acceptodds
Withdrawn from ICLR 2027

Learning What to Revise After a Prediction Error

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

We investigate sequential diagnosis, considering whether models can use new evidence to distinguish a faulty observation from a changed state or governing rule, and whether lightweight fine-tuning on synthetic examples can improve their ability to identify what revisions benefit real scenarios. We build matched triplets in which the three causes are indistinguishable until later evidence arrives, with an exact Bayesian reference at every step, and find that language models compute the diagnosis but do not report it. LocusProbe, a linear probe on Llama-3.1-8B's hidden states, identifies the cause on real weather series with 0.97 accuracy against 0.44 for the same model's answers, and beats every tested open direct-answer system on matched items, including the nine-times-larger Qwen2.5-72B by 0.125 [0.080, 0.173]. Posterior-matched tuning (PMT) on synthetic arithmetic and logic alone transfers to real weather and industrial series with no real labels; applied to the unembedding only, with every other weight frozen, it raises rule-change recall from 0.03 to 0.72. When the diagnosis selects the revision, probe-selected revisions recover 98.5–99.9% of the gap to an oracle over 60 steps and attain a worst-case loss of 6.63 against 14.98 for an observer given the exact generative family of all three causes. Recall on rule changes, not overall accuracy, predicts that recovery ( versus ), because an uncorrected rule compounds. These results identify a learnable component of revision, i.e., using subsequent evidence to determine what went wrong, what should change, and whether that revision allows the system to recover.

Primary area:
foundation or frontier models, including LLMs
Keywords:
sequential diagnosis, belief revision, fault attribution, prediction error, world models, language models, hidden-state probing, synthetic-to-real transfer, time series, anomaly diagnosis, recorded faults, fine-tuning

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_DHbnOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies the classification of surprising observations in time series as faulty observations, state displacements, or rule changes. The authors contribute synthetic benchmarks in which these three possibilities occur with prior uniform probabilities. They find that a linear probe, LocusProbe, trained on Llama 3.1 8B's hidden states, can make this diagnosis more reliably than the base LLM prompted without thinking on the synthetic benchmarks. Alternatively, they train a LoRa adapter to match the exact posterior on the synthetic data, and they show that the adapter beats the base model on real (recorded-fault) weather data, water-circulation data, computing cluster data, an industrial-control testbed, and aircraft flights with failures (Table 7). The impact of this diagnosis is further measured in revised forecasts based on the updated state level or new rule.

Critical strengths
  1. The authors develop a synthetic dataset such that training produces an adapter that is more useful than non-adaptation to classify surprise observations with an LLM.
Critical weaknesses
  1. No serious baselines are reported. The LLMs are forced to choose a diagnosis in the next token with no chain of thought. At the very least, chain of thought should be compared.

  2. Additionally as baselines, a statistical or rule-based method, operating on the values of the time series and implementing the principles used to compute the exact posterior, should be compared. Given the possibility of using the time series values directly, it is not clear why an LLM or a probe should be used to begin with, and hence the impact of the research is questioned.

  3. The modeling and synthetic data assume that observation faults or state disturbances operate at a single reading. In practice, future observations are also likely to be corrupted if a sensor becomes faulty, and a state disturbance such as turbulence of an aircraft may displace the dynamics differently at many consecutive observations.

  4. The family of rules being considered is never stated. In the synthetic example in Table 1, the surprise reading at time point 4 could be explained by multiplication by 11/6, not just by addition of 5. However many time points are observed, infinitely many rules may exist to explain all the readings, unless the rule family is limited to begin with.

Questions for authors
  1. Why is only a PMT adaptor and not LocusProbe tested on the real (recorded-fault) datasets (Table 7)?

  2. Layer choices seem inconsistent. Layer 16 is shown to have minimal effect compared to later layers for residual patching (line 275, Table 18), but layer 16 is used for the probe (lines 297, 798, 1215).

