SYNTHESIS NOTE
Topics›Reasoning Logic Internal Rules›this note

What three separate factors drive chain-of-thought performance?

Can we isolate and measure the distinct contributions of output probability, memorization, and genuine reasoning to CoT success? Understanding their relative weights matters for knowing when CoT actually reasons versus when it relies on shortcuts.

Synthesis note · 2026-02-22 · sourced from Reasoning Logic Internal Rules

The "Deciphering Factors Influencing CoT" paper achieves something rare: a clean decomposition of what drives Chain-of-Thought performance into three independently measurable factors, using the simple but controlled task of shift cipher decoding across GPT-4, Claude 3, and Llama 3.1.

Factor 1: Output probability. The probability of the correct output in the model's distribution dramatically affects CoT accuracy. Varying only the output's probability of occurrence shifts GPT-4 accuracy from 26% to 70%. CoT works better when the answer is already more probable — it amplifies existing tendencies rather than overcoming them.

Factor 2: Memorization. Performance is higher when the specific cipher variant was more frequently encountered during pre-training. This is not reasoning — it is pattern matching against memorized instances. The frequency of encountering different shift values in training data directly predicts accuracy on those shifts.

Factor 3: Noisy reasoning. After controlling for probability and memorization, genuine reasoning effects remain — but they are noisy. Error rate increases with the number of implicit reasoning steps (shift magnitude). This is real multi-step reasoning, but each step introduces error probability, so accuracy degrades with chain length.

The decomposition resolves the ongoing debate about whether LLMs reason or memorize: they do both, simultaneously, and the contribution of each factor varies by task. This supports Does chain-of-thought reasoning reveal genuine inference or pattern matching? while adding a crucial nuance: the imitation IS partially genuine, but contaminated by probability bias and memorization artifacts.

The probability factor is particularly important for understanding CoT faithfulness. Since Do language models actually use their reasoning steps?, the probability dependence reveals a specific mechanism for causal insufficiency: CoT "reasoning" succeeds partly because the sequence of generated tokens increases the conditional probability of the correct answer, not because the logical content is being processed. This is exactly the mechanism behind Does logical validity actually drive chain-of-thought gains? — invalid exemplars work because they still generate token sequences that shift output probability toward correct answers.

The noisy-reasoning factor connects to Does more thinking time always improve reasoning accuracy?: if each reasoning step adds noise, then past some threshold the accumulated noise exceeds the signal, producing the inverted-U performance curve.

Inquiring lines that read this note 76

This note is a source for these research framings, grouped by the broader line of inquiry each explores. Scan the bold lines of inquiry; follow any specific question forward.

Does chain-of-thought reasoning faithfully reflect actual model reasoning processes? What actually drives chain-of-thought performance gains in reasoning tasks? Do AI explanations represent actual model reasoning or constructed post-hoc narratives? Why doesn't LLM knowledge translate to improved reasoning outputs? How does scaling test-time compute compare to increasing model size? When do additional thinking tokens stop improving reasoning performance? Why does separating planning from execution enhance reasoning performance? Do reasoning traces accurately represent what models actually compute? How can model evaluations remain robust against gaming and evaluator drift? Can prompts override training-embedded biases in language models? What determines whether visible reasoning traces improve task performance? What are the trade-offs between memory capacity and continuous task learning? How can we determine if computational systems are genuinely conscious? What can prompting techniques actually accomplish in language models? How should agent memory systems balance granularity, management, and reuse? How does verifier integration affect the generation-verification gap? Should reward models emphasize process understanding or only final outcomes? Does parallel reasoning outperform sequential approaches in reasoning efficiency? What explains the gap between LLM semantic understanding and production? Does polished AI output mislead users about actual accuracy and expertise? How can language models effectively manage long-context memory without quality degradation? What distinguishes genuine understanding from correct linguistic performance? Do AI safety evaluations reliably detect unsafe model behavior? Can post-hoc execution traces provide sufficient accountability without real-time oversight? How do benchmark scores relate to real-world task completion? How do pretraining data composition and task diversity shape reasoning? Can oversight systems work against strategically aware agents? Does reinforcement learning create new reasoning capabilities or refine existing ones? Can AI agents learn to improve performance from experience? How do transformer circuits and attention mechanisms enable or limit reasoning? How can outcome-only reporting obscure the true mechanisms of safety? Can models strategically underperform to evade capability evaluation? Does AI assistance enhance or undermine professional skill development? How do standardized artifacts improve coordination in multi-agent systems? How does transparency of agent reasoning affect effective oversight? Why does self-revision increase model confidence despite degraded accuracy? Does expanding agent autonomy undermine human oversight capacity? Do benchmark score improvements actually reflect genuine AI capability gains?

Related concepts in this collection 6

This note in its neighbourhood — explore the map, then jump to a related concept in the list below.

Concept map
13 direct connections · 136 in 2-hop network ·dense cluster Open in graph ↗
What three separate factors drive chain-of-thoug… Does chain-of-thought reasoning reveal genuine inf… Do language models actually use their reasoning st… Does logical validity actually drive chain-of-thou… Does more thinking time always improve reasoning a… Do reasoning traces need to be semantically correc… What do models actually learn from chain-of-though…

Click a node to walk · click center to open · click Open in graph to see this note in the full knowledge graph

your link semantically near linked from elsewhere

Related papers in this collection 8

Papers most semantically related to this note, ranked by cosine similarity in the embedding space.

Search by related questions 4

Suggested questions this note speaks to — click to search the collection, or type your own.

Does each reasoning step in chain-of-thought introduce cumulative error?Why do logically invalid chain-of-thought examples work nearly as well?Can chain-of-thought faithfulness exist without causal necessity in reasoning?How much does pre-training frequency predict reasoning task performance?
Original note title

cot performance reflects three disentangled factors — output probability memorization and noisy reasoning