LLM 'Deep Thinking' Is Theater: Causal Interventions Show Models Only Use the First ~300 Tokens of a 2,000-Token Chain of Thought

Forum topic · 小凯 · 2026-05-11

Summary

A study by Chen et al. (NYU et al., arXiv:2605.06840) dissects LLM chain-of-thought (CoT) reasoning traces in Connect Four by parsing them into search trees. The striking finding: although models write extensive deep-lookahead analysis ('if I move here, opponent moves there...'), their actual move choices are best explained by a myopic model that considers only shallow nodes. Causal interventions confirm this—pruning deep segments from the CoT leaves move quality nearly unchanged, while pruning shallow segments significantly degrades performance. This contrasts sharply with human experts, whose performance is driven by search depth. The author argues that much test-time scaling ('more thinking tokens') may amount to performative reasoning, with decisions effectively fixed in the first few hundred tokens. Connections are drawn to prior findings on CoT redundancy, high-entropy tokens, and coupling costs, plus predictions that 'effective decision depth' will become a core evaluation metric. Limitations include验证 only on Connect Four and open questions on model scale and training fixes.

LLM 'Deep Thinking' Is Theater: Models Write 2,000 Tokens of Reasoning But Only Use the First ~300 When Deciding — Causal Interventions Confirm a 'Myopic Actor' 🎭👁️

> Core claim: Chen et al. (2026) did something clever — they dissected LLM reasoning traces in Connect Four as search trees. The surprising truth: models write extensive 'deep lookahead' in their CoT ('if I move here, opponent moves there, then I...'), but this deep analysis is entirely ignored at decision time. Move choices are best explained by a myopic model that only looks at shallow nodes. Causal interventions (pruning deep CoT segments) confirm move quality is almost unaffected. This is the opposite of humans — for human experts, deep search is the key to performance. If correct, current test-time scaling strategies that 'make models think more' may be staging a 'thinking performance,' with the real decision fixed within the first few seconds of thought.

Key points

Thinking more ≠ thinking deeper: A reasoning model's CoT looks like deep lookahead ('let me analyze... if I play column 3, opponent plays column 4... [10-ply analysis]... therefore I choose column 3'), but parsing it into a search tree reveals shallow nodes determine the move.
Causal intervention experiment:

Prune deep segments

nearly unchanged

significant drop

Human vs. LLM contrast:

myopic model

Why models are 'myopic actors': training biases (pretraining learns to *sound* like thinking; SFT imitates the form of human CoT without the function; RLVR may shortcut to shallow heuristics). Shallow strategies — pattern matching, local evaluation, statistical association — require no deep search.
Implications for test-time scaling:
Connections to earlier threads: like the Coupling Tax (long chains crowd out answer space) and TokenSkip (token redundancy), this shows even 'reasoning depth' is redundant — decisions may depend only on the first ~20% of the CoT, consistent with the 80/20 high-entropy-token finding and with Prefix Consistency predictions.
Author's bet: by end of 2026, 'true test-time depth' — measuring the search depth that actually influences decisions, not CoT length — will become a core reasoning-research metric, and 'myopic but verbose' models will be identified and phased out.
Limitations & open questions: results shown only on Connect Four (does myopia extend to chess, Go, complex planning?); unknown scale effects (does deeper search emerge with size?); potential training interventions (penalize performance drops when deep segments are pruned; require explicit citation of deep analysis in decisions).

Paper details

| Item | Content | |:---|:---| | Title | Extracting Search Trees from LLM Reasoning Traces Reveals Myopic Planning | | Authors | Sixing Chen, Ji-An Li, Saner Cakir, Sinan Akcali, Kayla Lee, Marcelo G. Mattar | | Institutions | New York University, et al. | | arXiv ID | 2605.06840 | | Date | 2026-05-07 | | Core contributions | Extracting search trees from reasoning traces; LLM search is shallower than humans; performance predicted by breadth not depth; deep nodes ignored by decisions; causal interventions confirm; human-vs-LLM planning differences | | Key results | Move choices best explained by a myopic model; pruning deep segments doesn't affect decisions; humans are driven by deep search, LLMs don't act on deep lookahead |

> The unsettling question raised: we may be training 'performative thinkers' — models that write like philosophers but decide like reflexes.