Parsa Mirtaheri: Let Me Think! A Long Chain-of-Thought Can Be Worth Exponentially Many Short Ones

Value of the Information & Strength of the Argument

The talk provides valuable insights into the fundamental trade-off between sequential and parallel scaling of chain-of-thought. The empirical results are compelling, showing a clear exponential advantage for sequential scaling in the short CoT regime. The theoretical analysis, including the vertex query model and TC0 reductions, strengthens the argument by providing formal evidence for the observed phenomena. The argumentation is coherent and well-structured, moving from empirical observations to theoretical explanations. The connection to RL and the potential explanation for the emergence of longer CoTs in models like DeepSeek-R1 adds practical relevance. However, the talk focuses on a specific synthetic task, and the generalizability to real-world reasoning tasks is not fully addressed, which limits the scope of the conclusions.

Scientific Rigor, Source Quality, Title Accuracy

The talk is based on a peer-reviewed paper (to appear in NeurIPS), which ensures a certain level of rigor. The authors build on previous work on transformer expressivity and chain-of-thought, citing relevant literature. The experimental setup is carefully designed, with controlled tasks and multiple models. The title accurately reflects the content, and the talk stays on topic. The description provides minimal context, but the talk itself is self-contained. The sources cited are primarily the paper itself and related works mentioned in the talk, but no external URLs are provided in the description, so the sources_citees field is limited to the paper reference.

The title accurately reflects the core message of the talk: that a single long chain-of-thought can be exponentially more effective than many short ones, supported by both experiments and theory.

8/10

The talk presents original research with both empirical and theoretical components, including formal proofs and experiments on transformers. The methodology is rigorous, and the results are supported by theoretical analysis. However, the presentation is a seminar talk, and the full details are in the paper, so some aspects are summarized.

Markers derived by PSI from the transcript: the creator did not define chapters.

• 00:00Introduction and overview of inference-time scaling and the two paradigms: parallel and sequential scaling.
• 02:00Definition of the ST1T2 connectivity task and the bridge graphs used in the study.
• 05:00Training models with different chain-of-thought strategies: shortest path, path, and DFS trace.
• 08:00Key empirical result: short CoT models fail exponentially, while DFS model succeeds; accuracy matches probability of in-distribution DFS traces.
• 11:00Introduction of parallel scaling and aggregation methods (majority voting and best-of-n verification).
• 14:00Trade-off results: sequential scaling has exponential advantage in short CoT regime; single long CoT worth exponentially many short ones.
• 17:00Theoretical evidence: vertex query model and proof that sequential scaling is necessary for bridge graphs.
• 20:00Theoretical results on TC0 complexity: polynomial-length CoT can solve task, constant-length CoTs cannot.
• 23:00RL experiments: reinforcing out-of-distribution traces leads to natural sequential scaling and improved accuracy.
• 26:00Discussion of pre-commitment phenomenon and potential barriers to sequential scaling.

• Let Me Think! A Long Chain-of-Thought Can Be Worth Exponentially Many Short Ones — The paper being presented, to appear in NeurIPS.

• Chain-of-Thought Prompting Elicits Reasoning in Large Language Models — Seminal work on chain-of-thought prompting, supporting the importance of CoT.
• Training Verifiers to Solve Math Word Problems — Related to verification-based aggregation methods.

• Let's Think Step by Step: Reasoning in Large Language Models — This work suggests that CoT may not always improve reasoning, contrasting with the positive results shown here.