A systematic study of 15 large reasoning models reveals that their intermediate reasoning traces pose higher safety risks than final answers, categorizes three failure modes, and proposes an adaptive multi‑principle steering mitigation that reduces unsafe outputs while preserving model capability.
The paper reveals that reinforcement‑learning post‑training flattens the entropy of the final layer in large reasoning models, making higher sampling temperatures ineffective, and introduces Latent Exploration Decoding (LED) to recover exploration from intermediate layers, yielding consistent pass@k gains without extra training.
SpecExit combines early‑exit and speculative decoding to let large reasoning models detect when they have almost finished thinking, trimming redundant chain‑of‑thought steps, reducing over‑thinking by 72% and achieving up to 2.5× faster end‑to‑end inference without noticeable accuracy loss.
Apple’s research reveals that large reasoning models, despite sophisticated self‑reflection mechanisms, experience a complete performance collapse when problem complexity exceeds a threshold, highlighting fundamental limits in their ability to achieve generalized reasoning.
