acceptodds
Under review as a conference paper at ICLR 2027

ChronoCheck: Scale Improves Per-Claim Faithfulness but Shifts Time-Series Rationales Toward the Claims Models Get Wrong

Abstract

Larger language models are more faithful on a fixed mix of claim kinds when they reason about time series, and they spend that gain on the kinds of claim they get wrong most. We measure this with ChronoCheck, a training-free oversight layer that decomposes a rationale into typed atomic claims (trend, extremum, periodicity, change-point, comparison, anomaly) and compiles each into a deterministic statistical check over the raw samples, so a rationale is refuted claim by claim. The checkers are conservative on the contradiction side: on a fresh suite of certified-true synthetic claims the contradiction criterion produced 0 false positives, and what they certify is consistency with the extracted claims rather than the truth of the rationale. Across three frozen scales chain-of-thought accuracy rises from 55.5% at 7B to 63.6% at 32B while the pooled contradicted-claim rate does not fall; held at the 7B mix of claim classes it falls from 36.5% to 31.5%, because claims naming a value or a location, the classes refuted most, take up a growing share of the rationale. Repair does not close the gap: no repair delta separates from zero at any Qwen2.5 scale (one ungated arm does on Llama-3.1-8B), and a large share of refuted claims disappear from the rewrite rather than being restated. Certify-abstain is the only arm to reach zero contradicted claims, at low coverage, with certified-subset accuracy gaps whose bootstrap intervals all cross zero. The checker measures time-series rationales; it does not repair them. We release it with the claim suite.

Primary area:
interpretability and explainable AI
Keywords:
time series, LLM reasoning, faithfulness, process supervision, claim verification, oversight

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_jw6wOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes ChronoCheck, a training-free layer that audits LLM chain-of-thought rationales about time series. The same frozen model runs a few-shot extractor that turns its own rationale into typed claims from six classes: trend, extremum, periodicity, change-point, comparison and anomaly. Each class compiles to a deterministic statistical checker on the raw series (Mann–Kendall/Theil–Sen, STL + ACF, PELT, robust z-scores).

Critical strengths
  • A verifier on the data side, with measured error. Each claim is compiled into a deterministic check on the raw series. The checkers’ contradiction precision and detection power are then measured on certified claims: 0 false contradictions out of 17,922 true claims (Table 3). Model-based verifiers cannot offer this. The released checkers and claim suite could be a useful auditing tool for time-series QA, and this is the main source of value I see in the paper.

  • Unusually careful and transparent reporting.

Critical weaknesses
  • The headline measurement is not validated on real rationales, and the paper’s own evidence suggests it is biased along the scale axis. The zero-false-positive calibration only covers slot-form claims emitted by the authors’ generator. It does not cover the natural-language→slot extraction, which is performed by the model under audit and differs across tiers. There is no human audit of extracted claims or of contradicted verdicts on real rationales. As a result, the 36.5–42.1% contradiction rates mix model unfaithfulness with extraction and schema error. The extremum class shows the problem. It drives the headline: ablating it removes 16.8 of the 36.5 points. Appendix A acknowledges that the schema turns local-extremum statements (“a trough around index 40”) into global-extremum claims. It also finds that, among refuted extremum claims carrying a value, 13.4% / 23.2% / 25.4% (7B/14B/32B) do name a local extremum within ±10 indices. These are plausibly true statements scored as contradictions, and their share grows with scale. I reclassified only those claims, using the paper’s own counts. The pooled rate moves from 36.5/42.1/39.0% to roughly 34.6/36.4/33.2%. The 7B→14B rise shrinks from +5.6 to about +1.8 points, and 32B ends up below 7B. That shift is the same size as the effect the title reports. The other refutations are checked only coarsely: a value within 2% of the range anywhere in the series, and a fixed ±10-index window. So the overall artifact could be larger or smaller than this. Without a human-validated sample per class and per scale, the claim that scale shifts rationales “toward the claims models get wrong” is not established.

  • The interventions give null or true-by-construction results, and the baselines are handicapped adaptations. With every intervention null, the contribution rests on the measurement in Weakness 1. All results also come from one greedy run per cell. Appendix B shows up to 153 of 887 questions change correctness across hardware.

