SYNTHESIS NOTE
Topics›Reasoning Critiques›this note

Why do reasoning models overthink ill-posed questions?

Explores why models trained for extended reasoning produce drastically longer, less useful responses to unanswerable questions—and whether this represents a fixable training deficit or inherent limitation.

Synthesis note · 2026-02-22 · sourced from Reasoning Critiques

The standard case for reasoning models: they think more, therefore they reason better. The missing-premise case inverts this completely.

When given questions with missing premises (MiP) — questions that are unanswerable because they lack necessary information — reasoning models produce responses that are drastically longer than for normal questions. The additional length is not useful thinking. It is redundant self-doubt: the model cycles through "alternatively," "wait," "check," and "but" without making progress, unable to resolve the contradiction introduced by the missing premise.

Non-reasoning models behave differently. They produce shorter responses and are significantly more likely to identify the question as ill-posed. They achieve better abstain rates. They do not ruminate.

The mechanism: reasoning-specific training optimizes for generating thinking patterns — for using reasoning steps — but does not develop the meta-capability to recognize when thinking cannot help. The training signal rewards chains that lead to answers. Questions without valid answers do not provide this signal, so no training pressure develops the critical thinking capability to disengage.

Three observations deepen this:

  1. Reasoning models show large increases in step count for MiP questions — most steps are redundant self-doubt
  2. The overthinking is contagious through distillation — models distilled from reasoning model responses inherit the overthinking pattern
  3. The problem generalizes beyond the "missing premises" framing — any question where the correct response is not to reason further will expose this deficit

This contradicts the naïve test-time scaling law assumption. Scaling thinking tokens is supposed to improve outcomes. For ill-posed questions, it does the opposite. The model is burning compute on questions that require no answer, only recognition.

The practical implication for deployed reasoning agents: well-formed questions from trusted sources are fine. Ill-formed, ambiguous, or manipulative questions are not — the reasoning model will not disengage, it will overthink.

Prompting-level mitigation: ISP2 (Iterative Summarization Pre-Prompting) demonstrates that pre-reasoning information gathering can partially address the implicit/missing information problem. The technique extracts entities and their descriptions from the question, rates the reliability of these information pairs, then iteratively merges the lowest-reliability pairs into new descriptions — building a key information pair that is fed alongside the original question into reasoning. The principle: "understanding before reasoning" — CoT emphasizes reasoning stages but neglects the critical prior step of gathering and extracting essential information. ISP2 addresses the missing-premise gap from the prompting side, while training-based approaches like Can models learn to ask clarifying questions instead of guessing? address it from the capability side.

QuestBench extends the picture from behavior to diagnostics: models can't even IDENTIFY what information is missing. At 40-50% accuracy on logic and planning clarification tasks, the information acquisition failure precedes the overthinking failure. See Can models identify what information they actually need? — the two findings describe a two-part deficit: (1) cannot detect what information is needed, (2) cannot disengage when information is absent.

"When Prompts Go Wrong" (2025) extends this to code generation with a systematic taxonomy. Ambiguous descriptions (multiple plausible interpretations), contradictory descriptions (conflicting requirements), and incomplete descriptions (omitted constraints) each cause distinct failure modes. Contradictory descriptions result in the most logical errors — models attempt to satisfy incompatible requirements simultaneously. Incomplete descriptions cause models to make incorrect assumptions (e.g., assuming a base area is provided when "triangular" is omitted). Even larger, more resilient models are not immune. The finding generalizes the missing-premises problem: it is not specific to reasoning tasks but a fundamental vulnerability wherever task specifications are imperfect. Source: Arxiv/Prompts Prompting.

Inquiring lines that read this note 105

This note is a source for these research framings, grouped by the broader line of inquiry each explores. Scan the bold lines of inquiry; follow any specific question forward.

Do AI explanations represent actual model reasoning or constructed post-hoc narratives? What makes reasoning models struggle with ill-posed and unfamiliar problems? Does reasoning fine-tuning trade off other critical model abilities? How can models learn to ask clarifying questions effectively? What actually drives chain-of-thought performance gains in reasoning tasks? Does prompting inject new knowledge into models or activate existing knowledge? Can safety evaluations accurately measure true capabilities when models obscure them? Does parallel reasoning outperform sequential approaches in reasoning efficiency? What reasoning limitations and biases do language models systematically exhibit? Why do users believe false claims despite fact-checking efforts? When should retrieval be triggered to augment language model reasoning? How do prompt injection attacks compromise agent reasoning and security? What determines whether visible reasoning traces improve task performance? When do additional thinking tokens stop improving reasoning performance? Do reasoning traces accurately represent what models actually compute? How does transparency of agent reasoning affect effective oversight? Does preference-optimized training sacrifice accuracy and calibration for perceived quality? How do pretraining data composition and task diversity shape reasoning? Why does increased reasoning effort fail to improve theory of mind performance? Can prompts override training-embedded biases in language models? Do language models truly understand their encoded knowledge or merely retrieve patterns? Does chain-of-thought reasoning faithfully reflect actual model reasoning processes? How do different reward signals affect model reasoning and calibration differently? Why does self-revision increase model confidence despite degraded accuracy? Does reinforcement learning create new reasoning capabilities or refine existing ones? What training data structures and signals produce robust learning across distributions? How should AI ethics integrate consciousness, values, and transparent disclosure? How can agents learn and retain useful knowledge across diverse environments? How does scaling test-time compute compare to increasing model size? How do LLMs generate false citations that fool readers and evaluators? How do clinicians evaluate the trustworthiness of AI medical recommendations? How do mechanistic features and representations drive model behavior? How does verifier integration affect the generation-verification gap? How has ChatGPT adoption actually reshaped labor markets and information ecosystems? What factors determine AI persuasion effectiveness and what controls mitigate it?

Related concepts in this collection 11

This note in its neighbourhood — explore the map, then jump to a related concept in the list below.

Concept map
21 direct connections · 210 in 2-hop network ·dense cluster Open in graph ↗
Why do reasoning models overthink ill-posed ques… Does more thinking time always improve reasoning a… Does more thinking time actually improve LLM reaso… Why do reasoning models fail under manipulative pr… Does reasoning fine-tuning make models worse at de… Can models identify what information they actually… How do users actually form intent when prompting A… Why do language models lose performance in longer … When should retrieval happen during model generati…

Click a node to walk · click center to open · click Open in graph to see this note in the full knowledge graph

your link semantically near linked from elsewhere

Related papers in this collection 8

Papers most semantically related to this note, ranked by cosine similarity in the embedding space.

Search by related questions 4

Suggested questions this note speaks to — click to search the collection, or type your own.

Can models learn to identify what information is missing from questions?Does distillation from reasoning models spread overthinking to smaller models?How can prompting help models gather information before attempting reasoning?What training signals would teach models when not to reason?
Original note title

missing premises exacerbate overthinking — reasoning models lack critical thinking to reject ill-posed questions