acceptodds
Withdrawn from ICLR 2027

ThesisAgent: Revising Persistent Theses for Long-Horizon Decision Making

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Long-horizon agents must revise decision-supporting assumptions while preserving unaffected beliefs. This requires controlling revision scope and separating belief changes from execution constraints. Behavioral responsiveness must also be distinguished from decision quality. We propose ThesisAgent, which maintains these assumptions as a persistent decision thesis governed by an explicit transition contract. To address Unscoped Thesis Revision, Evidence-Local Revision separates hypothesis selection from revision strength. Its numerical updater changes only a bounded subset and exactly preserves unselected confidences. To address Belief–Execution Entanglement, Decision–Execution Separation exposes desired actions before feasibility enforcement so that belief effects can be audited at fixed execution context. To address Sensitivity–Utility Ambiguity, Intervention-Based Utility Audit combines paired hypothesis interventions with oracle-referenced regret in the controlled setting to assess decision dependence and quality separately. ThesisWorld experiments show that local updates improve utility over matched dense updates and reveal a trade-off between rapid response and retained correction. In ALFWorld, persistent thesis state improves task success by 6.0 percentage points over a matched control using the same execution controller. In the 43-month return-based financial proxy, ThesisAgent exceeds the strongest evaluated baseline by 23.7 percentage points in cumulative net return. Our code is available at https://anonymous.4open.science/r/THESISAGENT-948D.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Long-horizon agents, persistent decision thesis, belief-action consistency, evidence-grounded belief revision, financial agents

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

The paper proposes ThesisAgent, which maintains a persistent set of assumptions about an ongoing task and revises them as new evidence arrives. It includes three mechanism to achieve persistent decision theses: Evidence-Local Revision, Decision–Execution Separation, and Intervention-Based Utility Audit. The paper evaluates these designs in ThesisWorld, ALFWorld, and a financial proxy, reporting improvements in belief and decision metrics and task success.

Critical strengths

In long-horizon tasks, new observations can invalidate some earlier assumptions while leaving others intact. How an agent maintains and revises its beliefs as conditions change is an important question.

Critical weaknesses

W1. The motivation and terminology are difficult to follow. Here are some examples:

  • The introduction presents three newly named challenges, but Figure 1 mainly contrasts step-wise rationale regeneration with persistent selective thesis revision; it does not illustrate the execution or decision-quality issues.
  • In the discussion of Unscoped Thesis Revision, the text moves from information omissions and incorrect updates (line 49), to prior memory-editing methods (lines 50–51), and then to evidence relevance and update magnitude (lines 52–70), without clearly distinguishing these failure modes or explaining how the cited methods motivate the specific scope problem.
  • Terms such as hypothesis, belief, thesis, desired decision, and executed action appear in the motivation before their relationships are made clear. A single worked example connecting each challenge to its mechanism and evaluation would improve readability.

W2. The ALFWorld baseline results require stronger validation. In Table 2, all six adapted baselines achieve only 17–25 successes out of 134 tasks (12.7–18.7%), while the paper’s own no-persistent-thesis variant achieves 103/134 (76.9%). The gap is particularly striking for ReAct, which scores 17/134 (12.7%) here. On 134 unseen ALFWorld tasks, the original ReAct paper [1] reports 57% averaged over six prompts, 71% with its best prompt, and 78.4% using GPT-3 in the appendix. A later study[2] reports 75.4% for ReAct using gpt-4.1-mini, 134 tasks, and a 50-step budget. Reflexion also scores only 17/134 in Table 2, whereas an independent reimplementation[3] reports 59.8% with GPT-3.5 Turbo and 85.9% with GPT-4 Turbo. These cross-paper numbers are not controlled comparisons: models, prompts, action interfaces, and—in Reflexion’s case—retry protocols differ. These unusually low baselines need explanation. The authors should disclose their prompts and protocols and rerun ReAct under the same setup. Until then, Table 2 cannot establish the size of ThesisAgent’s gain over prior methods.

[1]https://arxiv.org/html/2210.03629

[2]https://arxiv.org/html/2606.02304

[3]https://proceedings.neurips.cc/paper_files/paper/2024/file/0142921fad7ef9192bd87229cdafa9d4-Paper-Conference.pdf

Questions for authors

As shown in Weakness.