Questions for authors
  1. For a random sample of real claims, stratified by class and scale (a few hundred claims), how often: (a) does the extracted claim faithfully represent its source sentence? (b) would an annotator looking at the series agree with the contradicted or consistent verdict?

  2. What are the pooled and standardised rates per scale if local-extremum statements are extracted as windowed claims, or at least treated as unverifiable? Please also state the extremum location tolerance and the index convention used in the rendered series (0- vs 1-based; global vs window-relative). Index misreadings would inflate contradictions specifically for location claims.

  3. The rate’s denominator includes unverifiable verdicts. The per-class declines behind the standardised fall include period (40.3→17.6%) and comparison (36.4→27.7%). Do these reflect more consistent verdicts, or more abstentions? Please report the consistent/contradicted/unverifiable split per class and scale. By my arithmetic, extremum and anomaly account for only about 1.2 of the 5.0-point standardised drop at 32B. So this question bears directly on the “per-claim faithfulness improves” part of the title.

1 comment
AuthorsOpenReview ↗

Response to Reviewer jw6w

Thank you for the review.

Extraction and verdicts on real rationales (Q1). The revised paper reports one calibration and two automated diagnostics on the real rationales. Table 3 gives 0 false contradictions on 17,922 certified-true synthetic claims. Appendix A reports that, among refuted extremum claims with a value, 97.1%, 98.6% and 98.4% (7B, 14B, 32B) name a value the series attains within tolerance. Among those that also carry an index, 13.4%, 23.2% and 25.4% name a local extremum within 10 indices of the reported index. With the 32B extractor applied to all tiers, the standardized rate is 36.9%, 34.2% and 33.2%. Under the shared extractor, the extremum and numeric class shares keep their order across scales and the anomaly share does not, so the paper states the class-mix sentence under each tier's own extractor. No human annotation of the extracted claims or of the verdicts has been completed, so the paper reports no human accuracy for them.

Local-extremum recoding (Q2). On the stored records, we re-scored the extremum claims. If refuted extremum claims with a value and an index that match a local extremum within 10 indices become unverifiable (83, 291 and 295 claims), the pooled rates are 34.5%, 36.5% and 33.5%, and the standardized rates are 34.5%, 29.4% and 27.6%. If extremum claims are re-judged in a 21-sample window centered on the stated index, the pooled rates are 35.8%, 40.0% and 36.1%, and the standardized rates are 35.8%, 31.6% and 29.5%. The pooled change from 7B to 32B is -1.1 points [-3.2, +1.1] in the first reading and +0.3 [-1.9, +2.5] in the second. The standardized change at 32B in the second reading is -6.3 points [-8.1, -4.6]. In eight of nine readings, the standardized rate falls with its interval below zero at both scales and the composition effect is positive. In the ninth, which treats every refuted extremum claim as unverifiable, the standardized rate still falls at both scales with its interval below zero.

Index convention and tolerance (Q2). The prompt shows the series in order, and a claimed index counts the displayed points from 0 to N-1. The checker subtracts the window start before the lookup. The location tolerance is 5 samples around a near-global extremum. The near-extremum band is 5% of the window range. The value tolerance is the larger of 2% of the range and 1e-9.

Verdict split per class and scale (Q3). The counts are consistent / contradicted / unverifiable at 7B | 14B | 32B-AWQ. Trend 469/74/445 | 516/67/352 | 495/66/304. Extremum 122/720/22 | 350/1284/2 | 303/1213/7. Period 135/147/83 | 277/122/166 | 351/110/164. Change point 403/140/627 | 360/81/442 | 494/71/431. Comparison 102/84/45 | 151/80/40 | 227/113/68. Anomaly 186/396/82 | 229/549/114 | 267/490/110.

Source of the standardized fall (Q3). At 32B against 7B, the standardized fall is 4.9 points. It equals a consistent share up 10.2 points [+8.3, +12.2] and an unverifiable share down 5.3 points [-7.1, -3.6]. Period falls from 40.3% to 17.6% contradicted while its consistent share rises from 37.0% to 56.2% and its unverifiable share rises from 22.7% to 26.2%. Comparison falls from 36.4% to 27.7% contradicted, its consistent share rises from 44.2% to 55.6%, and its unverifiable share falls from 19.5% to 16.7%. Extremum and anomaly contribute 1.2 of the 4.9 points, and the other four classes contribute 3.7. On verifiable claims only, the standardized change is -8.4 points [-10.6, -6.2].

