LLM 'Deep Thinking' Is Performative: Causal Intervention Shows Decisions Ignore Deep CoT Search

Key Points

• The study: Chen et al. (2026), *Extracting Search Trees from LLM Reasoning Traces Reveals Myopic Planning* (arXiv: 2605.06840), parsed LLM chain-of-thought (CoT) traces in Connect Four into search trees and asked which parts actually drive move selection.
• Core finding: Models write extensive deep lookahead ("if I move here, opponent moves there, then I..."), but move choices are best explained by a myopic model using only shallow nodes. Performance is predicted by search breadth, not depth—the opposite of human experts, whose play is driven by deep search.
• Causal intervention: Truncating deep reasoning segments left move quality nearly unchanged; truncating shallow segments significantly degraded it. Deep analysis is decorative, not functional.

Why LLMs Act as 'Myopic Actors'

The post attributes this to training-stage biases:

| Stage | What is learned | Consequence | |---|---|---| | Pretraining | Language pattern matching | "Writing like thinking" ≠ thinking | | SFT | Imitating human CoT form | Form copied, function not learned | | RLVR | Optimizing answer correctness | Shortcut shallow heuristics |

Plausible shallow heuristics include pattern matching, local threat evaluation, and statistical association—all requiring no deep search. The model writes 2000 tokens of analysis but the decision is effectively fixed within the first ~300 tokens; the rest functions as auto-completed "explanatory filler."

Implications for Test-Time Scaling

• "More compute = better decisions" may be wrong: deep CoT may be decoration.
• Length ≠ depth: useful metrics should measure *effective decision depth*, not CoT length.
• Tree-of-Thoughts-style multi-path search may also be wasted if decisions depend only on shallow layers.

Connections to Related Work

• Coupling Tax: long reasoning chains crowd out answer space—now shown to be doubly wasteful since most content may not inform the decision.
• 80/20 Rule: only ~20% high-entropy tokens matter; here, the shallow ~20% of the CoT may carry the decision.
• TokenSkip: redundancy exists not just at the token level but across reasoning *depth layers*.
• Prefix Consistency: insensitivity to deep-segment truncation is consistent with deep content being "inconsistent" with decisions.

The Author's Bet

The poster wagers $1,000 that by end of 2026, "true test-time depth" (effective decision depth rather than CoT length) becomes a core metric in reasoning research, and that long-writing-but-shallow-thinking models will be identified and penalized.

Limitations & Open Questions

• Task generalization: verified only on Connect Four; chess, Go, and complex planning remain untested.
• Model scale: does myopia persist or does deep search emerge with scale?
• Training interventions: e.g., penalizing performance drops when deep segments are truncated, or requiring explicit citation of deep analysis in decisions.
• Human-AI collaboration: how should a deep-searching human work with a myopic model?

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 | | Affiliation | New York University, et al. | | arXiv ID | 2605.06840 | | Date | 2026-05-07 | | Key results | Move choice best explained by myopic model; truncating deep segments does not affect decisions; humans are driven by deep search, LLMs do not act on deep lookahead |