acceptodds
Under review as a conference paper at ICLR 2027

Belief-Trajectory Energy: Measuring the Path to a Prediction

Abstract

Large language models (LLMs) progressively revise their predictions across Transformer layers, yet we typically observe only the final output, discarding the trajectory through which it is formed. We introduce Belief-Trajectory Energy (BTE), a model-grounded measure that characterizes an input through the layerwise predictive revisions it induces in a model. By mapping intermediate states into a shared predictive space, BTE provides a principled measure of belief change that can be summarized as either a scalar or a structured depth profile. Theoretically, we show that local BTE corresponds to predictive revision under the Fisher–Rao geometry, while the sequence of revisions captures information beyond the initial-to-final belief change. Empirically, scalar BTE provides a model-relative signal of difficulty across diverse reasoning tasks, while richer BTE representations support human–LLM review detection and fine-grained generator attribution, reaching up to macro-AUROC and eight-way attribution accuracy. Further analysis shows that BTE develops throughout pretraining and is selectively reshaped by targeted training, demonstrating that the resulting measurement reflects what the scoring model has learned. Together, our results establish belief trajectories as a principled model-grounded signal and suggest a broader perspective in which learned models can themselves serve as instruments for measuring and characterizing the data they process.

Primary area:
interpretability and explainable AI
Keywords:
LLMs, Interpretation, LLM-generated peer review, information geometry

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_MAnTOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposed Belief-Trajectory Energy(BTE)that projects logis into predictive distribution and then calculate the SJD between layers

Critical strengths

1 The method is simple, requiring only a single forward pass and no training, and is applicable to any open-source model. 2 The experiments cover a wide range of areas—difficulty, detection, attribution, and training dynamics—and have been repeatedly validated across multiple scorers. 3 There are numerous controlled experiments, including length residualization, writer control, splitting the training and test sets by paper, transfer across generator families, and human rewriting. 4 The authors candidly report the limitations, including length heterogeneity, scorer dependency, and failed transfer to new generators.

Critical weaknesses

1BTE is essentially logit lens plus layer-wise JSD, so novelty is limited. 2The difficulty signal is largely explained by response length. 3Cross-task ordering reverses for three scorers, contradicting the authors’ explanation. 4Baselines are zero-shot while BTE is supervised, and some are sign-flipped. 5GPT-5.5-rewritten human reviews are still labeled human, suggesting format cues. 6The Ethics statement and Appendix A.4 contradict each other on who did the human evaluation. 7Training analysis is weak, with a non-significant final OLMo ρ and a single seed. 8 The 95.6% accuracy excludes the human class, and attribution is closed-set with no stylometric baselines.

Questions for authors

1 How does BTE compare with simple baselines such as final-layer NLL, entropy, and response length for difficulty?

2 Why was the middle-layer window chosen, and do results hold under a pre-registered or tuned-lens readout?

3 Why does weak Mistral also reverse the cross-task ordering if the cause is math training exposure?

4 How does BTE compare with supervised baselines trained on the same data, such as per-layer NLL or a fine-tuned RoBERTa detector?

5 Why are GPT-5.5-rewritten human reviews still classified as human, and does the detector rely on review format?

6 Was the human evaluation done by the authors or by six recruited evaluators?

7 Do the fine-tuning results hold across multiple seeds with confidence intervals?

Reviewer_gyneOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces Belief-Trajectory Energy (BTE), which measures changes in next-token predictive distributions across Transformer layers using the logit lens and Jensen–Shannon divergence. The authors evaluate scalar BTE as a model-relative difficulty signal and use richer trajectory features for review detection and generator attribution. They also examine how BTE changes during pretraining and targeted fine-tuning. The metric is simple, its interpretation is well motivated, and the detection results are strong. The main open questions concern the contribution of the specific trajectory representation, its difficulty information beyond length, and its robustness to shifts in the generator population.

Critical strengths
  1. BTE has a clear interpretation in predictive space. The local connection to Fisher–Rao geometry motivates measuring distributional revisions, while the trajectory retains information that cannot be recovered from the initial and final distributions alone.

