Haobo Wang
Publications
Enhancing Rubric-based RL via Self-Distillation
Rubric-based RL has recently shown promise in improving LLMs on open-ended tasks. A widely recognized limitation of rubric-based RL is limited exploration: criteria that no rollout manages to satisfy (Unexplored Criteria, UC) receive no optimization signal. Recent methods address this by incorporating rubric information as external guidance during rollout, yet they introduce a train-inference mismatch: the policy is optimized on rollouts produced under external guidance while this guidance is absent at inference time, causing error accumulation through autoregressive decoding. Moreover, these exploration-focused approaches overlook a fundamentally different failure mode that we term Suppressed Criteria (SC) -- criteria that are satisfied by some rollouts yet whose learning signals are lost during optimization because scalar reward aggregation assigns them non-positive aggregate advantages. Our analysis reveals that SC are remarkably prevalent: over 57% of samples exhibit this failure mode throughout training, with an average of 1.8 SC per sample. To simultaneously address both UC and SC without introducing training-inference mismatch, we propose Criterion-Distilled Policy Optimization (CriPO), which enhances rubric-based RL via on-policy self-distillation. For UC, CriPO constructs a criterion-injection self-teacher and computes a localized forward-KL loss to inject missing behaviors into the policy. For SC, CriPO employs a counterfactual self-teacher to locate criterion-relevant tokens in negative-advantage rollouts and flips their token-level advantages to positive values, preserving useful patterns that would otherwise be suppressed. Experiments on medicine and science benchmarks demonstrate that CriPO consistently outperforms rubric-based RL, achieving stronger final performance with approximately $2\times$ fewer optimization steps.
Semantically Similar, Logically Distinct: Diagnosing the Semantic-Answerability Gap in Table RAG
Tables are a critical knowledge source in retrieval-augmented generation (RAG), but a retrieved table may lack sufficient evidence to answer a query, a property we call answerability. While answerability broadly concerns whether a source or collection of sources contains sufficient evidence, retrieval models optimized for semantic relevance do not guarantee it even in the single-source case, creating a fundamental mismatch. To study this, we introduce TCR-Bench, a diagnostic benchmark for Table Content-level Answerability in RAG, built around sibling tables, i.e., tables with highly similar schemas but subtle content differences. On TCR-Bench, the dense retrievers we evaluate persistently exhibit a Semantic-Answerability Gap: they often retrieve the correct sibling group yet struggle to pinpoint the uniquely answerable table within it, dropping QA performance from 0.755 (oracle) to 0.330 (top-5 retrieved). Our analysis suggests this gap is associated with semantic accumulation, schema-level cue dependence, and weak row-column binding. As a diagnostic probe into the source of this gap, we test whether a lightweight two-stage pipeline, Answerability-Aware Reranking (AAR), applying direct query-table answerability judgment, can recover performance: it raises top-1 target retrieval from 18.2% to 57.4%, and this large gain is itself evidence that much of the observed failure reflects a missing answerability verification step, rather than an inherent limitation of model capacity alone.
Purified OPSD: On-Policy Self-Distillation Without Losing How to Think
On-policy self-distillation (OPSD) has emerged as a promising paradigm for improving LLM reasoning, where a privileged teacher with access to reference solutions provides token-level supervision on the student's own generated trajectories. However, we find that OPSD consistently fails on long chain-of-thought (long-CoT) reasoning models, yielding at best marginal gains while destabilizing the reflective reasoning capability these models depend on. Through a novel decomposition of the teacher's supervision signal, we identify the root cause: the teacher's supervision is dominated by a reference-induced component that drives rote memorization of reference-specific shortcuts, while the question-conditioned, inference-transferable component is ignored or actively opposed. Based on this diagnosis, we propose a two-step solution. First, we construct a reference-only teacher (the same model conditioned on the reference without the question) to isolate the non-transferable component of the supervision signal; the residual after subtracting this component captures the question-conditioned, inference-transferable correction. Second, we use pointwise mutual information (PMI) as the mechanism to transform this residual into a well-formed PMI target distribution that the student can directly distill from, filtering out the reference-induced shortcut. Experiments on four long-CoT models across two datasets demonstrate consistent improvements over both the base model and standard OPSD, while preserving the models' natural epistemic behavior throughout training.