Interventions and baselines. The registered 7B comparison after the reviews (Appendix E) compares the checker-guided repair with two alternatives. After round one, the mean per-question fraction of original contradicted claims with no match in the rewrite is 18.8 points lower under an explicit correction prompt and 19.6 points lower without feedback than under the original repair, and both intervals exclude zero. After two rounds the original repair gives +27 correct answers on the 394 questions relative to the archived chain of thought, against +3 for each alternative. On the three Qwen2.5 tiers, certify-abstain is the only arm that reaches zero contradicted claims, at 9.2-22.9% coverage; it is an offline selection rule. Matching each refuted claim to the rewrite leaves more refuted claims unmatched than restated as consistent (689 against 278 at 7B for ungated repair). The adapted baselines are tool-free adaptations, as stated in Section 4. These are the results of the specific variants under the current pool and budget.

Single greedy runs. Appendix B measures the launch-to-launch floor. Two launches on one box agree on every question (spread 0.0 points). Across boxes, up to 153 of 887 answers change correctness, and the net accuracy spread is at most 2.0 points. The five older comparisons that cross machines are marked with a dagger in the paper; the three new Appendix E arms ran in one generation.

Reviewer_8iAKOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper introduces ChronoCheck, a training-free framework for evaluating the faithfulness of LLM-generated reasoning about time-series data. The framework uses a frozen LLM to extract structured atomic claims from chain-of-thought rationales and verifies them against the underlying time-series data using deterministic statistical checkers. The authors evaluate ChronoCheck across three Qwen2.5 model scales. The results show that while answer accuracy improves with scale, the overall contradiction rate does not consistently decrease. The paper further investigates whether checker-generated feedback can help LLMs correct reasoning errors and improve answer accuracy, finding that the proposed repair strategies do not yield consistent improvements.

Critical strengths
  1. The paper proposes a structured and auditable framework for detecting reasoning errors. Compared with relying solely on LLM-based verification, the proposed framework uses deterministic statistical checkers, making the verification process more transparent.

  2. The paper proviee detailed analysis of reasoning errors. The framework can identify and record contradictions in LLM-generated rationales. The paper provides a detailed analysis of these errors.

  3. Interesting findings on the limitations of feedback-based repair. The authors investigate how contradicted claims change after feedback and find that models frequently delete these claims rather than correct them. More importantly, the finding that providing statistical counter-evidence does not consistently improve LLM answer accuracy is particularly interesting.

Critical weaknesses
  1. The writing of the paper is quite obscure, which hiders the comprehension and appreciation of the work. Section 1 Introduction is particularly unclear. For instance, It doesn't clearly formulate the problem that is being studied; It directly discuss some experiment results when nothing about the experimental setup has been explained. It doesn't clearly point out what's new in the paper and what's significantly, and it doesn't discuss such topics in the context of the literature.

  2. The paper demonstrated limited novelty and technical contribution. The paper does not clearly establish what is fundamentally new about the proposed problem formulation or methodology. Decomposing LLM-generated rationales into atomic claims, verifying them against external evidence, and using verification results to guide answer revision are established ideas. ChronoCheck adapts this general framework to time-series reasoning using deterministic statistical checkers, but the overall approach appears relatively straightforward. The introduction does not sufficiently explain what fundamental limitation of existing approaches is overcome or what substantive methodological advance ChronoCheck contributes.

  3. The paper has limited significance and generalizable insights. The empirical findings do not appear to provide sufficiently substantial new understanding of LLM reasoning, factual verification, or feedback-based correction. Although the paper reports changes in claim-level contradiction rates across model scales and limited accuracy improvements from repair, the broader significance of these observations remains unclear. In particular, the negative repair results do not establish whether the models struggle to utilize statistical counter-evidence or whether the particular feedback and revision protocol is ineffective. Prior research suggests that feedback presentation and correction strategies can substantially influence correction behavior (e.g., Chen et al., 2026; Wu et al., 2025), while feedback-assisted time-series reasoning has also produced positive results in a different, training-based setting (Su et al., 2025). Without disentangling these factors, it is difficult to draw generalizable conclusions about when and why feedback helps or fails.