  3. On lines 144-150, it would be better to say the exact posterior matches the prior. It is uniform only in the setting of the synthetic benchmark. The comments about the next reading depend on weaknesses #3 and #4. A study of the robustness of LocusProbe and PMT when the class priors or fault duration are varied, or with repeated disturbances, or with richer rule families, would strengthen the paper.

Reviewer_x2E8OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper investigates whether LLMs can identify the causes of unexpected observations in time-series prediction tasks and determine what should be revised to improve future predictions. The authors define three possible causes: observation faults, state disturbances, and rule changes, and introduce a synthetic benchmark for diagnosis. The paper presents two relatively independent methodological contributions. First, the authors train a linear probe on LLM hidden states to predict the underlying error's cause. They show that the hidden representations contain the error cause information that can be decoded more accurately than the LLM's textual answers. Second, they introduce Posterior-Matched Tuning (PMT), which SFT fine-tunes LLMs to align their predicted probability distribution over the three error causes with a pre-calculated Bayesian posterior from the benchmark's data-generating mechanisms.

Critical strengths
  • The problem of identifying what to revise after a prediction error is interesting and practically relevant.
  • The paper includes extensive experiments across synthetic and real-world datasets, including downstream recovery evaluations.
Critical weaknesses
  • Extremely poor clarity and organization. The paper is exceptionally difficult to follow. The task formulation, methodological pipeline, and connections between different experiments are not clearly explained.

  • Unclear central contribution. The paper combines probing, posterior-matched fine-tuning, synthetic-to-real transfer, fault diagnosis, and forecasting recovery without establishing a coherent central narrative.

  • Limited novelty and conceptual contribution. The proposed methods primarily combine existing techniques, such as linear probing and fine-tuning with soft targets, without introducing substantial methodological novelty. Moreover, the empirical findings do not appear to offer fundamentally new conceptual insights beyond existing observations about the gap between LLMs' internal representations and their textual outputs.

  • Interpretation of probing results. Linear decodability of causal information does not necessarily demonstrate that LLMs internally perform causal diagnosis.

  • Insufficient methodological clarity. The construction of Bayesian reference posteriors, training supervision, and differences between experimental settings are not explained clearly enough in the main text.

  • Limited generality of downstream evaluation. Forecast recovery relies on predefined revision operators, making it unclear how well the results generalize to more realistic settings requiring learned revisions.

Questions for authors

Please refer to the weaknesses

Reviewer_PDnwOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper seeks to identify the root cause of a prediction mistake in a time series problem, creating a benchmark for the problem and an LLM probe for the identification. Three types of causes are considered: "observation", "state disturbance", and "rule change". An experiment is reported for a range of real and synthetic datasets, comparing how LLMs can natively perform the identification against a Llama model equipped with the proposed probe.

Critical strengths

Root cause analysis is an important problem, and testing / understanding when LLMs can do the task well is critical for understanding current AI systems' limitations.

Critical weaknesses

Root cause analysis has a very long history, which is all but ignored in the paper. Moreover, Granger causality is especially developed for the time series setting, yet "Granger" never appears in the paper. Two recent, related papers that are not cited are: Root Cause Analysis of Anomalies in Multivariate Time Series through Granger Causal Discovery, and A Causal Approach to Detecting Multivariate Time-series Anomalies and Root Causes.

Unclear primary contribution: the paper claims that both the benchmark and the probe are a contribution. However, when proposing a benchmark, it is common to explain why exactly it is necessary and to report previous approaches' results on that benchmark. On the other hand, probe are a common tool, and thus the paper should explain why this particular form of probe is especially well-suited for the root-cause problem and improves upon previous neural-network-based approaches (that I have linked above).

Questions for authors

Can the probe actually do causal identification? Or is 'causal' used in the paper in only a non-technical sense?

What is the relationship to Granger causality?

Why weren't any previous methods for this problem tested on the benchmark? If they are not relevant / suitable, explain why.

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

Related papers

Discussion (0)

Sign in to comment.