- Rating
- 2
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper argues that behavioral forecasts for large reasoning models (LRMs) are hard to obtain, since long reasoning trajectories are often unfaithful when read as text, and local attribution or uncertainty methods operate on single generation steps. The authors propose treating behavior forecasting as a supervised learning task, by training a Behavior Forecaster model to predict statistics that would otherwise require resampling the LRM from a single sampled trajectory.
In their experiment, the authors resample an LRM 10 times per prompt and 10 times per modified prompt to obtain gold behavioral labels, using OLMo-3-7B-Think across three datasets. The evaluation focuses on two main aspects, namely 1) Rerun consistency, i.e. the fraction of other runs on the same prompt that reproduce the observed answer; and 2) Counterfactual sensitivity, i.e. the probability drop for the observed answer when prompt segments are removed independently.
Compared to two capable baselines (GPT-5.4 and Claude Opus 4.6), the trained forecaster obtains higher correlation scores in both metrics. The paper also reports transfer to held-out data and ablations on input format, initialization and freezing.
Critical strengths
-
The problem of future prediction is well defined and motivated, and the proposed evaluation approach is well designed. The setup in Section 2.1 is general and could cover other behaviors that can be labeled automatically in a similar fashion. The authors estimate uncertainty at the level of original samples, which I believe is the right choice since segments and runs from the same prompt are not independent. They also report negative results, such as weak results on TreeCut and weak zero-shot transfer across datasets. This makes the main comparisons credible as reported.
-
The rerun consistency results show that the reasoning trace contains useful signal about answer stability that is not picked up by frontier models reading the raw trace output. Gains over baselines are relatively large and have confidence intervals well above zero (Table 13). Removing the reasoning from the input significantly affects performance (Table 4), confirming that the signal comes from the reasoning tokens. The same trend holds on a held-out FEVEROUS subset and on Qwen3.5-2B. As discussed below, however, most of this gain comes from the RuleTaker dataset.
Critical weaknesses
-
The experiments show that the proposed behavior forecaster works, but it is not clear whether it truly relies on the LRM's reasoning. The paper claims that the forecaster learns how the target LRM encodes information in its reasoning, but removing the reasoning from the input barely changes the counterfactual sensitivity results (0.723 vs. 0.714 in Table 4), while still beating both frontier model baselines even when these can see the reasoning. This suggests that the forecaster might be picking up spurious shortcuts, e.g. learning which premises are logically needed for the answer, while ignoring the actual LRM behavior. This seems especially likely for RuleTaker, which makes up 72% of the test segments. For rerun consistency, reasoning seems to matter, but it remains unclear whether the forecaster uses patterns specific to the LRM, cues such as trace length, or whether the answer is correct. The fact that a randomly initialized model already beats the frontier models by a wide margin on this metric suggests that cues may be sufficient for obtaining these results. Since the paper’s framing depends on this point, I don't consider its central claim to be supported as of now.
-
The comparison with the frontier models is not fully fair, since trained models can learn typical label values for each dataset while the zero-shot prompted baseline models do not. Per-dataset results (Table 10) show that the picture is not as clear-cut, with Opus outperforming the forecaster on FEVEROUS, and all methods performing poorly on TreeCut. Since RuleTaker makes up most of the test segments, the gap can be attributed to the trained predictor specializing in this setup, also in light of the shortcut mentioned in the previous point.
-
Training models to predict quantities that normally require resampling is also a common approach: for example, Semantic entropy probes [1] predict how much answers vary across reruns from one generation, FastSHAP [2] learns explainers to predict the effect of removing input features and, more recently, Future Probes [3] predict the proportion of LRM traces resulting in specific behaviors of interest from intermediate activations at a given reasoning position. I see the generalization tests as the most valuable contribution that distinguishes this work from e.g. [3], but results show that, without labels from a new dataset, the trained forecaster does worse than the frontier models on all three datasets for counterfactual sensitivity, and on two of three for rerun consistency (Table 2 vs. Table 10), so the generality claim is not well supported.
