Kambhampati Calls LLM 'Thinking Traces' a Mythos at Tübingen

The position paper's title is a command: "Stop Anthropomorphizing Intermediate Tokens as Reasoning/Thinking Traces!" Subbarao Kambhampati and nine co-authors published it at ICML 2026, and he carried the argument into a talk at the Max Planck / Ellis Institute in Tübingen on October 9, 2026, posted to YouTube soon after.

The paper defines what reasoning models actually emit as "the unfiltered tokens emitted by the LLM before the solution." Training uses triples of problem, intermediate tokens, and solution, and the external verifiers that score the model only check the final answer. The trace itself is never graded.

The experiments do the damage. Models trained on deliberately incorrect or swapped traces still produce correct answers at high accuracy. Reinforcement learning improves the answer while degrading trace coherence. Trace length does not track problem complexity, a finding Kambhampati's group formalized separately in a paper titled "Performative Thinking? The Brittle Correlation Between CoT Length and Problem Complexity." On the DeepSeek "aha moment" that circulated as evidence of emergent insight, the paper is deadpan: "The correlation here may be spurious rather than causal."

The reframe in the Tübingen talk is that traces are "better understood as prompt augmentations ('incantations') that the model learns to use to improve the solution token probability." Two of the researchers we track in our Who's Who directory posted the link in the days after it went up.