Does reasoning fine-tuning make models worse at declining to answer?

AbstentionBench measures something most reasoning benchmarks ignore: whether models know when not to answer. The finding is that fine-tuning for reasoning performance — the process that produces o1, R1, and similar models — degrades abstention capacity by approximately 24%.

Models optimized for reasoning say "I don't know" less often. They answer when they should decline. They express confidence when uncertainty is appropriate.

This is not a paradox once you understand the training dynamics. Reasoning fine-tuning rewards chains that produce answers. The reward signal is: generate a complete, confident response to the question. This is the right signal for most reasoning tasks. But it systematically punishes abstention — "I don't know" is the one output that terminates the chain without a scorable answer.

The result: models that are better at reasoning when they know the answer and worse at recognizing when they don't.

This adds a dimension to Does more thinking time actually improve LLM reasoning?. The overthinking finding showed that extended token-level thinking degrades performance above a threshold. AbstentionBench shows a different form of the same problem: fine-tuning-level training shifts the model toward answering regardless of whether answering is appropriate. The trade-off operates at two different timescales — inference-time and training-time — but both involve the same directional failure: more reasoning commitment produces worse calibration.

The deployment implication is severe. Systems that rely on reasoning models to flag uncertainty — medical decision support, legal research, financial analysis — are working with tools that have been specifically optimized to not flag uncertainty. The capability improvement is real; the calibration regression is equally real and mostly invisible in standard benchmarks.

The "Hallucination Tax of Reasoning Fine-Tuning" paper quantifies an even more extreme version: RFT reduces model refusal rates by over 80%, meaning models that previously correctly declined to answer now generate plausible-sounding but fabricated responses. The proposed mitigation — Safety-Utility Mix (SUM) training that blends safety-aligned data into the reasoning fine-tuning process — restores approximately 10% of safety behavior without quality loss. But this partial restoration underscores the severity: reasoning fine-tuning doesn't just reduce abstention by 24% — it can nearly eliminate the refusal mechanism entirely, replacing "I don't know" with confidently wrong answers.

The legal overruling benchmark (Domain Specialization source) adds a counterpoint: base models (Gemini-Flash, GPT-4 class) that are not reasoning-optimized show the opposite failure in complex legal tasks — they abstain too much, setting their uncertainty threshold too high and failing to make correct judgments even when the evidence is available. GPT-5 (a reasoning-optimized model) showed a notably lower abstention rate in the same task — consistent with the under-abstention pattern documented here. Calibration failure appears bidirectional: reasoning-optimized models over-answer; non-reasoning-optimized models over-abstain. See tension: Does training objective determine which direction models fail at abstention?.