3 comments
AuthorsOpenReview ↗

Response to W1 and W2: Clarifications on Motivation, ALFWorld Baselines, and Evaluation Scores

Response to Reviewer wqnp

We thank the reviewer for the feedback and provide the following clarifications and new experimental evidence.

W1. Motivation and Figure 1

We respectfully disagree that Figure 1 is inadequate because it does not illustrate all three challenges.

The Introduction explicitly states: "Figure 1 illustrates how maintaining this thesis differs from regenerating a rationale at each step."

Figure 1 specifically motivates Unscoped Thesis Revision, not all three challenges. We never claim otherwise. The remaining challenges, Belief–Execution Entanglement and Sensitivity–Utility Ambiguity, are separately explained in the Introduction and illustrated through their corresponding mechanisms in Figure 2.

A focused motivation figure should not be considered inconsistent merely because it does not illustrate every technical challenge. We therefore disagree with this particular criticism.

Regarding terminology, a Thesis contains persistent, identifiable hypotheses. Desired Action precedes feasibility enforcement, while Executed Action follows it. Unscoped Thesis Revision distinguishes which hypotheses may change from how much they change. We agree that these definitions and their connections to prior work can be introduced more clearly and will revise the presentation accordingly.

W2. ALFWorld Baseline Reliability

First, the 1,800-character reasoning-retention limit was applied to ALL methods, not specifically to ReAct. Thus, the original 12.7% result reflects ReAct under a common restricted-context protocol, rather than its full-history implementation. Cross-paper success rates under different models, prompts, and execution protocols cannot alone establish unfair treatment.

Nevertheless, we acknowledge that common restrictions do not guarantee faithful reproduction of every original method.

To directly address this concern, we completed the ReAct implementation, relaxed history restrictions, and reran all 134 tasks.

ConfigurationSuccessLLM CallsTokens
Restricted-context ReAct17/134 (12.7%)6,1014.65M
Extended-history ReAct56/134 (41.8%)4,66211.00M

Success improved by 29.1 percentage points, while token consumption increased 2.37×. Since implementation details were also completed, we do not attribute the entire gain solely to history retention.

We acknowledge that 41.8% remains below some published results and will clarify the prompting and execution protocols. The two configurations should not be conflated.

Crucially, ThesisAgent's contribution does not depend solely on ReAct's performance. Our subsequent mechanism-matched ALFWorld evaluation, using the same PhaseGuard controller, yields:

  • ThesisAgent + PhaseGuard: 132/134 (98.5%)
  • No-persistent-thesis + PhaseGuard: 124/134 (92.5%)

This +6.0 percentage-point gain directly supports the incremental contribution of thesis persistence. We acknowledge that this evaluation is post-hoc, not an untouched test. Independent ThesisWorld experiments also evaluate the revision mechanism without relying on ReAct.

We therefore distinguish baseline reproduction fidelity, comparisons under common restrictions, and the independent contribution of persistent thesis state. Limitations in one comparison do not automatically invalidate separate controlled evidence.

Clarification on the Lowest Scores

We respectfully request clarification regarding the minimum scores assigned to all three dimensions.

Presentation (1): Figure 1 fulfills its explicitly stated purpose. What additional fundamental presentation issues justify the lowest score?

Informativeness (1): The paper introduces a persistent thesis-transition contract, scoped revision, and intervention-based utility auditing. Which prior work already establishes these contributions sufficiently to justify the lowest informativeness score?

Soundness (1): Beyond the ALFWorld baseline concerns, are there fundamental issues with the independent ThesisWorld experiments, analytical results, or matched ablations?

We acknowledge genuine limitations but believe that concerns about specific comparisons should not automatically be extended to all methodological contributions. We respectfully ask the reviewer to reconsider the scores and overall recommendation in light of the new evidence.

Reviewer_wqnpOpenReview ↗

Response to the authors

Dear Authors,

Thank you for clarifying Figure 1. I accept that it illustrates only the first challenge; that is a minor point. My broader presentation concern remains. Lines 49–70 are one example of a difficult argument, and the paper does not consistently distinguish thesis, belief, hypothesis, desired action, and executed action. Your rebuttal helps, but these connections need to be clearer in the manuscript before I can raise my Presentation score.