  2. The empirical evaluation includes paper-disjoint review detection, cross-family transfer, human-review rewriting, and length controls. Generator attribution is also strong within the evaluated source set: the eight-class classifier reaches 95.6% source accuracy on generated reviews from held-out papers, with human recall reported separately.

  3. The training analyses help explain what the metric measures. Pretraining and targeted fine-tuning change BTE in different ways, providing evidence that scorer familiarity contributes to the signal.

Critical weaknesses
  1. The experiments do not fully isolate the contribution of the specific trajectory construction. The paper includes existing detection baselines and several BTE ablations, but matched comparisons with alternatives such as layerwise entropy profiles or initial-to-final JSD would help establish what is gained specifically from consecutive-layer JSD. Using the same scorer, data splits, and classifier would make this comparison more informative.

  2. The difficulty association is substantially reduced after controlling for length. On MATH-500, controlling for both response and problem length leaves significant partial correlations of 0.151 for Llama and 0.191 for OpenMath2, while the other two reported scorers are no longer significant. The paper acknowledges that BTE is not length-independent. The remaining question is how much useful difficulty information it adds beyond simple length features.

  3. Detection transfer remains sensitive to the training setting and scorer. Qwen-BTE performs very well when a generator family is held out, but legacy-to-new transfer is less reliable, especially for Llama. The paper discusses this limitation and the role of scorer familiarity. The strong detection results should therefore be interpreted in light of the chosen scorer and training-generator coverage; the observed legacy-to-new drop limits the evidence for cross-generation robustness.

Questions for authors
  1. Under the same scorer, data splits, and classifier, how do consecutive-layer JSD features compare with layerwise entropy profiles and initial-to-final JSD for review detection?

  2. Could the authors compare length-only and length-plus-BTE predictors of difficulty on held-out data, and add bootstrap confidence intervals for the length-adjusted correlations in Table 16?

Reviewer_iLtrOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes BTE to measure changes in an LLM’s intermediate predictions. It applies the model’s final output head to each layer’s hidden states to obtain token distributions. It measures differences between consecutive layers using Jensen–Shannon divergence. These differences are averaged across tokens to form a depth profile. A middle-layer average is used to estimate difficulty. Simple classifiers are trained on the profile to detect generated text and identify its source. The scoring LLM remains frozen.

Critical strengths
  1. Each coordinate measures predictive change at a particular depth. This makes the features easier to inspect than an arbitrary high-dimensional hidden-state vector.
  2. The authors compare raw profiles with normalized profiles that retain only their shape. This helps investigate whether useful information comes from how much predictions change or where those changes occur.
Critical weaknesses
  1. The interpretation of BTE as belief revision is insufficiently validated. The final output head may decode intermediate layers with different accuracy. BTE could therefore reflect changes in representation compatibility rather than genuine changes in predictive beliefs. The paper does not experimentally separate these effects.
  2. The paper does not compare JSD with KL or symmetrized KL. Its theoretical justification does not establish that JSD produces better measurements or downstream results.
  3. On MATH-500, Llama’s correlation with difficulty drops from 0.444 to 0.151 after controlling for problem and solution length. This weakens the evidence that BTE captures difficulty beyond these simpler properties.
  4. The difficulty estimate depends strongly on layer selection. Middle and late layers can show opposite relationships with difficulty. The chosen middle-layer window has an empirical motivation, but its reliability across unseen tasks and architectures remains uncertain.
  5. Strong results with known generators or held-out families do not guarantee performance on future model generations. Detectors trained only on older generators sometimes reverse their predictions on newer ones.
Questions for authors
  1. JSD measures change without identifying its direction. Can BTE distinguish a useful correction from movement toward an incorrect answer?
  2. For the same problems, can it predict success across multiple independent generations after controlling for solution length, token likelihood, and benchmark identity?
  3. If one generator produces reviews with different prompts, quality levels, and writing styles, does its BTE signature remain stable?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.