acceptodds
Under review as a conference paper at ICLR 2027

Belief Maintenance under Retraction in Large Language Models

Abstract

Retracting an inferred belief requires withdrawing its support while reassessing the conclusions that depend on it. Such changes are nonmonotonic (NMR): removing one justification can invalidate a conclusion, restore an alternative, or enable a new inference. We investigate how large language models (LLMs) perform this belief maintenance by formalizing two procedures: editing an existing reasoning state and rebuilding it from revised information. Within a temporal argumentation framework, we compare these procedures under shared counterfactuals and support cuts. We establish conditions for agreement on belief values and show that identical current values can conceal differences in support histories that emerge under subsequent retraction. Using NMR-Retraction, we evaluate whether LLMs perform support withdrawal and belief revision or preservation in accordance with the semantics of each procedure. Experiments reveal substantial difficulties in downstream retraction and history-dependent revision, despite stronger performance in accepting counterfactual assumptions. Supervised fine-tuning improves performance on unseen graphs, showing that these maintenance capabilities can be strengthened through training. Together, these findings highlight the importance of procedural knowledge for reliable belief maintenance in LLM reasoning.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
Large Language Models, Belief Maintenance, Belief Retraction, Nonmonotonic Reasoning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_QSd5OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper focuses on two approaches to belief maintenance in LLM reasoning: editing and rebuilding. Editing revises existing conclusions and their supporting inferences, whereas rebuilding rederives new conclusions from the initial facts and rules under changed assumptions. To compare these two approaches, the paper conducts experiments within a temporal framework that simultaneously records the ordered process of rule application and the current belief values.

The paper introduces NMR-RETRACTION to evaluate the modifications required and the reasonable retention involved in each maintenance procedure. Formal semantics benchmarks are used to provide symbolic reference answers for belief computation that support target beliefs and their derived conclusions. Existing models do not consistently accept counterfactual assumptions, and their average accuracy in predicting the impact of retraction behavior on related beliefs is low. Further analysis indicates that it is difficult to use retained reasoning history to determine which beliefs should be revised or retained, and that supervised fine-tuning can improve models’ belief maintenance capability.

Critical strengths

The paper introduces a framework for nonmonotonic belief maintenance in LLM reasoning based on support dependencies and ordered rule-application histories, which helps researchers better understand how the LLM retains or changes its beliefs.

Critical weaknesses

Regarding problem modeling, the explanation of the modeling process is not sufficiently clear, making it difficult to fully understand. In terms of experiments, the evaluation of the experimental methods lacks cross-method comparison, and the experimental evaluation process lacks details; for example, no specific evaluation criteria are provided for the counterfactual acceptance rates and independent retraction accuracies of the various models and methods.

Questions for authors

In what aspects do you think your method is superior to other related research methods?

3 comments
AuthorsOpenReview ↗

Response to Questions for Authors

We thank the reviewer for their careful review and positive assessment, and for highlighting where our problem formulation and experimental evaluation need clearer explanations.


What advantages does your approach offer over related methods?

We thank the reviewer for this question. Related work approaches belief change from different perspectives. LLM knowledge editing examines whether changing a fact (e.g., “France’s capital is Paris” → “France’s capital is Rome”) produces consistent changes in semantically related knowledge. Symbolic belief revision and defeasible reasoning study how beliefs and conclusions change when facts, rules, or their justifications are withdrawn.

Our work focuses on a practical question from the perspective of a user observing an LLM’s reasoning: when a particular inferred belief is withdrawn or counterfactually changed, which parts of the existing reasoning should be revised, and which should remain valid? A user may be interested in a particular conclusion and its consequences without needing to inspect the entire belief set. Our framework supports this selective examination by making the relevant support dependencies and maintenance operations explicit. Its advantage lies in connecting a user’s targeted revision request with the operations required to carry it out and their verifiable semantic consequences.

We formalize Rebuild and Edit to provide procedure-specific reference labels for the queried beliefs. These labels distinguish counterfactual acceptance from correct downstream revision and preservation, while revealing when the two procedures yield different outcomes.

We believe our formulation provides a precisely specified starting point for examining practical belief maintenance and extending evaluation to broader representations and alternative policies for retraction and counterfactual revision. Such extensions are important because our rule-based semantics do not capture every form of belief maintenance in LLMs: beliefs may involve uncertainty, priorities may differ from recency, and some reasoning may not be fully expressible through propositions and rules.

Weakness 1. Ambiguity in the research setting

We begin with grounded beliefs G={b1,b2,…,bk}\mathcal G=\{b_1,b_2,\ldots,b_k\} and rules R\mathcal R that produce an inference chain:

