Jeonhwi Kang

Author directory

2026

Large reasoning models have recently demonstrated emergent “aha moments” that enable reflective reasoning. However, a non-trivial portion of reflective steps are “low-utility,” where reasoning errors are not corrected, or even correct reasoning trajectories are shifted toward incorrect ones. To address low-utility reflective continuation, this paper proposes RUAR, a utility-guided advantage rescaling method for reflective reasoning that estimates step-level utility from changes in correct-answer probability, without human process labels or a separately learned process reward model. RUAR augments outcome-supervised RL by rescaling the token-level advantages within each reflective reasoning step according to the estimated utility of the reflective continuation, rather than applying a uniform trajectory-level advantage across the reasoning trace. The utility is estimated through answer-forcing utility estimation, which measures how the model’s correct-answer probability changes before and after the reflective step by forcing answer generation from both reasoning states using answer-forcing prompts. Furthermore, RUAR introduces an overthinking-mitigating rescaling mechanism, where utility-guided rescaling is no longer applied from the earliest answer-ready reflection step onward. Experiments demonstrate that RUAR improves the accuracy–efficiency trade-off, substantially reducing generated token usage while often preserving or improving final-answer accuracy.
Reasoning-oriented language models can generate thousands of tokens before answering, yet the internal factors that determine their reasoning length remain poorly understood. This paper studies whether reasoning length can be controlled through sparse neuron interventions, complementing prompt-based, budget-based, and controller-based approaches. The proposed method discovers reasoning-length-associated neurons from paired long and concise correct trajectories on a training split, ranks MLP neurons by long-minus-concise activation contrast, and tests the selected neurons on disjoint evaluation splits by scaling their activations during generation. The evaluation covers math, science, and commonsense reasoning tasks, including GSM8K, MATH500, ARC-Challenge, CommonsenseQA, and QASC, under a strict split-separated protocol that never discovers neurons on held-out evaluation examples. In several, but not all, task/model/budget regimes, selected neurons modulate generated length differently from layer-matched random controls, with a modest single-run specificity result on CommonsenseQA accompanied by a small accuracy increase. These results provide task-, model-, and prompt-protocol-specific evidence that sparse MLP neurons can, in some split-controlled regimes, provide neuron-level control signals for generated reasoning length.
Standard semi-autoregressive decoding for full-attention masked diffusion language models (MDLMs) computes predictions for future blocks but discards them until those blocks become active. When introduced as soft-future context, these predictions reveal a stability contrast: hypotheses later revised during decoding exhibit larger distributional shifts than those eventually retained. We introduce FUSA (Future-aware Semi-Autoregressive Decoding), a training-free decoder that exploits this contrast by alternating the next block between masked and soft representations and committing active-block tokens that remain stable across the two views, without additional model forward calls. On LLaDA-8B-Instruct, FUSA improves scores across reasoning and code-generation benchmarks while delivering 2.10– 3.94× wall-clock speedups over Top-1 decoding, with gains of up to 5.1 points. Additional results show that FUSA transfers to other MDLMs and remains effective under KV-cached inference and in long-form instruction following. Our code is available at https://github.com/yangryeol/FUSA.
Recent advances in large language models(LLMs) have given rise to large reasoning models that achieve strong performance through explicit multi-step reasoning. However, models often generate long reasoning chains even for simple inputs. We present Difficulty-Aware Long-to-Short LLM Reasoning with Model Merging (DALSA), an inference-time framework that adaptively controls reasoning depth based on input difficulty. DALSA integrates (i) representation-based difficulty estimation, which predicts a soft failure-rate score from the prompt hidden state, and (ii) difficulty-aware model merging, which uses a quantile remapped score to modulate the interpolation coefficient between short- and long-reasoning models. DALSA trains a lightweight probe while keeping both LLMs frozen. On mathematical benchmarks with the evaluated Qwen pairs, the accuracy–length trade-off depends on scale: at 7B, DALSA preserves static ACM-TA average accuracy (66.60% vs. 66.61%) while reducing average response length by 22.35%; at 1.5B, it improves average accuracy from 56.75% to 58.24% with a 2.27% length in- crease. Additional merging baselines, routing-threshold and merge-range sweeps, and component ablations show sensitivity to the operating configuration. Domain-matched commonsense and Llama-family diagnostics do not establish broader efficiency gains, limiting the primary empirical claim to the evaluated Qwen mathematical setting.