test time. We introduce **DemoPSD**, a novel framework that resolves such problems through the idea of *selective adoption of teacher guidance*. Instead of fitting the full teacher distribution, DemoPSD steers the student toward a *reverse-KL barycenter target*, a weighted geometric combination of the teacher and student distributions, that naturally balances learning from the teacher with preserving the student's own reasoning capacity. We measure the difference between their distributions and use such a discrepancy to adaptively control the blending at each token position. We provably show that DemoPSD achieves **(1)** *leakage attenuation*, i.e., effective mitigation of privileged information leakage; and **(2)** *exploration preservation*, i.e., preservation of exploration capacity under dense token-level distillation. Extensive experiments on SciKnowEval across four scientific fields show that DemoPSD outperforms both GRPO and SDPO while maintaining higher training entropy and robustly generalizing to out-of-distribution GPQA benchmarks.
Building general-purpose role-playing agents that faithfully portray any character from a natural-language profile remains challenging. The dominant paradigm -- supervised fine-tuning -- encourages behavioral mimicry without deep, human-like internal thought processes, resulting in poor out-of-distribution generalization. Therefore, we propose \textbf{Psy-CoT}, a psychology-grounded chain-of-thought framework that decomposes pre-response reasoning into three role-specific steps -- \emph{Interaction Perception}, \emph{Psychological Empathy}, and \emph{Logical Construction} -- so that the model \emph{thinks dynamically} from the profile rather than merely mimicking surface patterns. While structured reasoning provides a foundation, it alone is insufficient; reinforcement learning is essential to further align the model with character fidelity. However, we observe that under LLM-based reward models, both generic phrases that hack the reward model and genuinely role-specific phrases receive identical gradient signals -- this hacking accumulates over training, misleading the model into treating both as equally optimal choices. To address this, we propose \textbf{Role-Aware Policy Optimization (RAPO)}, which uses profile--token mutual information to weight gradients asymmetrically -- amplifying role-specific tokens under positive advantage while attenuating them under negative advantage. Experiments on CoSER, CharacterBench, and CharacterEval demonstrate that Psy-CoT outperforms existing role-playing CoT methods, and RAPO consistently surpasses GRPO across multiple model scales.
Distilling reasoning capabilities from strong to weak language models typically involves imitating specific solution trajectories, effectively transferring what to answer rather than how to reason. This trajectory-level imitation encourages memorization of instance-specific steps rather than acquisition of transferable problem-solving skills, limiting generalization to novel problems. We propose Strategy-Guided Policy Optimization (SGPO), which replaces instance-level trajectory imitation with reusable strategy distillation. SGPO extracts structured strategy descriptions from strong-model responses and, for each problem, constructs both autonomous and strategy-guided trajectories to enable direct comparison of the model's behavior with and without strategic guidance. The framework then addresses two key questions. For how to distill, a token-level forward-KL objective selectively transfers the distributional shift induced by strategy conditioning into the unguided policy, with proximal constraints ensuring stability. For when to distill, adaptive instance-level weighting strengthens guidance when autonomous exploration falls short and reduces it as the model's own competence grows. Experiments on four mathematical benchmarks across two model families show that SGPO consistently outperforms SFT, on-policy RL, and hybrid-policy baselines, improving the average score by 2.2 points over the strongest baseline on Qwen2.5-7B-Instruct. Analysis reveals that the forward-KL objective provides an inherently selective distillation signal that outperforms direct trajectory imitation, and that strategy distillation exhibits complementary scaling with base model capability.