\qquad A \xrightarrow{r_{AB}} B \xrightarrow{r_{BC}} C.$$ Here, $A$ is derived from the grounded beliefs, $B$ from $A$, and $C$ from $B$. A counterfactual changes $B$’s value; we use $B'$ as shorthand for $B$ with its newly assigned value. Under explicitly supplied support cuts and maintenance instructions, we examine three operations: | Operation | Evaluation question | |---|---| | Forward retraction | When $B$ is withdrawn, is $C$ also withdrawn? | | Backward retraction | When $B$ is withdrawn, is its support from $A$ also withdrawn? | | Counterfactual propagation | When $B$ changes to $B'$, is $B'$ correctly propagated? | For this example, the specified cut removes $r_{AB}$, and no alternative support preserves $C$. The two maintenance procedures can be illustrated as follows: - **Rebuild:** Start from $\mathcal G\cup\{B'\}$, use $\mathcal R\setminus\{r_{AB}\}$, and recompute. - **Edit:** - withdraw $C$’s support from $B$ - withdraw $B$’s support from $A$ - replace $B$ with $B'$ - propagate $B'$. - **Theoretically**, we ask which propositions receive the same values under Rebuild and Edit, and which may receive different values. Our analysis characterizes conditions for agreement and explains how differences in retained support histories can lead to different outcomes, including under subsequent retraction. - **Experimentally**, we evaluate whether LLMs correctly predict these proposition values under the specified procedure, using symbolic reference labels. ---- **Weakness 2. Insufficient detail on evaluation procedures and metrics** **1. Independent retraction evaluation** Forward and backward retraction are evaluated independently, with the correct preceding state supplied. We score the queried belief values immediately after the designated operation, before subsequent inference. This isolates retraction accuracy from errors in earlier stages. **2. Counterfactual preservation and propagation** After assigning $B'$, we evaluate whether the model preserves this value through subsequent reasoning and correctly determines other belief values. | Component | Evaluation question | |---|---| | CF preservation | Does $B$ retain its assigned value $B'$ after subsequent reasoning? | | On-CF | Are beliefs reached by propagation from $B'$ correctly updated or preserved? | | Off-CF | Are beliefs outside that propagation path correctly updated or preserved? | The CF target is excluded from the On-CF and Off-CF evaluations and scored separately. Importantly, Off-CF does not mean unaffected: withdrawing support can change beliefs outside the CF propagation path. Conversely, a proposition reached by CF propagation may retain its original value despite a change in support. We therefore distinguish required value changes from justified preservation in both regions. **3. Evaluation metrics** Each queried proposition has a symbolic reference value $y_i\in\{\mathrm T,\mathrm F,\mathrm U\}$ under the specified procedure and evaluation stage. Here, $\mathrm U$ denotes an undetermined value. Let $\hat y_i$ be the model prediction and $b_i$ the baseline value. - **Belief Update (BU):** Does the model correctly change a value that should change? - **Belief Maintenance (BM):** Does the model correctly preserve a value that should remain unchanged? $$\mathcal Q_{\mathrm{BU}}=\{i:y_i\neq b_i\}, \qquad \mathcal Q_{\mathrm{BM}}=\{i:y_i=b_i\}.$$ For either subset, we measure exact-match accuracy: For either subset, we measure exact-match accuracy: $$\operatorname{Acc}(\mathcal Q)=\frac{1}{\operatorname{card}(\mathcal Q)}\sum_{i\in\mathcal Q}\mathbf{1}[\hat y_i=y_i].$$ Here, $\operatorname{card}(\mathcal Q)$ denotes the number of evaluated queries. For final-state evaluation, $b_i$ is the initial belief value. For independent retraction diagnostics, $b_i$ is the value immediately before the evaluated operation. Counterfactual preservation measures whether the final prediction at the CF target matches its assigned value $c_i$: $$\operatorname{CFPres}=\frac{1}{\operatorname{card}(\mathcal Q_{\mathrm{CF}})}\sum_{i\in\mathcal Q_{\mathrm{CF}}}\mathbf{1}[\hat y_{i,\mathrm{final}}=c_i].$$ Cross-procedure comparisons distinguish matched queries with identical reference answers from queries scored against each procedure’s own reference answer.

Based on our theoretical framework and experimental results, we make the following empirical observations:

  1. Belief updates remain difficult. Models preserve unchanged values more accurately than they perform required retraction and revision, even with explicit instructions.

  2. Rebuild and Edit have different strengths. Edit generally performs better on backward retraction and shared-gold Off-CF updates; Rebuild performs better or comparably on On-CF updates and better on Off-CF preservation.

  3. SFT improves belief revision. Direct-answer supervision yields substantial gains that transfer to CoT and MultiTurn prompting.

We will make these evaluation stages and metric definitions explicit in the revised manuscript.

Reviewer_mcA4OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper is concerned with belief maintenance as applied by large language models. It does so by making very specific assumptions about how belief maintenance is performed and checking whether, and under what conditions, belief maintenance performed by LLMs correspond to these. The paper show that deviations from what is expected can be mitigated through training. The upshot of this is that what the author(s) refer to as procedural knowledge plays an important part in getting LLMs to perform belief maintenance.

Critical strengths

The topic of the paper is very relevant for the area of belief maintenance and, more generally belief revision, and a detailed investigation of the extent to which LLMs perform belief revision is sorely needed.

Critical weaknesses

The main weakness of the paper is the very specific assumptions being made about the type of revision that it studies. One of the main strengths of belief revision theory is the broad class of revision (and contraction or withdrawal) operations that have been studied, and the abstract frameworks for describing revision operators in terms of rationality postulates. The canonical example here is the AGM postulates for revision and contraction (the author(s) are aware of this (the original paper describing these is the first reference in the bibliography), but since their original work, there has been a number of different, but related frameworks proposed as well. These frameworks allow for the evaluation of whole classes of revision operators, rather than focusing on a few specific operators. By making specific assumptions, instead of focusing on an abstract framework, describing a class of operators, the impact of the evaluation being performed is mitigated, and the significance of the results is decreased.

