- Rating
- 3(+1)
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper proposes a stopping rule for streaming hallucination detection that aims to control the probability of incorrectly interrupting a faithful generation, independently of its length. The method converts scores from existing hallucination detectors into p-values, transforms them into e-values, and accumulates evidence across generation checkpoints using a sequential framework. The authors provide theoretical guarantees under conditional calibration assumptions and evaluate the method across several hallucination detection benchmarks
Critical strengths
- The paper addresses a practically relevant problem: controlling false alarms across arbitrarily long generations is important for deploying streaming hallucination detectors.
- The proposed framework is principled and detector-agnostic, building on established techniques from sequential hypothesis testing.
- The experimental analysis is detailed and covers multiple benchmarks configurations. I also appreciate that the authors explicitly discuss cases where their method underperforms simpler baselines.
Critical weaknesses
-
The theortical guarantee provided in the papaer relies on conditional calibration assumptions that are not ensured by the marginally calibrated variant used in the main experiments. The authors introduce a certified conditional variant, which comes at a noticeable cost in detection power. This means there is a gap between the theoretical contribution and the empirical results.
-
My main problem with the paper is that the empirical improvements over the baselines are very limited. SeqGuard does not consistently outperform calibrated threshold-based methods in terms of detection power. Overall, it is unclear whether the additional complexity is justified by the empirical gains.
-
The evaluation does not sufficiently establish the effectiveness of the method on long generations. The main experiments use relatively short responses, with a median of approximately five sentence-level checkpoints. This is particularly important given that length independence is a central motivation of the paper.
-
Detection delays are relatively large (32–73 tokens) and most successful detections occur after the hallucinated span has already been generated. So unclear what the practical benefits of the approach are.
Questions for authors
- Can the authors evaluate the method on substantially longer generations?
3 comments
I'd like to thank the authors for the detailed rebuttal and extra experiments. The new long-form evaluation addresses my main concern, and I appreciate the clearer distinction between the marginal and certified variants.
I still have two reservations:
- Limited practical gains: The new experiments show substantially lower detection power for SeqGuard (14.2% vs. 39.3% for the threshold), although at different false-alarm rates. It remains unclear whether a more conservatively calibrated threshold could achieve similar false-alarm control with better power.
- Practical utility. Detection delays remain substantial (~155 tokens), which means most hallucinations are already generated before detection. This limits the practical benefits of the approach.
Overall, the rebuttal strengthens the paper, particularly the evidence for length-independent false-alarm control. While some concerns remain, I am inclined to raise my score from 2 to 3.
Thank you for re-reading the paper and for raising your score. On your two remaining points:
- Pooled over the long-form set, a threshold tuned at half the budget does match SeqGuard: realized FAR .043 and power .222 at nominal α = .05, against .046 and .195 for SeqGuard at α = .10. The difficulty is knowing the right correction in advance. The threshold’s realized FAR runs from 0 to 1.9 times its budget depending on α and output length, so it can only be set with labelled long outputs. SeqGuard needs no adjustment. We will add this comparison to the paper.
- We agree the delay is substantial. The threshold itself takes 120 to 130 tokens on these outputs, so most of the delay comes from the detectors, and about 25 to 35 tokens from accumulation. The rule is a stop signal that saves the rest of a long output, not a preventer, and the paper says so.
Thank you again; the long-form experiment exists because you asked for it, and it is now the paper's main result.
Reviewer Z4ik: can you say whether you are convinced by the rebuttal, particularly the question of how close the experiments are to the theory adn whether that is enough?