Does training objective determine which direction models fail at abstention?

LLM abstention calibration fails in both directions depending on the training objective, not the model's general capability:

Reasoning-trained models under-abstain. RL/RLHF training for reasoning optimizes answer generation. Abstention is penalized because "I don't know" receives no reward. Since Does reasoning fine-tuning make models worse at declining to answer?, the result is overconfident models that answer when they shouldn't.

Safety-trained models over-abstain. RLHF with safety emphasis raises uncertainty thresholds too high. Models refuse benign prompts or decline complex but answerable open-ended tasks. TrustLLM demonstrates safety-training-driven over-refusal on completely safe questions.

Base models split by domain complexity. In simple templated tasks, base models calibrate reasonably. In complex open-ended domains (legal reasoning, medical diagnosis), base models set their uncertainty threshold too conservatively, under-answering questions they could handle.

The implication: "calibration" is not a single axis that can be fixed by one technique. The training objective creates a characteristic failure signature. A model that was tuned for both reasoning and safety faces contradictory calibration pressures — one pushes toward answering, the other toward refusing. This may explain why reasoning fine-tuning degrades abstention: it actively counteracts the safety training's conservative bias. A potential resolution exists: Does binary reward training hurt model calibration?, suggesting the axis conflict can be addressed at the reward design level.

For post-writing: connects to "the critical thinking problem" (reasoning training optimizes narrow thinking while degrading meta-cognitive judgment about when not to think) and the broader theme that training optimizes a target metric while degrading adjacent capabilities.