  • Chen, Kuan-Yen, Fang-Yi Su, and Jung-Hsien Chiang. "The Self-Correction Illusion: LLMs Correct Others but Not Themselves." arXiv preprint arXiv:2606.05976 (2026).
  • Wu, Zhenyu, et al. "Enhancing mathematical reasoning in llms by stepwise correction." Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.
  • Su, Chen, Yuanhe Tian, and Yan Song. "Chain-of-thought Reviewing and Correction for Time Series Question Answering." arXiv preprint arXiv:2512.22627 (2025).
  1. Potential confounding from the feedback prompt and repair design. The repair prompt (in supplementary materials) explicitly instructs the model not to repeat refuted claims, while the paper reports that contradicted claims are more frequently deleted than corrected. This raises the possibility that the observed deletion behavior is partly induced by the prompt design. Moreover, the repair process regenerates the answer without preserving the original full rationale or explicitly identifying supported claims. These choices may influence the effectiveness of repair, but their effects are not isolated through experiments.

  2. Insufficient validation of the end-to-end verification pipeline. The framework relies heavily on the quality of atomic claims extracted by an LLM. Errors in extraction, or the intended semantics of a claim could lead to incorrect contradiction judgments. Although the authors evaluate the deterministic checkers extensively on synthetic claims, this does not fully establish the reliability of the complete extraction and verification pipeline on natural language rationales.

Questions for authors
  1. Have the authors investigated whether the limited effectiveness of repair stems from the models' difficulty in utilizing statistical counter-evidence, sensitivity to feedback prompt formulation, or instability introduced by answer regeneration?

  2. The repair prompt explicitly instructs the model not to repeat refuted claims, which may contribute to the observed tendency to delete rather than correct claims. Would a correction-oriented prompt that preserves the original rationale produce different results?

  3. Could the authors specifically analyze cases where the original final answer is correct but its rationale contains contradicted claims? How often does repair correct the reasoning while preserving the answer, remove contradicted claims without correcting them, or change an initially correct answer into an incorrect one?

  4. The current approach uses single-round repair. Have the authors considered allowing multiple rounds of verification and correction?

  5. Could the authors compare checker-guided repair against answer regeneration without feedback and evaluate the robustness of the results across multiple independent runs or alternative feedback prompts?

2 comments
AuthorsOpenReview ↗

Response to Reviewer 8iAK (1/2)

Thank you for the review.

Registered comparison (Q1, Q2, Q4, Q5). After the reviews we ran a registered post-review experiment on Qwen2.5-7B (Appendix E of the revision). It covers the 394 questions whose archived rationale holds a load-bearing contradicted claim. Each question starts from the same archived rationale in three arms: the original checker-guided repair, an explicit correction prompt that shows the previous reasoning and asks the model to keep supported claims, and no feedback. Each arm runs up to two rounds. We compare the repair arms with the no-feedback arm of the same run, which reruns the archived prompt. The correction arm changes the instruction and the access to the previous reasoning together. The intervals are descriptive paired percentile intervals.

Prompt and rationale (Q2). Yes, the results differ. After round one, the mean per-question fraction of the original contradicted claims with no match in the rewrite is 18.8 points lower under the explicit correction prompt than under the original repair (correction minus original: -18.8 points, interval [-23.6, -14.0], 375 complete pairs). Without feedback it is 19.6 points lower (-19.6 points, interval [-25.1, -14.2], 378 pairs). Both intervals exclude zero. The registered bounds, which set every unknown outcome to its least and most favorable value, also exclude zero: [-20.5, -15.4] and [-21.0, -16.6] points. In 19 and 16 of the 394 pairs an outcome is unknown. A claim with no match is a matcher outcome, not a judgment about the factual accuracy of the rewrite. Pooled over the 774 original load-bearing contradicted claims, 347 have no match after round one and 424 after round two. For the correction arm the counts are 195 and 225. For the no-feedback arm they are 170 and 170.

