New Papers Probe the Hidden Costs and Risks of LLM Reasoning Traces

Chain-of-thought reasoning is no longer being judged only by the final answer. Six new arXiv preprints treat the intermediate reasoning trace itself as an object of study, and they arrive at different conclusions about what it is for. GUARD argues that protected facts and unsafe rationales can surface in chain-of-thought before the final answer, making unlearning harder, and proposes a distillation-based method for 'natural forgetting' in large reasoning models. LogicTrack similarly focuses on intermediate steps, but from a quality angle: it uses formal logic solvers to audit whether reasoning trajectories are logically valid, rather than relying only on outcome-based feedback.

Other papers emphasize cost and efficiency. OverThink describes a slowdown attack that induces reasoning models to emit large numbers of hidden reasoning tokens, threatening latency and cost in chatbots and coding agents. DRT tries to make multimodal reasoning more efficient by moving beyond natural-language chain-of-thought to a dense reasoning trace. A separate hierarchical analysis of machine translation finds that reasoning traces do not always help: their value depends on models, languages, domains, and datasets.

One paper adds a behavioral angle: rewarding efficient reasoning improves abstention on underspecified tasks, addressing the observation that reasoning models often fail to know when to decline to answer. The papers agree that hidden reasoning traces matter beyond final accuracy, but they diverge on what to do with them: forget parts of them, attack them, audit them, compress them, or use them to decide when to abstain.