Questions for authors

The main question I have for the author(s) is this: What was your motivation for defining belief maintenance in the particular way in which you decided to do it (based on Hansson's kernel-based support withdrawal), rather than, looking at a class (or classes) of revision operators based on abstract principles?

4 comments

Hi everyone,

I don't see any author responses to questions raised by any of the reviewers. The other reviews also do not change my views about the paper substantially. I'm therefore keeping my scores unchanged.

Dear Reviewer mcA4,

We apologize for the delayed response. We are organizing the results of our additional GRPO experiments, which took more than 60 hours to run. We provide our response to your questions below and will share the complete GRPO results within a day.


We sincerely thank the reviewer for their careful reading and thoughtful feedback, particularly for situating our work within the broader literature on belief revision and contraction. As the reviewer notes, we cite the foundational AGM work and recognize the importance of postulate-based frameworks for characterizing rational belief change operators. However, our goal is to investigate support-based belief maintenance in LLMs, and we will clarify its relationship to these broader frameworks in the revised manuscript.

We would like to distinguish the characterization of admissible belief changes from the operational problem of maintaining a reasoning state under a specified withdrawal. For instance, partial meet contraction determines which beliefs to retain by selecting and intersecting maximal subsets that no longer entail a target proposition. The kernel-based perspective motivates our focus on explicit support dependencies. In our evaluation, the support cut is supplied, and we examine how the resulting reasoning state is maintained.

Partial Meet ContractionRebuildEdit
FocusDetermine a contraction resultRecompute from revised informationUpdate an existing reasoning state
MechanismSelect and intersect maximal subsetsReconstruct conclusions and historiesRetract affected supports and update dependencies
HistoryNot explicit in the classical belief-set formulationReconstructedPreserved where valid

Our theoretical contribution lies in formalizing and distinguishing Rebuild and Edit under temporal argumentation semantics. Drawing on defeasible logic and argumentation theory, we investigate how support withdrawal can invalidate previously derived conclusions or reinstate previously defeated ones.

The focus of this analysis is retraction and its consequences, rather than an axiomatic justification of the rationality of belief maintenance as a whole. Edit involves proposition-value retraction, forward retraction, backward retraction, counterfactual-value assignment, and subsequent inference. The ordering of these operations is a design choice, and our framework additionally specifies a temporal priority policy in which more recently applied rules take precedence in determining current proposition values. These choices define the semantics under which we analyze retraction. Our theoretical results characterize the consequences of these choices and the relationship between the two procedures; they do not establish that the chosen ordering or priority policy satisfies a general set of belief-change postulates.

Our evaluation asks whether LLMs can correctly perform retraction under these explicitly specified conditions. We adopt deterministic semantics to isolate this reasoning challenge without introducing the additional problem of selecting among alternative contraction results. The reference outcomes are uniquely defined relative to the prescribed procedure, rather than asserted to be uniquely rational across all belief change operators.

We will clarify this scope in the revised manuscript and make the operation-ordering and temporal-priority assumptions explicit. We consider a postulate-based characterization of these procedures and an extension to broader operator classes important directions for future research, and we are exploring how to pursue them. These extensions would complement our analysis but fall beyond the scope of the present paper, which focuses on the reasoning required for retraction under specified semantics.

AuthorsOpenReview ↗

Response to the Reviewer’s Questions

On the Reviewer’s Question: Why Support-Based Retraction?

Our motivation was to study the reasoning required for support withdrawal and subsequent belief maintenance, both theoretically and empirically. We distinguish grounded beliefs from inferred beliefs because changing an inferred belief requires addressing its supporting inference paths. Simply overwriting an intermediate proposition value leaves the original justification intact, so subsequent inference may restore the withdrawn value. Our formulation therefore interrupts the relevant support paths to maintain consistency with the counterfactual change.

This introduces a choice of which supporting elements to withdraw. Hansson’s kernel-based approach addresses this support-selection problem by identifying minimal support sets and selecting elements whose removal breaks those supports. Drawing on this perspective, we specify the support cut and temporal semantics in our evaluation to isolate whether LLMs correctly execute the withdrawal and maintain the resulting reasoning state.

A postulate-based analysis across operator classes would complement this work by examining broader rationality requirements.
To our knowledge, directly adapting existing frameworks to defeasible reasoning with explicit support histories and temporal rule priorities is not straightforward.

In particular, how should an LLM use retained reasoning history to decide which beliefs to retract, revise, or preserve in accordance with appropriate rationality postulates? We consider this an important but challenging question, reminiscent of the frame problem’s concern with determining what should change and what should remain unaffected. We do not yet see a straightforward general answer and are considering this as a future research direction beyond the present paper’s focus on reasoning under specified retraction semantics.

We would therefore appreciate suggestions for relevant belief change operators or references that could guide this extension.

AuthorsOpenReview ↗

Gentle Follow-Up Before the Discussion Deadline

Dear Reviewer mcA4,

Thank you again for your insightful comments on the theoretical scope of our work.

In the revised manuscript, we have clarified our focus on support-based belief maintenance under specified retraction semantics, while acknowledging the importance of broader postulate-based approaches.

We hope these revisions help clarify our research objectives and would welcome any further thoughts you may have before the discussion period ends.

We sincerely appreciate your time and consideration.