Multiple rounds (Q4). Two rounds raise the original repair from 202 to 217 correct answers on the 394 questions, against 193 for the no-feedback arm and 190 for the archived chain of thought. The correction arm reads 193 after both rounds. The no-feedback arm reads 193 after both rounds, and its second round reproduces its first. On the full 887-question pool, the original repair reaches 519 correct answers after two rounds, against 495 for each alternative and 492 for the archived chain of thought.

Answer changes (Q3). The checker-guided repair has the largest net gain of the three arms. After two rounds it is +27 correct answers on the 394 questions relative to the archived chain of thought (85 wrong to right, 58 right to wrong). The explicit correction arm gives +3 (47 and 44). The no-feedback arm gives +3 (24 and 21), so regeneration alone changes about as many answers in each direction. For questions that were originally correct, the single-round 7B experiment gives this breakdown: of 344 questions with an originally correct answer and a contradicted claim, one round of the original repair leaves 237 correct (131 with a contradicted claim recurring, 44 with at least one claim restated as consistent, 62 with every old claim unmatched or unverifiable) and 107 wrong.

AuthorsOpenReview ↗

Response to Reviewer 8iAK (2/2)

Why repair is limited (Q1). The same experiment compares three full treatments against the archived chain of thought. Regeneration alone, the no-feedback arm, gives a net of +3. A different prompt formulation, the correction prompt that keeps the rationale, gives a net of +3. The original checker-guided prompt gives +27 and has the largest matcher outcome. The experiment does not separate the individual factors inside each prompt.

Independent runs (Q5). The no-feedback arm is a second generation of the archived prompt under the same decoding. Appendix B measures the launch-to-launch floor directly. On one box and one software stack, two launches agree on every question (spread 0.0 points, 0 changed rows). Across boxes, up to 153 of 887 answers change correctness, and the net accuracy spread of the five arms is at most 2.0 points. The experiments do not cover repair outcomes under independent sampling seeds.

Writing and contribution. The Introduction states the question in its first paragraph: whether the claims in a model's time-series rationale are contradicted by the series. The second paragraph sets out the position. Atomic-claim verification usually relies on a learned verifier, whereas a trend, change-point, period or spike is a statistical question with deterministic checkers whose error we calibrate. Section 3 defines the claim schema and the six checkers, and Section 4 gives the setup. Three of the contributions bear on your question. The checkers return consistent, contradicted or unverifiable, with 0 false contradictions on 17,922 certified-true claims. The scale measurement separates the pooled rate from the rate at a fixed class mix. The claim-level matching shows what repair does to refuted claims.

Extraction and the full pipeline. Appendix A reports the extremum audit and the fixed-extractor check on real rationales. With the 32B extractor on all tiers, the standardized rate is 36.9%, 34.2% and 33.2%, and the 32B change against 7B is -3.7 points [-5.4, -2.0]. Under the shared extractor, the extremum and numeric class shares keep their order across scales and the anomaly share does not, so the paper states the class-mix sentence under each tier's own extractor.

Reviewer_p6yCOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes ChronoCheck, a training-free framework that verifies LLM-generated time-series rationales using typed atomic claims and deterministic statistical checkers. The authors find that larger models have lower contradiction rates when controlling for claim-type distributions, even though the overall contradiction rate does not improve. They also show that self-correction often removes incorrect claims rather than fixing them, while certify-abstain achieves zero detected contradictions at the cost of low coverage.

Critical strengths
  • Interpretable verification framework. Checking atomic claims directly against raw time-series data is a sensible approach. The deterministic checkers and synthetic calibration make the verification process relatively transparent.

  • Insightful empirical analysis. The distinction between pooled and standardized contradiction rates is interesting. I also appreciate the analysis showing that repair often deletes incorrect claims rather than correcting them.

Critical weaknesses
  • Uncertainty in claim extraction. The reliability of the claim extraction process is not sufficiently evaluated. As acknowledged in the appendix, local extrema may be interpreted as global extrema. Given the high contradiction rate in this category, extraction errors could meaningfully affect the main findings.

  • Limited generalizability of the scaling findings. The main conclusion relies on a post-hoc standardization of claim distributions. Moreover, controlling for six broad claim categories may not fully account for differences in claim difficulty, and the scaling analysis is mainly limited to Qwen2.5.