The ReAct rerun strengthens my concern about Table 2. Its success rises from 17/134 to 56/134 after you both complete the implementation and relax the history restriction. Because both changed, the rebuttal does not explain the original result. The manuscript does not disclose the 1,800-character limit. Appendix A.1 describes different recent-history windows but does not specify what each literature baseline receives. A common limit also does not ensure comparable access to information when ThesisAgent retains a thesis and PhaseGuard retains a separate ledger. What does “Full-History” mean under this limit?

The within-paper contrast is striking: all six adapted literature baselines solve 17–25 tasks, while your no-persistent-thesis variant without PhaseGuard solves 103. The matched 124-to-132 result supports thesis persistence, but it does not explain this much larger gap, and the evaluation includes development-touched tasks. Published ReAct scores are also substantially higher, though their protocols differ. Please provide the baseline prompts, exact history rules, paired task outcomes, and separate tests of the implementation and history changes. Until then, Table 2 does not reliably establish the size of the improvement over prior agents.

AuthorsOpenReview ↗

### Final Response to Reviewer wqnp

Dear Reviewer,

We sincerely thank the reviewer for the thoughtful follow-up and detailed clarification.  Your latest response has helped us better understand the limitations of our current manuscript and, more importantly, provided concrete guidance on how to improve our work. 

After carefully considering your feedback, we recognize that the manuscript requires further refinement, particularly in terms of conceptual clarity, baseline implementation transparency, and experimental validation. 

We have therefore decided to withdraw our submission from ICLR 2027 to undertake a more thorough revision. We will clarify the relationships among the core concepts, systematically re-examine the experimental protocols and baseline implementations, and conduct additional controlled experiments to ensure that our conclusions are adequately supported by the evidence. 

We sincerely appreciate the time and effort you have devoted to reviewing our work. Your detailed comments and follow-up discussion have provided valuable direction for improving the rigor, clarity, and reproducibility of our research. 

Thank you again for your careful evaluation and constructive feedback.

Reviewer_K3fUOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

ThesisAgent organizes decision-supporting assumptions as persistent hypotheses with explicit identities and revision rules. Its numerical updater restricts ordinary revisions to a bounded relevance-selected subset; desired actions are separated from execution constraints; paired interventions distinguish behavioral dependence from oracle-referenced decision quality. Evidence comes from a synthetic numerical ThesisWorld, prompted ALFWorld adaptations, and a 43-month return-based financial proxy. The controlled experiments support local revision and reference-dependent residual filtering. ALFWorld persistence gains are measured with and without the task-specific PhaseGuard controller. Financial gains concern tuned numerical signal adapters, rather than native financial LLM agents or textual fundamental reasoning.

Critical strengths
  1. The distinction between dependence, consistency, and decision quality is valuable. Desired-action interventions fix context, controller state, and randomness, while the numerical utility audit uses an oracle distinct from the belief-based transition reference. The reference-error analysis and matched controls explicitly show when residual denoising can fail and why lower residual error alone does not imply better decisions. This helps avoid overinterpreting intervention sensitivity.
  2. Several controls improve interpretability: dense revision changes update eligibility under shared noise, ALFWorld thesis removal retains the same recent history and PhaseGuard ledger, and the finance section includes a common portfolio projection and same-evidence reset/EMA comparisons. The paper also discloses implementation equivalence, post-hoc refinements, and limitations of external-agent adaptations. These are useful diagnostics and make the narrow numerical findings easier to assess.
Critical weaknesses
  1. The main controlled results do not yet validate the difficult semantic part of the proposed agent framework. ThesisWorld supplies numerical prototypes, reliability/duplicate annotations, a fixed action map, and an exact transition reference; its baselines are hand-specified numerical adapters. Appendix G reproduces the complete state/action trajectories with an ordinary array implementation and identifies consistency as residual filtering. These facts do not make the experiment incorrect, but they leave limited evidence for a distinct general agent mechanism beyond sparse accumulation and known filtering. ALFWorld uses prompted textual revision without the numerical hard gate or consistency rule, so it does not directly validate those central guarantees in a semantic setting.
  2. The domain results lack a clean independent validation of the resulting system. Appendix C states that the post-hoc ALFWorld census includes valid-unseen tasks touched during development. The finance result is a single inspected 43-month path, follows substantially greater ThesisAgent parameter-search effort than baseline development, and is exactly reproduced by a numerical accumulator. These limitations are disclosed, but disclosure does not resolve the generalization question. Together they make the evidence insufficient for the breadth of the claimed long-horizon-agent contribution, motivating weak rejection pending stronger independent validation.
Questions for authors
  1. Which ALFWorld tasks were never inspected during development? Can paired thesis-persistence gains, with uncertainty and intervention counts, be reported separately on that untouched subset using the frozen controller and prompts?
  2. What existing experiment tests evidence attribution/local revision with realistically imperfect semantic interpretation while retaining the same update and decision mechanism? A targeted comparison against a simple sparse accumulator/filter under the same interface would help identify what the thesis representation adds beyond the numerical rules.
  3. Can the finance conclusion be supported on an untouched period or market with matched development budgets, and against a tuned conventional accumulator/filter family? Otherwise, would the authors narrow this part to an exploratory illustration rather than evidence of general financial-agent superiority?
Reviewer_TX1UOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The problem the paper address relates to ho LLMs handle new information. Currently, LLMs typically process the new request from scratch, while there may be many previous observations that maybe relevant. This sometimes cause major decision changes. Authors propose ThesisAgent, which keeps the assumptions of the previous state, and as soon as a new information arrives, it only modifies the assumptions that are related to that new information, estimating how much previous assumptions should change. Additionally, the approach differentiates between what the agent would do and what it is allowed to do, hence checking whether a change in belief propagates to the decision.

