Learning to Refine Hidden States for Reliable LLM Reasoning

Large language models show strong reasoning ability, but their internal reasoning process can remain unstable in complex multi-step settings, where early hidden-state errors may propagate to incorrect predictions. We propose ReLAR, a reinforcement-guided latent refinement framework that iteratively updates hidden representations before decoding. ReLAR maintains a compact latent reasoning state and uses learned depth and action controllers to adaptively determine both the number and direction of refinement steps. The controllers are trained with a policy gradient objective based on step-wise likelihood improvement, enabling efficient input-dependent reasoning without explicit chain-of-thought generation. Experiments on medical, mathematical, multi-hop reasoning, and open-ended generation benchmarks show that ReLAR improves accuracy, generation quality, and reasoning stability with substantially lower inference overhead than explicit reasoning baselines.

Token Constraint Decoding Improves Robustness on Question Answering for Large Language Models

Large Language Models (LLMs) have demonstrated impressive performance on multiple-choice question answering (MCQA) benchmarks, yet they remain highly vulnerable to minor input perturbations. In this paper, we introduce and evaluate Token Constraint Decoding (TCD). This simple yet effective inference-time algorithm enforces alignment between token-level predictions to enhance robustness in noisy settings. Through extensive experiments on CommonsenseQA, MMLU, and MMLU-Pro, we show that TCD, especially when paired with prompt engineering (PE) fixes, significantly restores performance degraded by input noise, yielding up to +39\% absolute gains for weaker models like Gemma3 1B. Penalty sweep analyses further reveal that TCD implicitly regularizes overconfident outputs, with different models requiring distinct penalty schedules to maximize resilience. Our findings establish TCD as a practical, model-agnostic approach for improving reasoning stability under real-world imperfections and pave the way for more reliable deployment of LLMs in safety-critical or user-facing applications.