← the graph
Sector 1 · 1 sub-themes

Reasoning & Test-Time Compute

Fresh evidence keeps chipping away at the premise that chain-of-thought reasoning is real reasoning. Researchers at IIIT Allahabad and NIELIT Delhi tested a wide range of proprietary and open models, including ChatGPT and Claude, and found the step-by-step traces often don't reflect what actually drives the answer. That echoes earlier Anthropic and Apple work on unfaithful reasoning, a debate that has since dissolved into dueling rebuttals. The response is to move reasoning off the page. A new survey argues that externalizing every step as text is wasteful and pushes latent, non-text reasoning. Chelsea Finn and collaborators go further, using RL to optimize the reasoning strategy itself, running parallel trains of thought and aggregating them.

IIIT AllahabadNIELIT DelhiChatGPTClaudeAnthropicAppleChelsea Finnchain-of-thought faithfulnessLLM reasoning transparencylatent reasoningreinforcement learning for reasoningparallel reasoning aggregationinterpretability critiqueexplicit CoT efficiencyopen-source vs proprietary models
1.1

LLM Reasoning Validity Debated Across CoT, Latent, and RL Methods

  • A paper from researchers at the Indian Institute of Information Technology Allahabad (IIITA) and the National Institute of Electronics and Information Technology (NIELIT) in Delhi found that, across a significant range of proprietary and open-source LLMs including ChatGPT and Claude, chain-of-thought reasoning presented to users is "decorative" — invented after the model has already decided its answer, rather than driving it. The paper offers a cheap and easily replicable method to test whether step-by-step reasoning outputs reflect the actual inference process. [5]
  • This finding follows a series of earlier high-profile studies from Anthropic and Apple indicating that so-called "reasoning AIs" often produce step-by-step explanations that don't reflect what actually informed their answers, though that prior debate devolved into conflicting rebuttals and interpretations without resolution. [5]
  • A survey on latent reasoning (circulated via a dedicated thread) argues that forcing LLMs to externalize every reasoning step in text is computationally wasteful, and proposes keeping reasoning in latent (non-text) space rather than requiring explicit chain-of-thought output. [2][1]
  • Chelsea Finn and collaborators describe an RL-based approach that optimizes not just sequential reasoning but also the reasoning strategy itself, including parallel trains of thought and aggregation of parallel reasoning paths. [3]
  • Google Research found that allowing an LLM to reason (rather than directly retrieve) helps the model recall facts it already knows, suggesting reasoning improves factual memory access and is not limited to solving novel hard problems. [4]
Indian Institute of Information Technology AllahabadNIELITChatGPTClaudeAnthropicAppleChelsea FinnGoogle Researchchain-of-thought reasoninglarge language modelslatent reasoningreasoning faithfulnessreinforcement learning for reasoningparallel reasoning pathsfactual memory retrievalLLM inference transparency
Sources
  1. [1]che_shr_cat — 1/ Chain-of-Thought works, but forcing LLMs to write out every step in text is a massive computational waste. What if we could keep the reas★twitter
  2. [2]burny_tech — A Survey on Latent Reasoning "Large Language Models (LLMs) have demonstrated impressive reasoning capabilities, especially when guided by extwitter
  3. [3]chelseabfinn — LLM RL optimizes for sequential reasoning We also optimize over the reasoning strategy, incl parallel trains of thought, aggregation of paratwitter
  4. [4]VukRosic99 — Reasoning doesn't just solve hard problems - it helps an LLM *remember facts it already knows*. New Google Research finding 👇 Letting an LLMtwitter
  5. [5]Ronald_vanLoon — Chain-Of-Thought Reasoning Proven ‘Decorative’ in Major Language Models by Martin Anderson @UniteAi Learn more: https://t.co/YJDNvgOoAj #AI twitter