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

The paper studies how LLMs maintain beliefs when an inferred belief is retracted. It formalizes two maintenance procedures in a temporal argumentation framework with ordered rule-application histories: EDIT, which retracts support in the existing reasoning state and then propagates a counterfactual, and REBUILD, which recomputes beliefs from the revised belief base. Proposition 2 gives sufficient conditions under which the two procedures agree on current belief values in regions of the reasoning graph, and Theorem 1 shows that states with identical current values can diverge under a later retraction because their support histories differ. The NMR-Retraction benchmark consists of six-layer synthetic graphs with signed unary rules, with reference answers computed by a symbolic engine. On six Qwen3 and Gemma-3 models, counterfactual preservation is 84.0% (Direct) and 81.9% (CoT), backward retraction 38.7% and 68.8%, and forward retraction of downstream beliefs 16.0% and 47.1% (R-Test, 360 graphs). On BU queries outside counterfactual paths where EDIT and REBUILD prescribe different values, the six-model mean stays below 40%. Supervised fine-tuning on 900 graphs improves accuracy on 1,800 held-out graphs, and a one-epoch GRPO run gives smaller gains.

Critical strengths
  1. The theory is solid and makes the evaluation interpretable. Proposition 2 gives sufficient conditions for regional value equivalence between EDIT and REBUILD, and Theorem 1 shows that current-value equivalence is not sequential equivalence. Together they let the paper separate differences between the procedures' prescribed outcomes from errors in executing them. This is the main reason the benchmark results can be read as procedural failures of the models.

  2. The benchmark is carefully built and the results are clearly layered. Reference answers come from a symbolic engine under the stated semantics. The results separate three outcomes across six Qwen3 and Gemma-3 models: counterfactual preservation (84.0% Direct, 81.9% CoT), backward retraction (68.8% with CoT), and forward retraction of downstream beliefs (only 47.1% with CoT).

Critical weaknesses
  1. External validity is not tested. All tasks are synthetic graphs with signed unary rules between adjacent layers, and there is no natural-language or real-task evaluation. The paper lists natural-language settings as future work (Section 6), so it is not known whether the observed failures, or the SFT gains, carry over beyond this formal setting.

  2. The evidence for the central history-dependent revision claim is thin. The test that separates EDIT from REBUILD outside counterfactual paths (Section 5.3, Figure 5) uses only 82 BU queries (77 in Table 6). Off-CF accuracy on the larger set of 1,800 BU queries appears only as the before-training baseline in Tables 7 and 9 and does not isolate the queries where the two procedures disagree. Appendix D.4 also states that this single-update test cannot directly test the sequential divergence of Theorem 1. The paper also reports no confidence intervals, and each query gets one greedy response, so small differences such as the slightly higher agreement under EDIT instructions in Figure 5 cannot be separated from noise. This amount of evidence does not support a claim of this strength.

  3. The training results are weaker than the claim that training strengthens these maintenance capabilities, and the RL comparison is weak. After two SFT epochs, accuracy rises strongly on counterfactual paths, but under CoT, Off-CF BM accuracy falls on four of the six main models, by up to 18.7 points (Qwen3-4B, from 39.2 to 20.5), and on Llama3.3-70B from 72.8 to 36.6 (Table 7b). GRPO is trained for one epoch on five models and gives average gains below 10%, with some regressions (Appendix D.6). SFT is trained for two epochs, both use seed 0 only, and the paper notes that the two training budgets are not compute-matched. The comparison therefore says little about whether SFT is more efficient than RL for this task.

Questions for authors
  1. (Clarification) For the 82 queries in Figure 5, how many distinct graphs do they come from, and what are the per-model confidence intervals for agreement under EDIT and REBUILD instructions? With this sample size, can the difference between the two instructions be separated from noise?

  2. (Clarification) Have you run any natural-language version of the task, even a small one, for example the same graphs written as everyday statements? A result of this kind would change my assessment of external validity.

  3. (Clarification) Did you try GRPO with a budget matched to SFT, for example two epochs? Do the Off-CF BU regressions in Table 9 remain?

7 comments

We sincerely thank the reviewer for their detailed feedback on the experimental evaluation and concrete suggestions for improvement. We first respond to the reviewer’s questions, then explain how the expanded experiments address the concerns raised in the weaknesses.

Question 1. Sample size and statistical reliability

The 82 queries in Figure 5 come from 82 distinct graphs; the 77-query subset in Table 6 likewise contains 77 distinct graphs. Figure 5 measures agreement with the Edit reference under both instructions, rather than procedure-specific accuracy. We have now rerun the evaluation on revised datasets with explicitly defined matched-query cohorts and compare Backward Edit as an ablation of Edit that omits forward retraction. We report 95% graph-level bootstrap confidence intervals for these expanded experiments, while distinguishing same-gold and different-gold comparisons. These intervals do not establish the statistical significance of the original Figure 5 comparison.

Question 2. Natural-language evaluation

The original submission did not report a natural-language evaluation. For the rebuttal, we added matched symbolic and natural-language versions of 500 Synthetic test graphs, together with evaluations on adapted ProofWriter and BoardGameQA datasets containing 944 and 291 test graphs, respectively. The Synthetic versions share the same graphs and gold labels; the native datasets use explicitly documented CF-rule adaptations.

Question 3. GRPO training budget and regressions

The original comparison used two SFT epochs and one GRPO epoch, both with seed 0, and was not compute-matched. We have extended GRPO to a single three-epoch trajectory and evaluate checkpoints at epochs 1, 2, and 3. Matching the epoch count does not match computational cost. In the new evaluation, Because these experiments use revised data and training protocols, they are additional evidence rather than a direct replication of Table 9.


