Can Aha Moments Be Fake? Towards Quantifying Decorative and True Thinking in Chain-of-Thought
arXiv:2510.24941v4 Announce Type: replace Abstract: Large language models can generate long chain-of-thought (CoT) reasoning, yet prior work suggests that CoT can be post-hoc rationalization rather than a faithful reflection of the computation through explicitly designed settings. In this work, we go further and propose a True Thinking Score (TTS) to quantify the causal contribution of each step in CoT to the model's final prediction in realistic reasoning problems. Across eleven models ranging from 1.5B to 1.1T parameters on common reasoning benchmarks, we find that CoTs often interleave true
The proliferation of advanced large language models necessitates a deeper understanding of their reasoning processes beyond superficial chain-of-thought outputs.
Quantifying 'true thinking' directly impacts the trustworthiness, explainability, and reliability of AI agents and systems, particularly for critical applications.
We can now quantitatively assess the genuine problem-solving contribution of each step in an LLM's reasoning, rather than merely observing its generated output.
- · AI developers focused on transparency
- · Auditors of AI systems
- · Researchers in AI interpretability
- · Developers relying solely on CoT for performance metrics
- · Systems with opaque AI decision making
Improved debugging and optimization of large language models' reasoning capabilities.
Development of more robust and verifiable AI agents that can clearly demonstrate their reasoning.
Increased societal trust in AI systems due to enhanced transparency and reduced 'post-hoc rationalization'.
This signal links to a primary source. Continuum Brief monitors and indexes it as part of the live intelligence stream — we do not republish source content.
Read at arXiv cs.LG