Critical strengths
  1. The problem is well motivated. Separating "which beliefs may change" from "how much they change," and "did the action depend on the belief" from "was the change good," is a genuinely helpful perspective. The audit design follows good causal-probe concepts: paired copies, fixed context, reused randomness, no updater rerun.

  2. The residual-filter analysis is correct and matches the data

Critical weaknesses

1.The paper links to an anonymous repository, but it is currently empty, so the implementation cannot be checked. If the code is to be released later, this should be stated explicitly.

  1. Persistent memory and belief revision of related works needs a better writing in my opinion. After introducing the works, it is needed to say what is wrong with them, and then say how your approach is improving. The same happens with the third part, Financial agents and evaluation.

  2. Section 3.2, When more than M hypotheses pass the threshold, the ones ranked below M are simply dropped. That evidence is ignored for them, with no deferral or partial update. The paper never discusses this false-negative side. Its analysis (Appendix F) only covers false positives, meaning unrelated hypotheses wrongly admitted. In ThesisWorld the problem is essentially hidden by design. M = 1, and the simulator generates each observation to target exactly one hypothesis, with orthogonal prototypes. So "more than M relevant hypotheses" almost never happens except through noise. Additionally, the authors are introducing two important hypeparameters that they did not study their effect. The values are fixed in Appendix A: γr = 0.45 and M = 1 for ThesisWorld, with no justification for 0.45.

  3. Under-specified baselines. Several baselines are adaptations of existing systems, but the paper gives little detail on how they were built. The ThesisWorld adapters get one sentence of coefficients each, with no justification for the values. The financial agents are described in a sentence each, mostly without formulas. The ALFWorld baselines are called "prompt adaptations," but no prompts are given. This matters most for ALFWorld, where the baselines' 12.7–18.7% success is unusually low for gpt-4.1-mini. No tuning protocol is described for the baselines, while ThesisAgent was tuned through validation sweeps. Some adapter choices, such as cross-claim coupling, also directly inflate the unaffected-belief metric on which ThesisAgent wins. The authors should provide full baseline specifications and explain how each adapter preserves the original method's core mechanism, or else stop using the original systems' names.

Questions for authors
  1. How were γr = 0.45 and M = 1 chosen? Can you report a sensitivity analysis that varies each one separately, rather than removing both at once as the dense-revision ablation does?

  2. When more than M hypotheses pass the threshold, the lower-ranked ones are not updated at all. How often does this happen, and how does the gate behave when an observation is genuinely relevant to several hypotheses?

  3. Can you provide the prompts used for the ALFWorld baselines, and explain why they reach only 12.7–18.7% success, well below what these methods usually achieve with comparable models?

  4. How were the baseline adapters' parameters chosen, and were they tuned with effort comparable to ThesisAgent? For each adapter, what core mechanism of the original method does it preserve?

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

Related papers

Discussion (0)

Sign in to comment.