University of Tokyo's CSA Makes LLMs Accurately Admit When They're Wrong
A new self-assessment method turns sampled LLM outputs into a multiple-choice question, beating semantic entropy with as few as two extra samples
Takeaways
- New method: Clustered Self-Assessment (CSA) turns sampled LLM outputs into a multiple-choice question and uses the model's own token probabilities as a confidence score.
- How it works: Sample N responses, cluster semantically equivalent ones via NLI, present clusters as answer options, read the model's probability over each option letter.
- Beats baselines: Outperforms Semantic Entropy, tokenSAR, EigV, and perplexity-based methods across TriviaQA, Natural Questions, and multiple model families.
- Sample efficient: Achieves competitive performance with as few as 2 additional samples, far fewer than most competing methods require.
- Limitations: Requires logit access (not purely black-box), inherits NLI clustering noise, and cannot detect confident hallucinations where the model is consistently wrong.
- Related work: Companion paper Semantic Token Clustering (EACL 2026) from the same lab achieves UQ in a single generation pass with no auxiliary model.
LLMs are remarkably good at sounding right even when they are wrong. The deeper problem is that most models have no reliable way to signal when they are guessing. A new paper from the University of Tokyo proposes a surprisingly clean fix: instead of measuring uncertainty indirectly through entropy math, ask the model to pick its own best answer from a structured menu of options it generated itself.
The method is called Clustered Self-Assessment (CSA), and it consistently outperforms a wide set of baseline uncertainty quantification (UQ) methods across multiple models and open-domain question-answering datasets, while requiring as few as two additional samples beyond the original generation.
Why knowing what you don't know is so hard
Uncertainty quantification can be used to assess the reliability of LLM outputs, and as LLMs are rapidly deployed in high-stakes domains such as healthcare and law, it becomes increasingly important to have meaningful measures for the trustworthiness of model-generated content. The challenge is that the open-ended nature of LLM generation hinders direct application of traditional UQ methods, which treat LLM outputs as autoregressive sequences and consider only lexical uncertainty. Since response correctness fundamentally depends on semantics, and distinct token sequences can convey identical meanings, uncertainty in the semantic space serves as a more reliable indicator of trustworthiness.