Questions for authors
  • Could the authors provide a manual evaluation of claim extraction accuracy on real model-generated rationales, particularly for extremum claims?

  • Does the scaling trend remain consistent when controlling for finer-grained differences in claim difficulty within each category?

1 comment
AuthorsOpenReview ↗

Response to Reviewer p6yC

Thank you for the review.

Extraction on real rationales (Q1). The revised paper reports one calibration and two automated diagnostics on the extraction. First, Table 3 gives 0 false contradictions on 17,922 certified-true synthetic claims for the checkers. Second, Appendix A compares every refuted extremum claim with the series by script. Among the 620, 1,263 and 1,203 refuted claims that carry a value (7B, 14B, 32B), 97.1%, 98.6% and 98.4% name a value that the series attains within the 2%-of-range tolerance. Among the 620, 1,253 and 1,162 that also carry an index, 13.4%, 23.2% and 25.4% name a local extremum of the 10-index window around the reported index. Third, Appendix A re-extracts the 7B and 14B rationales with the 32B extractor. The standardized rate is 36.9%, 34.2% and 33.2% (pooled 36.9%, 42.5%, 39.0%). Against 7B, it moves -2.7 points [-4.3, -1.1] at 14B and -3.7 [-5.4, -2.0] at 32B. Under the shared extractor, the extremum and numeric class shares keep their order across scales and the anomaly share does not, so the paper states the class-mix sentence under each tier's own extractor. No human annotation of the extracted claims or of the verdicts has been completed, so the paper reports no human accuracy for the extraction.

Local extrema read as global. On the stored records, we re-scored the extremum claims under nine readings, the submitted one included. In the first, refuted extremum claims that match a local extremum within 10 indices become unverifiable (83, 291 and 295 claims at 7B, 14B, 32B). The pooled rates are 34.5%, 36.5% and 33.5%. The rates standardized to the 7B class mix are 34.5%, 29.4% and 27.6%. In the second, extremum claims are re-judged in a 21-sample window centered on the stated index. The pooled rates are 35.8%, 40.0% and 36.1%. The standardized rates are 35.8%, 31.6% and 29.5%. The pooled change from 7B to 32B is -1.1 points [-3.2, +1.1] in the first reading and +0.3 [-1.9, +2.5] in the second. In the second reading, the standardized change at 32B is -6.3 points [-8.1, -4.6] and the composition effect is +6.6 [+5.5, +7.8]. In eight of the nine readings, the standardized rate falls with its interval below zero at both scales and the composition effect is positive. In the ninth, which treats every refuted extremum claim as unverifiable, the standardized rate still falls at both scales with its interval below zero.

Finer difficulty control (Q2). On the stored records, we split the six classes into finer cells in three ways: by claim content (19 cells), by question source group (18 cells) and by series length (18 cells). The standardized change against 7B at 14B and 32B is -4.1 [-6.0, -2.2] and -5.1 [-6.8, -3.3] for claim content. It is -3.2 [-4.8, -1.5] and -4.8 [-6.5, -3.0] for source group. It is -3.4 [-5.0, -1.7] and -4.8 [-6.6, -3.1] for series length. The interval lies below zero at both scales in all three splits.

Post-hoc standardization and model family. The standardization fixes the class weights at the 7B mix and reports each tier's own per-class rates under those weights. Figure 1 shows the pooled rate and the standardized rate together. The three scales are Qwen2.5-7B, 14B and 32B-AWQ. The 32B bf16 control gives the same direction: -5.8 points [-7.5, -4.1] against 7B. The abstract and the body state the scaling results for these Qwen2.5 scales and for claims extracted by each tier's own model.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Title change in the revision

Title change. The revised PDF is titled "ChronoCheck: Scale Improves Per-Claim Faithfulness at a Fixed Claim Mix, and Shifts Time-Series Rationales Toward the Claims Models Get Wrong". The submitted title was "ChronoCheck: Scale Improves Per-Claim Faithfulness but Shifts Time-Series Rationales Toward the Claims Models Get Wrong". The added phrase states the condition of the first finding, as the abstract does: the contradicted-claim rate falls with scale when the mix of claim kinds is held fixed, and the pooled rate does not fall.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.