[1] Kossen et al., 2024. Semantic Entropy Probes: Robust and Cheap Hallucination Detection in LLMs. https://arxiv.org/abs/2406.15927 [2] Jethani et al. 2022. FastSHAP: Real-Time Shapley Value Estimation. https://arxiv.org/abs/2107.07436 [3] Kortukov et al. 2026. Predicting Future Behaviors in Reasoning Models Enables Better Steering. https://arxiv.org/abs/2606.11172
Questions for authors
-
How much of the rerun consistency label depends on whether the answer is correct? It would be helpful to present the forecaster's results on correct and wrong answers separately to highlight whether this is a confounder in the analysis.
-
Did you test whether the OLMo forecaster predicts Qwen’s behavior as well as OLMo’s? There are some results in the literature regarding models' "priviledged access" to their own computations (see: https://arxiv.org/abs/2511.08579v1). This would help to show whether initializing from the target model really matters.
-
For counterfactual sensitivity, could you compare against a simple baseline based on the RuleTaker gold proofs to see how much the LRM reasoning is actually helpful, as opposed to general resolution traces?
2 comments
Typical labels, novelty and generalization (2/2)
The comparison with the frontier models is not fully fair, since trained models can learn typical label values for each dataset while the zero-shot prompted baseline models do not. Per-dataset results (Table 10) show that the picture is not as clear-cut [...]
The typical label of each dataset cannot explain the forecaster's lead. Typical label values alone give a Spearman correlation of at most 0.22 and 0.30, the score of a predictor that outputs each dataset's mean test label on counterfactual sensitivity and rerun consistency, far below the forecaster (0.731 and 0.571). Also, simpler forecasters trained on the same labels can learn the typical values just as well, and they are less accurate: a prediction head trained on the frozen LRM reaches 0.245 and 0.475, and a randomly initialized backbone 0.534 and 0.477 (Table 5, previously Table 4).
Training models to predict quantities that normally require resampling is also a common approach [...]
Thank you for these pointers. These works share one ingredient with ours, training a model to predict a quantity that otherwise requires resampling, and the revised paper cites all three (lines 118-119, 483 and 489-490). Their question is different though. They take a quantity that already has an accepted way to compute it and make it cheaper. We take a question that in practice is answered by reading the LRM's reasoning, and propose to answer it by learning from the LRM's past behavior instead, for any behavior that can be labeled automatically. We show that learning does better: a 7B model trained on the LRM's past behavior forecasts it more accurately than frontier models that read the full trajectory. This matters most for LRMs, whose behavior unfolds over long autoregressive generations: measuring it directly takes many runs of thousands of tokens each, while the forecaster needs one forward pass over the observed trajectory and, on rerun consistency, matches the accuracy of one to two full reruns (lines 1230-1259).
without labels from a new dataset, the trained forecaster does worse than the frontier models on all three datasets for counterfactual sensitivity, and on two of three for rerun consistency (Table 2 vs. Table 10), so the generality claim is not well supported.
The forecaster generalizes zero-shot within a dataset family: trained on three FEVEROUS subsets, it stays ahead of both readers on the held-out numerical subset (0.565 against 0.322 and 0.463 on counterfactual sensitivity, and 0.311 against 0.223 and 0.173 on rerun consistency; Table 2, previously Table 1). Zero-shot transfer to an entirely different domain, for example from logic rules and fact verification to math word problems, is a different matter: with only three training datasets, we would not expect it from any model on any language task. The transfer section now states this, and its table includes the readers' results (lines 376-380 and Table 3, previously Table 2).
Did you test whether the OLMo forecaster predicts Qwen's behavior as well as OLMo's? [...] This would help to show whether initializing from the target model really matters.