Issue 1. Clarifying the 82-versus-77 Query Counts

The two counts refer to different evaluation subsets. Figure 5 uses 82 Off-CF update queries where Edit and Rebuild prescribe different values. Table 6 additionally requires the Edit reference to differ from Backward Edit, leaving 77 queries within the original set of 82. Although these counts follow different selection criteria, their presentation creates avoidable ambiguity. We therefore no longer use these original evaluation subsets in the revised analysis and replace them with explicitly defined matched cohorts.

In the expanded evaluation, we match the same graph, activation variant, intervention, cut, and query across Edit and Backward Edit, then report same-gold and different-gold cohorts separately. Same-gold queries allow direct comparisons on identical update or maintenance targets. For different-gold queries, each procedure is scored against its own oracle answer.

Matched final-query accuracy: same gold

DataCohortEditBackward EditBoth correct
SynCommon Update0.17 [0.16, 0.17]0.18 [0.17, 0.18]0.08 [0.07, 0.08]
SynCommon Maintenance0.80 [0.80, 0.81]0.85 [0.85, 0.85]0.74 [0.74, 0.75]
Syn NaturalCommon Update0.04 [0.04, 0.05]0.04 [0.04, 0.04]0.01 [0.01, 0.01]
Syn NaturalCommon Maintenance0.80 [0.79, 0.80]0.81 [0.81, 0.82]0.72 [0.72, 0.73]
PWCommon Update0.23 [0.22, 0.23]0.25 [0.25, 0.26]0.13 [0.12, 0.13]
PWCommon Maintenance0.83 [0.83, 0.84]0.88 [0.88, 0.89]0.79 [0.79, 0.79]
BGQACommon Update0.24 [0.22, 0.25]0.26 [0.24, 0.27]0.14 [0.12, 0.15]
BGQACommon Maintenance0.86 [0.85, 0.86]0.89 [0.88, 0.90]0.81 [0.80, 0.81]

On queries with identical gold labels, Backward Edit generally achieves slightly higher accuracy, suggesting that the additional forward-retraction stage may introduce further opportunities for error.

Issue 2. Strengthening the Evidence for History-Dependent Revision (Theorem 1)

We evaluate paired states with identical current proposition values but different valid support histories, applying the same CF intervention and support cut. The primary cohort requires different answers after retraction; a separate control cohort requires the same answers.

Our screening found only 17 eligible cases in ProofWriter and none in BoardGameQA. We therefore constructed 1,000 Synthetic test graphs, each providing one different-answer pair and one same-answer control pair: 2,000 pairs, or 4,000 individual trajectories per representation and evaluation condition. Symbolic and natural-language versions share the same graphs.

Synthetic: Symbolic

ModelDifferent Answers After RetractionSame Answers After Retraction
Average across models0.02 [0.02, 0.03]0.22 [0.21, 0.23]
Qwen/Qwen3-14B0.01 [0.01, 0.02]0.01 [0.00, 0.01]
Qwen/Qwen3-4B-Instruct-25070.04 [0.03, 0.05]0.63 [0.61, 0.65]
google/gemma-3-12b-it0.04 [0.03, 0.04]0.01 [0.01, 0.02]
google/gemma-3-4b-it0.01 [0.01, 0.01]0.24 [0.22, 0.25]

Synthetic: Natural Language

ModelDifferent Answers After RetractionSame Answers After Retraction
Average across models0.01 [0.01, 0.02]0.31 [0.30, 0.32]
Qwen/Qwen3-14B0.01 [0.01, 0.02]0.01 [0.01, 0.02]
Qwen/Qwen3-4B-Instruct-25070.00 [0.00, 0.00]0.50 [0.50, 0.51]
google/gemma-3-12b-it0.04 [0.03, 0.04]0.44 [0.42, 0.46]
google/gemma-3-4b-it0.01 [0.00, 0.01]0.28 [0.26, 0.29]

Support-History Input Ablation

We additionally omit the support histories while retaining the graph, current values, and intervention instructions. This tests the effect of providing history information; omitting it removes information needed to distinguish the two states in different-answer pairs.

DatasetSupport HistoriesDifferent Answers After RetractionSame Answers After Retraction
Synthetic (symbolic)Provided0.02 [0.02, 0.03]0.22 [0.21, 0.23]
Synthetic (symbolic)Omitted (current values only)0.01 [0.01, 0.01]0.19 [0.18, 0.19]
Synthetic (natural language)Provided0.01 [0.01, 0.02]0.31 [0.30, 0.32]
Synthetic (natural language)Omitted (current values only)0.00 [0.00, 0.00]0.20 [0.19, 0.21]

Theorem 1 shows that identical current belief values can lead to different outcomes after the same retraction because their support histories differ. Our pair-level evaluation tests whether models correctly handle both histories: a pair receives credit only when every final query is answered correctly for both states. The very low accuracy on different-answer pairs indicates substantial difficulty executing the history-dependent retraction highlighted by the theorem, even when valid histories are provided.

Please refer to “Common Response: Experimental Extensions and Evaluation Design” and “Common Response: Extended Experimental Results” for the revised experimental setup and results.