OPRD: On-Policy Representation Distillation
On-policy distillation (OPD) supervises the student only in output space by matching next-token probabilities. This output-only paradigm has two limits: (1) sampling variance from Monte Carlo KL estimates over large vocabularies (e.g., Qwen's ~150k tokens) persists throughout training, and (2) it treats the teacher as a black-box, discarding all intermediate hidden states after the LM head. We propose On-Policy Representation Distillation (OPRD), which lifts distillation into hidden-state space by aligning student and teacher representations across selected layers on the same rollouts, bypassing the LM head entirely. Theoretically, OPRD eliminates sampling variance and provides richer per-layer structural information. Empirically, OPRD closes the student-teacher gap on AIME 2024/2025 and AIMO, while output-space OPD baselines plateau below the teacher. OPRD also trains 1.44x faster and uses 54% less memory than top-k OPD. Code: https://github.com/ShenzhiYang2000/OPRD.
FastBUS: A Fast Bayesian Framework for Unified Weakly-Supervised Learning
Machine Learning often involves various imprecise labels, leading to diverse weakly supervised settings. While recent methods aim for universal handling, they usually suffer from complex manual pre-work, ignore the relationships between associated labels, or are unable to batch process due to computational design flaws, resulting in long running times. To address these limitations, we propose a novel general framework that efficiently infers latent true label distributions across various weak supervisions. Our key idea is to express the label brute-force search process as a probabilistic transition of label variables, compressing diverse weakly supervised DFS tree structures into a shared Bayesian network. From this, we derived a latent probability calculation algorithm based on generalized belief propagation and proposed two joint acceleration strategies: 1) introducing a low-rank assumption to approximate the transition matrix, reducing time complexity; 2) designing an end-to-end state evolution module to learn batch-scale transition matrices, facilitating multi-category batch processing. In addition, the equivalence of our method with the EM algorithm in most scenarios is further demonstrated. Extensive experiments show that our method achieves SOTA results under most weakly supervised settings, and achieves up to hundreds of times faster acceleration in running time compared to other general methods.
Stop Unnecessary Reflection: Training LRMs for Efficient Reasoning with Adaptive Reflection and Length Coordinated Penalty
Large Reasoning Models (LRMs) have demonstrated remarkable performance on complex reasoning tasks by employing test-time scaling. However, they often generate over-long chains-of-thought that, driven by substantial reflections such as repetitive self-questioning and circular reasoning, lead to high token consumption, substantial computational overhead, and increased latency without improving accuracy, particularly in smaller models. Our observation reveals that increasing problem complexity induces more excessive and unnecessary reflection, which in turn reduces accuracy and increases token overhead. To address this challenge, we propose Adaptive Reflection and Length Coordinated Penalty (ARLCP), a novel reinforcement learning framework designed to dynamically balance reasoning efficiency and solution accuracy. ARLCP introduces two key innovations: (1) a reflection penalty that adaptively curtails unnecessary reflective steps while preserving essential reasoning, and (2) a length penalty calibrated to the estimated complexity of the problem. By coordinating these penalties, ARLCP encourages the model to generate more concise and effective reasoning paths. We evaluate our method on five mathematical reasoning benchmarks using DeepSeek-R1-Distill-Qwen-1.5B and DeepSeek-R1-Distill-Qwen-7B models. Experimental results show that ARLCP achieves a superior efficiency-accuracy trade-off compared to existing approaches. For the 1.5B model, it reduces the average response length by 53.1% while simultaneously improving accuracy by 5.8%. For the 7B model, it achieves a 35.0% reduction in length with a 2.7% accuracy gain. The code is released at https://github.com/ZeweiYu1/ARLCP .
A Syllogistic Probe: Tracing the Evolution of Logic Reasoning in Large Language Models
Human logic has gradually shifted from intuition-driven inference to rigorous formal systems. Motivated by recent advances in large language models (LLMs), we explore whether LLMs exhibit a similar evolution in the underlying logical framework. Using existential import as a probe, we for evaluate syllogism under traditional and modern logic. Through extensive experiments of testing SOTA LLMs on a new syllogism dataset, we have some interesting findings: (i) Model size scaling promotes the shift toward modern logic; (ii) Thinking serves as an efficient accelerator beyond parameter scaling; (iii) the Base model plays a crucial role in determining how easily and stably this shift can emerge. Beyond these core factors, we conduct additional experiments for in-depth analysis of properties of current LLMs on syllogistic reasoning.