We added this experiment to the appendix (lines 1375-1418, Tables 19 and 20). The answer depends on the task. For counterfactual sensitivity, a forecaster trained on OLMo's behavior forecasts Qwen's as well as one trained on Qwen's own (0.782 against 0.784, same Qwen3.5-4B backbone), and OLMo's own weights are not needed to forecast OLMo (0.746 from Qwen3.5-4B, against 0.731). For rerun consistency, the target matters on both counts: on RuleTaker, where Qwen is almost always consistent and OLMo is not, a forecaster trained on Qwen's behavior reaches 0.113 on OLMo, against 0.455 for one trained on OLMo's, and initializing from Qwen2.5-7B reaches 0.416, against 0.568 from OLMo's own weights. We made this clearer in the abstract (lines 24-27) and the ablation section (lines 465-470), and now cite Li et al. (2025).
For counterfactual sensitivity, could you compare against a simple baseline based on the RuleTaker gold proofs to see how much the LRM reasoning is actually helpful, as opposed to general resolution traces?
It is true that a solver built on the gold proofs may do well on RuleTaker, but it has to be designed by hand for one dataset and one behavior. What the paper demonstrates is that an LRM's behavior can be learned from its past behavior: the same training recipe, with no knowledge of how any dataset is built, gives forecasters that are more accurate than frontier readers on both behaviors.
Please let us know if this resolves your concerns. If it does, we would appreciate it if you would update your score.
Shortcuts do not explain the forecaster’s lead (1/2)
Thank you for the careful and detailed review, and we are glad that you found the problem well motivated, the evaluation well designed and the main comparisons credible. Your questions about what the forecasters learn made the paper more precise. Following your comments, we uploaded a revised PDF that makes clearer what the reasoning tokens contribute in each task, adds the two analyses you asked for (rerun consistency on correct and wrong answers, and forecasters trained on one LRM and evaluated on another), moves the per-dataset results into the main text, states the limits of the generalization results, and cites the three works you pointed to.
It is true that for counterfactual sensitivity the reasoning tokens add little. To clarify our claim: a trajectory is the prompt, the reasoning and the answer together (Section 2), and we claim that naive reading of a trajectory does not tell the full story about the LRM's future behavior, and that the rest can be learned from the LRM's past behavior. We do not claim that the forecaster always needs the reasoning tokens. For rerun consistency it does: removing them drops the forecaster from 0.568 to 0.316. For counterfactual sensitivity, what stands is the result that you describe: a forecaster that never sees the reasoning (0.714) is more accurate than two frontier readers that do (0.672 and 0.690). We made it clearer in the paper that the contribution of the reasoning tokens applies to rerun consistency and not to counterfactual sensitivity (lines 458-463), and the setup sentence that you refer to is now limited to rerun consistency (lines 323-325).
The forecaster might learn the logic to some degree, but it is important to remember that it is a 7B model that runs one forward pass and the readers are far stronger solvers of logic problems than any 7B model. If the task were only to find which statements are logically needed, they would be at least as accurate as the forecaster. They are not (0.649 and 0.601, against 0.668). The result also does not rest on RuleTaker: the forecaster's margin over the better reader is larger on FEVEROUS (0.660 against 0.621) and on TreeCut (0.840 against 0.804). We moved the per-dataset results into the main text (Table 1, previously Table 10).
We ran this analysis and added it to the appendix (lines 1260-1284 and Table 16, referenced from the main results at lines 342-343). Correctness is informative about the label, but it does not explain the forecaster's accuracy. Within correct answers, its Spearman correlation is 0.443, against 0.073 for GPT-5.4 and 0.004 for Claude Opus 4.6. Within wrong answers it is 0.526, against 0.008 and 0.162. It is the readers' pooled correlation (0.305 and 0.308) that mostly reflects correctness. Trace length does not explain it either: controlling for both length and correctness, the forecaster's correlation with the label is 0.41, against 0.04 and 0.07 for the readers.
We read this result as support for the paper's thesis. A randomly initialized forecaster knows only what it learned from the target LRM's trajectories and their behavior labels, and this alone takes it past the frontier readers (0.477 against 0.305 and 0.308). Whatever it learned, it is a regularity of this LRM's behavior that reading does not pick up.