Issue 3. Statistical Uncertainty and Reliability

We agree that differences based on single greedy responses require cautious interpretation. During the rebuttal period, we prioritized consistent dataset construction, oracle annotations, and matched evaluation cohorts, and expanded the evaluation across datasets and prompting formats. We now report 95% confidence intervals from graph-level bootstrap resampling. These intervals quantify variation across graphs under fixed experimental settings, but do not capture stochastic decoding or training-seed variability. We therefore plan to add repeated temperature-based generation and extend SFT and GRPO evaluation to three independent training seeds, updating the manuscript accordingly. Until these checks are complete, our conclusions about procedure differences and trainability remain limited to the evaluated configurations.

Issue 4. External Validity Beyond Synthetic Graphs

We added matched natural-language Synthetic examples and adapted ProofWriter and BoardGameQA evaluations. The current Synthetic natural-language version uses one formulation family, with two additional families planned to assess sensitivity to wording. ProofWriter and BoardGameQA provide useful conflict and multiple-support structures, but only a limited subset satisfies the requirements of our retraction interventions, and controlled CF-path adaptations were still necessary. Related work such as STALE [1] examines whether agents recognize and act on memories invalidated by later observations, providing complementary evidence about dynamic belief maintenance. Our task additionally requires explicit support dependencies, specified interventions, and reference support histories. We will discuss the need for natural-language benchmarks combining these properties, alongside the limitations of our current adaptations.

[1] Chao et al. (2026). STALE: Can LLM Agents Know When Their Memories Are No Longer Valid?

Issue 5. Uneven Improvements from SFT

We agree that the training claim should distinguish improved belief updating from preservation of unaffected beliefs. Our intended claim is that SFT can teach retraction-related maintenance under specified semantics, rather than uniformly improve every metric. We investigate whether Direct supervision teaches models which information to revise or retain, and whether this capability transfers to Structured CoT and MultiTurn prompting. Separately, GRPO tests whether rewards for intermediate and final states improve execution of the maintenance procedure. Direct SFT generally improves update accuracy across the evaluated datasets and transfers across prompting formats, but preservation gains remain uneven, including the decrease in Edit On-CF maintenance under CoT. We will report these regressions explicitly and frame the contribution as evidence that retraction-related reasoning can be strengthened through training, with remaining limitations.

Issue 6. GRPO Training Duration and Budget Comparison

We extended GRPO to three epochs and evaluated each epoch checkpoint, alongside two-epoch SFT. In the completed Rebuild runs, mean training costs are approximately 1.2 GPU-hours for SFT and 25.2 GPU-hours for three-epoch GRPO. Although GRPO improves some metrics over Base, its gains are not consistent across epochs, and Off-CF update regressions remain. SFT achieves higher update accuracy across the reported prompting formats at substantially lower measured training cost. (Our claim is that SFT achieved better outcomes for retraction learning in the evaluated setting, rather than that it is inherently superior to GRPO; SFT may have limitations in longer-horizon reasoning beyond these maintenance tasks.)

Comments on the paper

5 comments
AuthorsOpenReview ↗

Common Response: Experimental Extensions and Evaluation Design

We thank the reviewers for their comments on the theoretical aspects of our work and their concrete suggestions for expanding and strengthening the experimental evaluation.

In this common response, we summarize how we expanded and refined our experiments to address the concerns raised across the reviews.


Dataset Extensions (by Reviewer CTPg and QSd5)

We replaced the original fixed six-layer Synthetic graphs (widths 3–5) with graphs spanning 5–13 layers and maximum layer widths of 5–9 in the evaluation set, added matched natural-language versions (e.g., “P1 is true” → “Alice has access to the server”), and constructed a separate history-matched dataset.

For ProofWriter (PW) and BoardGameQA (BGQA), we exhaustively screened the source corpora for valid retraction cases. Since most examples lacked suitable outgoing dependencies for propagating a proposition’s T/F change, we connected available propositions already present in the input with two additional rules to construct CF propagation paths, preserving the original facts, rules, priorities, and CF targets.

After fixed split allocation and input-length filtering, the train/test sizes are 100/500 for Synthetic, 100/944 for ProofWriter, and 84/291 for BoardGameQA; the symbolic and natural-language Synthetic versions share the same graphs.


Evaluation Formats (Reviewer CTPg)

We evaluate both procedures using single-response CoT and MultiTurn prompting. CoT produces all requested stages in one response, whereas MultiTurn uses six successive turns: withdrawal, forward maintenance, backward support removal, CF assignment, CF propagation, and final-query answering. Each turn requests a structured JSON response for the current stage, with previous model responses retained in the conversation and no gold-state correction between turns. Both formats use the same initial state, procedure, and query scope.

  • The original Direct and free-form CoT evaluations showed low accuracy on inputs containing many retraction and revision rules. We therefore adopted Structured CoT and MultiTurn formats to make the required maintenance stages explicit.

Extended Experiments: Research Questions and Methods

RQ1. How accurately do LLMs perform forward/backward retraction and preserve counterfactual values?

We evaluate forward and backward retraction using independent stage diagnostics with correct preceding states. We also measure whether the assigned CF value is preserved after subsequent inference.

RQ2. How accurately do LLMs update and preserve beliefs on and off CF-activated inference paths?

After counterfactual (CF, T->F or F->T) assignment and propagation, we evaluate final belief update and preservation separately for propositions reached by CF propagation (On-CF) and those outside that region (Off-CF), excluding the CF target. We assess both whether required value changes are correctly applied and whether beliefs that should remain unchanged are preserved.

RQ3. Can training improve retraction and transfer across prompting formats?