Masked diffusion language models decode by iteratively unmasking tokens, where the unmasking order defines an "order of thought" that strongly influences generation quality yet is typically chosen heuristically. We derive a tractable upper bound on the sequential decoding mismatch, measured by the Kullback-Leibler divergence and expressed in terms of the model's pathwise log-likelihood, with tightness under sufficient model expressivity. This bound induces a dense self-aware reward over ordered trajectories, casting order selection as a principled policy optimization problem with a frozen denoiser. We instantiate this idea as Self-Aware Scheduling (SAS), which learns a lightweight order policy using Group Relative Policy Optimization and applies seamlessly to both any-order and semi-autoregressive decoding. On Sudoku with 1B MDM, SAS improves puzzle accuracy from 82.0% (best heuristic schedule) to 91.8%, and reaches 97.5% with second-stage fine-tuning along learned trajectories. On mathematical reasoning with LLaDA-8B, SAS improves pass@1 on GSM8K from 64% to 76% and on MBPP from 39.5% to 41%, consistently matching or exceeding heuristic schedules across generation lengths and block sizes. Project page: https://jimmyxu123.github.io/SAS
Large reasoning models typically follow a read-then-think paradigm: they observe the complete input, reason over a static context, and then produce the answer. Yet many real-world scenarios are inherently dynamic, such as audio and video stream, where information arrives as a continuous stream and models must reason, update, and respond under partial observations. Recent streaming reasoning methods allow models to think while reading, but they largely rely on supervised imitation of pre-constructed trajectories, which limits their flexibility. In this paper, we propose AdaSR, an adaptive streaming reasoning framework that enables models to reason during input streaming and perform final deliberation once the stream is complete, learning when to think, and how much computation to allocate across different stages. To optimize this hierarchical reasoning process, we introduce Hierarchical Relative Policy Optimization (HRPO), which decomposes policy optimization into streaming reasoning and deep reasoning phases, providing more fine-grained advantage assignment instead of uniformly distributing a single sequence-level advantage over all tokens. HRPO integrates format, accuracy, and adaptive thinking rewards to enforce valid reasoning protocols, preserve final task performance, and encourage latency-aware computation allocation. Experiments show that AdaSR achieves a better balance among reasoning accuracy, computational efficiency, and streaming latency compared with supervised fine-tuning baseline. We release our code at https://github.com/EIT-NLP/StreamingLLM/tree/main/AdaSR.
Latent reasoning has emerged as a promising alternative to discrete Chain-of-Thought (CoT) in large language models (LLMs), enabling more expressive reasoning by operating over continuous representations. However, the inherently deterministic nature of continuous representations limits policy exploration in reinforcement learning (RL). To address this, we propose TARPO (Token-Wise Latent-Explicit Reasoning via Action-Routing Policy Optimization), a pure RL framework that adaptively switches between discrete token generation and continuous latent reasoning at each step. TARPO introduces a lightweight action head router that observes the current hidden state and samples a routing decision from a binary mode-selection space, preserving the stochasticity of discrete token sampling from the vocabulary. The LLM backbone and router are jointly optimized end-to-end with a shared group-relative advantage signal. Extensive experiments across Qwen2.5 (from 1.5B to 7B) and Llama-3.1-8B backbones demonstrate that TARPO consistently outperforms existing explicit and latent reasoning RL baselines across diverse benchmarks. Further analysis shows that TARPO learns adaptive token-wise switching behaviors while maintaining stable training dynamics. Our code is available at https://github.com/NKU-LITI/TARPO-master.
Tool-integrated Text-to-SQL parsing has emerged as a promising paradigm, framing SQL generation as a sequential decision-making process interleaved with tool execution. However, existing reinforcement learning approaches mainly rely on coarse-grained outcome supervision, resulting in a fundamental credit assignment problem: models receive the same reward for any trajectory that yields the correct answer, even when intermediate steps are redundant, inefficient, or erroneous. Consequently, models are encouraged to explore suboptimal reasoning spaces, limiting both efficiency and generalization. To address this problem, we propose FineStep, a novel framework for step-level credit assignment in tool-augmented Text-to-SQL. First, we introduce a reward design with independent process rewards to alleviate the signal sparsity of outcome supervision. Next, we present a step-level credit assignment mechanism to precisely quantify the value of each reasoning step. Finally, we develop a policy optimization method based on step-level advantages for efficient updates. Extensive experiments on BIRD benchmarks show that FineStep achieves state-of-the-art performance and reduces redundant tool interactions, with a 3.25% average EX gain over GRPO at the 4B scale.