arXiv:2607. 28639v1 Announce Type: cross Abstract: We show that knowledge distillation in small instruction-tuned language models has asymmetric effects on bias.
The paper examines on‑policy self‑distillation, where a teacher scores a student’s rollout to provide dense token‑level credit. Experiments show that the raw scoring signals neither localize errors nor effectively penalize incorrect tokens, because locating errors is a search problem that the teacher is not explicitly asked to perform. Introducing a critique‑conditioned approach—first generating a critique of the rollout and then conditioning the teacher on it—produces signals that accurately localize errors, correct erroneous spans, and guide the policy toward backtracking tokens, improving performance on math problems and in cold‑start GRPO settings.
arXiv:2610.16318v2 Announce Type: replace Abstract: Large language models (LLMs) can reduce hallucination by abstaining from questions they would likely get wrong. Abstention can be implemented outsi...
The paper introduces a penalty‑aware evaluation framework for Retrieval‑Augmented Generation (RAG) systems that uses asymmetric scoring, knowledge‑gap canaries, and a failure‑attribution pipeline. Applying this framework to three commercial RAG products and a baseline on SimpleQA‑Verified, the authors find that while overall accuracy is similar across systems, canary violation rates vary dramatically, showing that systems differ more in when they answer than in what they answer. The study demonstrates that penalty‑aware scoring can reorder system rankings and is robust across different penalty settings.
DeclineCraft is a dynamic abstention benchmark for retrieval‑augmented generation (RAG) that tests whether models can recognize when neither retrieved documents nor parametric knowledge can answer a question. The benchmark uses unanswerable Minecraft feature‑addition queries that appear after a corpus cut‑off, allowing evaluation of future models. Across 27 large language models, the study measures Correct Abstention Rate (CAR), Answerable Attempt Rate, and Answer Accuracy, finding that single‑shot retrieval boosts CAR, iterative RAG reduces CAR with minimal impact on accuracy, and larger models improve accuracy but not CAR.
arXiv:2609.39702v1 Announce Type: new Abstract: Adapting a language model to a specialized corpus means choosing which instruction-tuning tasks to train on under a fixed budget, and testing one choic...
arXiv:2609.15467v1 Announce Type: cross Abstract: Adding answer options can lower multiple-choice scores without improving assessment validity. Turkish MMLU Pro examines this distinction using 12,000...
arXiv:2609.24885v1 Announce Type: new Abstract: When a language model answers from a curated corpus via graph-based retrieval, a large grounding uplift does not establish reasoning over the retrieved...
arXiv:2607. 14552v1 Announce Type: cross Abstract: A standard recipe for distilling the reasoning ability of large language models (LLMs) is to sample chains of thought from the model, keep those that reach the correct final answer, and fine-tune on the survivors.
arXiv:2607. 08268v1 Announce Type: new Abstract: High-volume structured extraction pays a large model's latency on every item, so distilling the task into a small on-device model is attractive: comparable output at a fraction of the time and cost.
The paper investigates how conversational Retrieval-Augmented Generation (RAG) benchmarks that rely on relevance judgments to determine context sufficiency can be misleading. By re‑running official baselines on the RETECO track and evaluating with both large‑language‑model judges and a blind audit, the authors show that stronger retrievers produce a higher proportion of contexts lacking gold documents yet still containing the answer, leading to a sharp drop in shortcut precision from 0.68 to 0.10. The study demonstrates that many false misses arise from hard negatives, gold content from other turns, or neighboring chunks, and that using retrieval scores alone as a sufficiency detector can underestimate answerability by up to 0.19 in AUROC.
The paper introduces CROP, a method for selective on‑policy distillation that prioritizes token‑level supervision based on task relevance. CROP uses paraphrase‑calibrated counterfactual sensitivity to measure how much each response token depends on the semantic content of the input, constructing validated original‑paraphrase‑counterfactual triplets for each prompt. Experiments in two teacher‑student settings show that CROP outperforms other selectors, improving aggregate performance by 1.92 and 2.96 points.