We jointly train on Synthetic, Synthetic Natural, ProofWriter, and BoardGameQA using Direct JSON supervision, then evaluate SFT transfer to CoT prompting. Independently, we train a separate model from the same base checkpoint using Structured CoT GRPO and evaluate both CoT performance and transfer to MultiTurn prompting. For each training approach, we report belief-maintenance performance.

Additional analyses in reviewer-specific replies

  1. Performance under single-response CoT versus MultiTurn prompting.
  2. Backward Edit: ablating forward retraction.
  3. Detailed history-matched results and support-history input ablations.
  4. Training budgets, computational costs, and limitations of the SFT–GRPO comparison.
AuthorsOpenReview ↗

Common Response: Extended Experimental Results

Results cover Qwen3-4B, Qwen3-14B, Gemma-3-4B, and Gemma-3-12B. Table entries report accuracy for predicting T/F/U values, with 95% confidence intervals in square brackets. We compute these intervals from 10,000 graph-level bootstrap replicates, retaining all observations from each sampled graph and recomputing the mean across the fixed model, prompting, and cut settings in each replicate.

Key Findings

  1. Edit generally shows higher accuracy on backward retraction and shared-gold Off-CF updates.

  2. Rebuild shows higher or comparable On-CF update accuracy and higher Off-CF maintenance accuracy.

  3. Across the reported comparisons, models are consistently more accurate at maintaining unchanged belief values than at applying required value changes.


RQ1. How accurately do LLMs perform forward/backward retraction and preserve counterfactual values?


1.1. Counterfactual Value Preservation

DatasetEditRebuild
Syn0.60 [0.60, 0.61]0.62 [0.61, 0.62]
Syn Natural0.70 [0.70, 0.70]0.68 [0.68, 0.69]
PW0.62 [0.61, 0.62]0.65 [0.64, 0.65]
BGQA0.61 [0.60, 0.62]0.64 [0.63, 0.65]

1.2. Forward Retraction

DatasetEdit: UpdateRebuild: UpdateEdit: MaintenanceRebuild: Maintenance
Syn0.02 [0.02, 0.03]*0.67 [0.66, 0.68]0.57 [0.56, 0.58]
Syn Natural0.02 [0.01, 0.02]*0.49 [0.48, 0.50]0.41 [0.40, 0.41]
PW0.08 [0.07, 0.09]*0.67 [0.66, 0.68]0.47 [0.47, 0.48]
BGQA0.04 [0.03, 0.06]*0.54 [0.53, 0.56]0.38 [0.37, 0.40]

* Under the local Cached Rebuild policy, there are no eligible value-changing queries for forward retraction.

1.3. Backward Retraction

DatasetEdit: UpdateRebuild: UpdateEdit: MaintenanceRebuild: Maintenance
Syn0.10 [0.09, 0.10]0.08 [0.08, 0.09]0.45 [0.45, 0.46]0.40 [0.40, 0.41]
Syn Natural0.08 [0.08, 0.09]0.05 [0.04, 0.05]0.41 [0.40, 0.42]0.27 [0.27, 0.28]
PW0.12 [0.11, 0.12]0.12 [0.11, 0.13]0.47 [0.46, 0.48]0.36 [0.36, 0.37]
BGQA0.15 [0.14, 0.16]0.13 [0.12, 0.14]0.47 [0.46, 0.49]0.30 [0.29, 0.31]

RQ2. How accurately do LLMs update and preserve beliefs on and off CF-activated inference paths?


2.1. On-CF: Belief Update and Maintenance

DatasetEdit: UpdateRebuild: UpdateEdit: MaintenanceRebuild: Maintenance
Syn0.17 [0.16, 0.18]0.19 [0.18, 0.20]0.78 [0.72, 0.88]0.81 [0.75, 0.94]
Syn Natural0.02 [0.02, 0.02]0.02 [0.02, 0.03]0.71 [0.62, 0.75]0.71 [0.62, 0.81]
PW0.25 [0.24, 0.26]0.29 [0.28, 0.29]Not reported (<1%)Not reported (<1%)
BGQA0.27 [0.25, 0.28]0.29 [0.28, 0.31]Not reported (<1%)Not reported (<1%)

2.2. Off-CF: Comparison on Queries with Identical Gold Answers

DatasetEdit: UpdateRebuild: UpdateEdit: MaintenanceRebuild: Maintenance
Syn0.15 [0.14, 0.15]0.12 [0.11, 0.13]0.80 [0.80, 0.81]0.83 [0.82, 0.83]
Syn Natural0.13 [0.12, 0.14]0.11 [0.10, 0.12]0.80 [0.79, 0.80]0.84 [0.84, 0.85]
PW0.13 [0.12, 0.13]0.11 [0.10, 0.12]0.83 [0.83, 0.84]0.87 [0.87, 0.87]
BGQA0.09 [0.08, 0.10]0.07 [0.06, 0.08]0.86 [0.85, 0.86]0.87 [0.86, 0.88]

2.3. Retraction-Affected Updates under Edit

DatasetEdit
Syn0.10 [0.09, 0.10]
Syn Natural0.07 [0.06, 0.07]
PW0.21 [0.20, 0.23]
BGQA0.14 [0.12, 0.17]
AuthorsOpenReview ↗

Common Response: Extended Experimental Results (Continued)

Due to computational constraints, we currently report Rebuild results only for Qwen3-4B and Qwen3-14B.

RQ3. Can SFT and GRPO improve retraction and transfer across prompting formats?

