The paper introduces Weakening-Induced On-Policy Self-Distillation (WISD), a method that improves large reasoning models by creating a weaker version of the original model and using the contrast between them as a training signal, without needing privileged information. WISD constructs the weaker model via low-bit quantization or magnitude pruning, and evaluates it on mathematical and coding reasoning tasks across various architectures and scales. The results show significant accuracy gains—up to 5.5 percentage points—over the original models, achieving state‑of‑the‑art performance on benchmarks such as HMMT 2025 and LiveCodeBench v6.
arXiv:2609.26708v1 Announce Type: new Abstract: Quantization-aware distillation (QAD) restores much of the short-form question-answering performance lost to sub-3-bit quantization, yet leaves mathema...
arXiv:2508. 09883v2 Announce Type: replace-cross Abstract: Large language models (LLMs) demonstrate remarkable reasoning capabilities in tasks such as algorithmic coding and mathematical problem-solving.
The paper introduces masked self‑distillation, a post‑training framework that trains a language model to internalize portions of its own intermediate reasoning traces. By varying the fraction of trace internalized, the authors demonstrate that models can achieve higher inference efficiency and improved task performance on math and graph‑coloring problems. Experiments on Qwen3‑4B and Qwen3‑8B show that the method generalizes well to in‑domain out‑of‑distribution cases without catastrophic forgetting, and that supervised fine‑tuning alone can reduce trace length at the expense of generalization.
arXiv:2605.10194v2 Announce Type: replace Abstract: On-policy self-distillation (OPSD) uses a model as its own teacher under privileged context, providing token-level supervision on the model's own r...
arXiv:2606. 31048v1 Announce Type: cross Abstract: This paper investigates knowledge distillation from a large reasoning model (DeepSeek-R1) to a compact student model (Qwen2.
ThinkBit is a two‑bit scalar quantization framework designed to preserve reasoning capabilities in language models. It assigns each layer a ternary, 2‑bit, or 3‑bit tier based on sensitivity, then uses a Gumbel‑softmax search to find optimal codewords block‑by‑block. After freezing codewords, it recalibrates shared scales via knowledge‑distillation and performs a brief quantization‑aware training, finally selecting the best checkpoint with paired reasoning tests.
The paper investigates why self‑distillation can sometimes worsen the reasoning abilities of large language models (LLMs). It finds that the process suppresses the model’s epistemic verbalization—its expression of uncertainty—leading to shorter but less accurate responses in mathematical reasoning tasks. Experiments on several LLMs show performance drops of up to 40%, especially on out‑of‑distribution problems where uncertainty expression is beneficial.
BoundaryEvo introduces a two‑stage data‑model co‑evolution framework for distilling long‑chain reasoning from large teacher models into small language models. First, Capability Priming distills a small set of teacher trajectories to give the student an initial reasoning baseline. Then, Capability‑Boundary Co‑Evolution iteratively selects questions the student has not mastered but the teacher can solve, updating the training data as the student’s capability grows. With only 1,000 examples, BoundaryEvo improves Qwen3.5‑4B accuracy from 62.9 % to 75.3 % on IMO‑AnswerBench and from 26.8 % to 46.0 % on AMO‑Bench, outperforming static and dynamic baselines across 11 benchmarks.
OBC‑Prune introduces an outcome‑based calibration approach for pruning large reasoning models, focusing on the causal importance of each reasoning sentence rather than uniform activation salience. By pairing correct and incorrect rollouts and using intervention‑based analysis, it assigns per‑token weights that guide one‑shot pruning methods such as SparseGPT, Wanda, and ALPS. Experiments on DeepSeek‑R1‑Distill‑Qwen models show consistent accuracy gains and shorter reasoning traces across multiple benchmarks at 40‑50% sparsity.
The paper introduces OnQ, a low‑bit recovery pipeline for quantized large language models that uses on‑policy distillation and candidate‑wise Dynamic KL to address trajectory divergence and distributional misalignment. OnQ starts with block‑wise calibration, then lets the quantized student generate responses and receive full‑precision teacher supervision on its own prefixes, dynamically weighting forward and reverse KL losses per candidate. Experiments on Qwen3 models show that OnQ outperforms baselines on nine benchmarks, especially long‑horizon reasoning, and produces shorter, less truncated responses.
arXiv:2605. 15532v3 Announce Type: replace-cross Abstract: Distillation enables compact Vision-Language Models (VLMs) to obtain strong reasoning capabilities, yet the prompts driving this process are typically chosen via simple heuristics or aggregated from off-the-shelf datasets.
arXiv:2610.18613v1 Announce Type: cross Abstract: Self-distillation is an emerging approach that trains a model to imitate its own predictions when conditioned on privileged information, such as corr...