Self-Study Reconsidered: The Hidden Fragility of Learning from Self-Generated QA

• URL: http://arxiv.org/abs/2606.32002v1
• Date: Tue, 30 Jun 2026 17:35:14 GMT
• Title: Self-Study Reconsidered: The Hidden Fragility of Learning from Self-Generated QA
• Abstract summary: Language models are increasingly taught from synthetic question--answer (QA) supervision.<n>We show that this generation step is not neutral preprocessing.<n>We show that this failure modes arise from choices made during QA generation, so they can be reduced without changing the training loop.
• Score: 40.94400211806987
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Language models are increasingly taught from synthetic question--answer (QA) supervision: a model generates questions about a document, answers them from the same text, and the resulting pairs are used to fine-tune, distill, or compress knowledge into another model. We show that this generation step is not neutral preprocessing. It is an implicit policy that both selects which evidence becomes training signal and decides how that evidence is answered, and it is fragile at both stages. When choosing what to ask, generators do not scan a document uniformly. Coverage saturates early and concentrates on salient spans, diverse prompts converge on the same regions, and what looks question-worthy is driven by local presentation. As a result, salient artifacts such as poorly cleaned markup can hijack question generation across model families and scales. When answering, the model that produces the supervision tends to obey instruction-like passages embedded in the text. This compliance depends on the intent and surface form of the passage rather than its strictness, and is worst under task conflict, where larger models comply more often. These failure modes arise from choices made during QA generation, so they can be reduced without changing the training loop. Tying each question to a fixed target reduces biased selection, and filtering instruction-like spans before answering lowers mean injection compliance from $88\%$ to $13\%$ in our evaluation while retaining nearly all clean text.

Related papers

• Answer-Conditioned Chains of Thought Degrade Verifiable-Reasoning Distillation in Large Language Models [21.049419375153914]
  We show that a common fix shows the generator the gold answer and asks it to write a chain that reaches that answer.<n>We show that this second step degrades the training data in a way that correctness filtering cannot catch.
  arXiv  Detail & Related papers  (2026-07-16T04:28:19Z)
• Implicit Reasoning Steering via Concept Chaining [29.121206944212556]
  We use natural-language text to bias a model toward a designated answer without explicit instructions, triggers, or direct answer cues.<n>We show that indirect, natural-looking text can systematically steer model predictions while remaining substantially less inferable than direct paraphrases.
  arXiv  Detail & Related papers  (2026-07-15T18:03:08Z)
• What Am I Missing? Question-Answering as Hidden State Probing [13.065059683491958]
  We propose to leverage question-asking as an inference-time intervention.<n>We train a probe on the student's hidden state before and after asking a question.<n>We find that the success of question-asking as an intervention is largely dependent on the model's self-consistency.
  arXiv  Detail & Related papers  (2026-05-29T17:27:07Z)
• Bridging the Know-Act Gap via Task-Level Autoregressive Reasoning [22.264933226031435]
  We present a comprehensive analysis using FaultyScience, a newly constructed large-scale, cross-disciplinary benchmark of faulty questions.<n>We show that the gap is pervasive and stems from token-level autoregression, which entangles task selection with content generation.<n>We propose DeIllusionLLM, a task-level autoregressive framework that explicitly models this decision.
  arXiv  Detail & Related papers  (2026-03-23T22:43:34Z)
• Answer, Assemble, Ace: Understanding How LMs Answer Multiple Choice Questions [103.20281438405111]
  Multiple-choice question answering (MCQA) is a key competence of performant transformer language models.<n>We employ vocabulary projection and activation patching methods to localize key hidden states that encode relevant information for predicting the correct answer.<n>We show that subsequent layers increase the probability of the predicted answer symbol in vocabulary space, and that this probability increase is associated with a sparse set of attention heads with unique roles.
  arXiv  Detail & Related papers  (2024-07-21T00:10:23Z)
• On Few-Shot Prompting for Controllable Question-Answer Generation in Narrative Comprehension [0.0]
  We propose a few-shot prompting strategy for controlling the generation of question-answer pairs from children's narrative texts. We show the effectiveness of controlling the generation process by employing few-shot prompting side by side with a reference model.
  arXiv  Detail & Related papers  (2024-04-03T15:17:21Z)
• Momentum Contrastive Pre-training for Question Answering [54.57078061878619]
  MCROSS introduces a momentum contrastive learning framework to align the answer probability between cloze-like and natural query-passage sample pairs. Our method achieves noticeable improvement compared with all baselines in both supervised and zero-shot scenarios.
  arXiv  Detail & Related papers  (2022-12-12T08:28:22Z)
• TASA: Deceiving Question Answering Models by Twin Answer Sentences Attack [93.50174324435321]
  We present Twin Answer Sentences Attack (TASA), an adversarial attack method for question answering (QA) models. TASA produces fluent and grammatical adversarial contexts while maintaining gold answers.
  arXiv  Detail & Related papers  (2022-10-27T07:16:30Z)
• Measuring and Narrowing the Compositionality Gap in Language Models [116.5228850227024]
  We measure how often models can correctly answer all sub-problems but not generate the overall solution. We present a new method, self-ask, that further improves on chain of thought.
  arXiv  Detail & Related papers  (2022-10-07T06:50:23Z)
• A Semantic-based Method for Unsupervised Commonsense Question Answering [40.18557352036813]
  Unsupervised commonsense question answering is appealing since it does not rely on any labeled task data. We present a novel SEmantic-based Question Answering method (SEQA) for unsupervised commonsense question answering.
  arXiv  Detail & Related papers  (2021-05-31T08:21:52Z)
• Counterfactual Variable Control for Robust and Interpretable Question Answering [57.25261576239862]
  Deep neural network based question answering (QA) models are neither robust nor explainable in many cases. In this paper, we inspect such spurious "capability" of QA models using causal inference. We propose a novel approach called Counterfactual Variable Control (CVC) that explicitly mitigates any shortcut correlation.
  arXiv  Detail & Related papers  (2020-10-12T10:09:05Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.