3.1. Rebuild

3.1.1. Counterfactual Value Preservation and On-CF Performance

TrainingCF Preservation (CoT)CF Preservation (MultiTurn)On-CF BU (CoT)On-CF BU (MultiTurn)On-CF BM (CoT)On-CF BM (MultiTurn)
CoTMultiTurnCoTMultiTurnCoTMultiTurn
Original0.98 [0.98, 0.98]0.99 [0.99, 0.99]0.41 [0.40, 0.42]0.31 [0.30, 0.32]0.72 [0.62, 0.88]0.81 [0.69, 1.00]
SFT0.99 [0.99, 0.99]1.00 [1.00, 1.00]0.83 [0.82, 0.84]0.67 [0.67, 0.68]0.91 [0.88, 1.00]1.00 [1.00, 1.00]
GRPO epoch 10.98 [0.98, 0.99]0.99 [0.99, 0.99]0.47 [0.46, 0.48]0.34 [0.33, 0.35]0.88 [0.62, 1.00]0.75 [0.69, 0.88]
GRPO epoch 20.99 [0.99, 0.99]1.00 [0.99, 1.00]0.55 [0.54, 0.56]0.40 [0.39, 0.41]0.72 [0.62, 0.88]0.91 [0.81, 1.00]
GRPO epoch 30.99 [0.99, 0.99]0.99 [0.99, 1.00]0.55 [0.54, 0.56]0.40 [0.39, 0.41]0.75 [0.62, 1.00]0.75 [0.62, 0.88]

3.1.2. Off-CF Performance

TrainingOff-CF BU (CoT)Off-CF BU (MultiTurn)Off-CF BM (CoT)Off-CF BM (MultiTurn)
CoTMultiTurnCoTMultiTurn
Original0.12 [0.12, 0.13]0.18 [0.17, 0.19]0.93 [0.92, 0.93]0.93 [0.92, 0.93]
SFT0.70 [0.69, 0.71]0.68 [0.67, 0.70]0.98 [0.98, 0.98]0.97 [0.97, 0.98]
GRPO epoch 10.11 [0.11, 0.12]0.17 [0.16, 0.18]0.94 [0.94, 0.94]0.93 [0.93, 0.94]
GRPO epoch 20.10 [0.09, 0.10]0.16 [0.15, 0.17]0.95 [0.95, 0.96]0.94 [0.94, 0.94]
GRPO epoch 30.10 [0.10, 0.11]0.16 [0.16, 0.17]0.95 [0.94, 0.95]0.94 [0.93, 0.94]

3.2. Edit

3.2.1. Counterfactual Value Preservation and On-CF Performance

TrainingCF Preservation (CoT)CF Preservation (MultiTurn)On-CF BU (CoT)On-CF BU (MultiTurn)On-CF BM (CoT)On-CF BM (MultiTurn)
Original0.98 [0.98, 0.98]0.99 [0.99, 0.99]0.34 [0.33, 0.35]0.28 [0.27, 0.29]0.84 [0.75, 1.00]0.88 [0.81, 1.00]
SFT0.99 [0.99, 1.00]1.00 [0.99, 1.00]0.84 [0.83, 0.84]0.76 [0.76, 0.77]0.78 [0.62, 0.88]0.81 [0.75, 0.88]
GRPO epoch 10.99 [0.99, 0.99]0.99 [0.99, 1.00]0.43 [0.42, 0.44]0.32 [0.32, 0.33]0.84 [0.75, 0.88]0.78 [0.62, 1.00]
GRPO epoch 20.99 [0.99, 0.99]0.99 [0.99, 1.00]0.54 [0.53, 0.55]0.41 [0.40, 0.42]0.81 [0.75, 1.00]0.84 [0.81, 0.88]
GRPO epoch 30.99 [0.99, 0.99]0.99 [0.99, 1.00]0.55 [0.54, 0.56]0.41 [0.40, 0.42]0.69 [0.50, 0.75]0.84 [0.75, 1.00]

3.2.2. Off-CF Performance

TrainingOff-CF BU (CoT)Off-CF BU (MultiTurn)Off-CF BM (CoT)Off-CF BM (MultiTurn)
Original0.12 [0.11, 0.13]0.19 [0.18, 0.20]0.96 [0.96, 0.96]0.96 [0.96, 0.96]
SFT0.55 [0.53, 0.56]0.57 [0.56, 0.58]0.97 [0.97, 0.97]0.97 [0.97, 0.97]
GRPO epoch 10.13 [0.12, 0.14]0.20 [0.19, 0.21]0.96 [0.96, 0.96]0.96 [0.96, 0.96]
GRPO epoch 20.14 [0.13, 0.15]0.22 [0.21, 0.23]0.95 [0.95, 0.96]0.95 [0.95, 0.96]
GRPO epoch 30.15 [0.14, 0.16]0.22 [0.21, 0.23]0.95 [0.95, 0.96]0.95 [0.95, 0.95]

Key Findings

  1. Direct SFT generally improves belief-update and maintenance accuracy, with gains transferring to CoT and MultiTurn prompting.

  2. GRPO produces some improvements with longer training, but gains remain limited and are not consistent across epochs and metrics.

  3. When exact symbolic targets are available, Direct SFT provides a practical and computationally cheaper approach in our evaluated configurations.

  4. On-CF BM accuracy remains high under Rebuild, whereas under Edit, both SFT and GRPO yield accuracy at or below the baseline.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.