acceptodds
Under review as a conference paper at ICLR 2027

A popperian neuro-symbolic agent that learns through LLM-based conjecture and refutation

Abstract

Recent advancements in Large Language Models (LLMs) have enabled new manners of integration with decision systems and autonomous agents, although they offer no guarantees for deductive reasoning nor plan generation. However, their capacity to provide probable, plausible answers to given inputs makes them suitable to be used as abductive reasoners, and brings the field to the era of popperian agents that weed out invalid possibilities in order to quicken learning. In this paper, we present a popperian agent architecture that uses an LLM-based component for epistemic queries in order to learn the dynamics of its environment: the LLM conjectures hypotheses, which are validated through thorough experimentation, and ultimately form a planning model that can be reused at near-zero cost, without the intervention of the LLM. We also present a provenance ontology that allows connecting each learnt piece of information with the experiments that validated it and the phenomenon that caused the piece of information to be conjectured at all.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
Artificial Intelligence, autonomous agents, neuro-symbolic agents, Large Language Models, abductive reasoning, planning, scientific discovery

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_tkiaOpenReview ↗
Rating
1
Confidence
4
Informativeness
3
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes HYPER, a neuro-symbolic agent architecture that uses large language models to generate hypotheses about unknown environment dynamics and validates them through explicit experiments before incorporating them into a symbolic planning model.

The framework separates hypothesis generation from decision-making. When the agent encounters an unexpected observation or cannot construct a plan, an LLM proposes candidate explanations and designs experiments to test them. Hypotheses that survive the validation process are incorporated into a PDDL-based action model, which is subsequently used by a classical planner. The authors also introduce a provenance ontology based on W3C-PROV to record the relationships among observations, hypotheses, experiments, and learned beliefs.

The approach is evaluated in four custom grid-world environments using three language models. The experiments compare the complete HYPER architecture against a variant without refutation and a direct LLM-based planner, reporting task success, output token consumption, and environment actions.

Critical strengths
  1. Explicit separation between hypothesis generation and validated planning knowledge.

The paper explores a meaningful architectural principle: LLM-generated explanations should be treated as fallible hypotheses rather than immediately accepted as reliable environment knowledge. The explicit conjecture–experiment–refutation loop provides an interpretable approach to learning symbolic action models.

  1. Potential for reusable and traceable environment knowledge.

The integration of classical planning with a provenance ontology is practically interesting. Recording which observations and experiments support each learned belief could facilitate debugging and model revision. Once a sufficiently accurate action model has been constructed, reusing it without repeated LLM inference may also offer computational benefits. However, these potential advantages require stronger empirical validation than currently provided.

Critical weaknesses
  1. Inadequate experimental validation of the central world-model learning and reuse claims.

My primary concern is that the experimental evaluation is too limited to support the paper's main scientific claims.

First, the experiments involve only four custom grid-world environments, three of which are closely related variations of the fire-crossing task. Each model–environment configuration is evaluated using only three or five runs. Such a small and narrowly constructed evaluation provides insufficient evidence for general conclusions about learning complex environment dynamics.

Second, Table 1 explicitly marks several experimental results as invalid due to an external error, with the authors stating that these results will be updated after the review process. Although the disclosure is appreciated, the submission must be evaluated based on the currently available valid evidence.

Third, the paper claims that HYPER learns reusable symbolic world models through hypothesis generation and refutation. However, the main experiments measure task completion rather than the correctness or completeness of the learned action models. Successfully completing a task does not establish that the agent has correctly identified the underlying environment dynamics. An agent may learn only a small subset of the required rules or rely on partially incorrect models that happen to support the evaluated task.

The evaluation does not systematically examine transition prediction accuracy, learned-rule coverage, generalization to unseen initial states or goals, or the ability to reuse learned dynamics across different tasks. Consequently, the central claim that the proposed refutation process produces reliable and reusable world knowledge remains insufficiently supported.

Finally, the computational efficiency analysis introduces an additional concern. Table 1 explicitly states that token consumption and environment actions are reported only for successful runs. This excludes potentially substantial costs incurred by unsuccessful learning attempts, particularly when failures result from reaching the process limit. The claimed amortization of learning costs over repeated deployments is therefore not convincingly established by the reported measurements.

These limitations concern the paper's primary contributions rather than minor implementation details. Substantially broader and more direct validation would be required to establish the effectiveness of the proposed learning procedure.

  1. Limited methodological novelty and insufficient comparisons with existing interactive world-model learning approaches.

The proposed architecture combines several established ideas: LLM-based hypothesis generation, environment interaction, experimental validation, symbolic action-model construction, classical planning, and provenance tracking.

Although the conjecture–refutation formulation provides a coherent conceptual organization, the paper does not clearly establish what new generalizable learning mechanism distinguishes HYPER from existing approaches to interactive symbolic world-model learning, hypothesis-driven exploration, and verification-guided agent learning.

The experimental comparison is particularly limited. The authors primarily compare HYPER against direct LLM planning and an internal variant that accepts hypotheses without refutation. These baselines do not adequately represent the broader literature on interactive action-model acquisition or verification-guided world-model construction.

For example, the paper discusses prior work on learning action models through environment interactions, yet does not include a competitive model-learning approach under matched interaction and computation budgets. It is therefore difficult to determine whether HYPER provides advantages

Questions for authors

Q1. How do the authors establish that HYPER learns correct and reusable environment dynamics?

The current evaluation primarily reports task success, which does not directly measure the correctness of learned action models. Can the authors provide quantitative evaluations of transition prediction accuracy, learned-rule coverage, or performance on unseen initial states and goals using the learned PDDL model without further LLM-based learning?

Such evidence is essential to distinguish genuine environment-model acquisition from learning a limited set of rules sufficient for the evaluated tasks.

Q2. Can the authors provide a complete and unbiased evaluation of computational efficiency?

Table 1 reports token consumption and environment actions only for successful runs, while several configurations have substantial failure rates. Can the authors provide learning costs across all runs, including unsuccessful and timeout cases, and clarify how these costs affect the claimed amortization over repeated deployments?

Additionally, can the authors provide empirical repeated-deployment results rather than relying solely on estimated break-even points?

Q3. What is the independent advantage of the proposed refutation procedure over existing interactive model-learning approaches?

The current baselines are limited to direct LLM planning and a no-refutation variant. Can the authors provide evidence that HYPER outperforms an existing interactive symbolic model-learning or verification-guided learning approach under comparable environment interaction and computation budgets?

Furthermore, can the authors explain why refutation improves performance in some environments but provides no benefit or leads to lower success in others?

Addressing these questions would be necessary to substantiate the central methodological and empirical claims.

1 comment

We thank the reviewer for their questions and extensive comments. We will proceed to address them, one by one.

Q1

We would like to take the opportunity to clarify that the objective of HYPER, as well as similar agents, is not to produce a complete-coverage model of the environment it exists in, but instead to learn how to perform a task. It could be fine-tuned in order to make that its objective, and it could be an interesting line to explore, but it is not its current objective.

As there is no guarantee of the existence of a unique minimal set of dynamics required to learn a task (e.g. only one of turn_right or turn_left is technically required to solve all MiniGrid tasks, we belief that finding a theoretical measure of task dynamic coverage is quite complicated. HYPER will learn, at least, enough dynamics to cover a plan from the initial state to the goal, with no guarantees of generality. An empirical approach to gauge how general or robust the learnt model is, is by counting the amount of novelties encountered in unseen variations of the learned task. Further studies regarding generality could involve fusing various learned models, but it is out of scope for the paper.

Q2

After HYPER learned a task, we observed that further re-runs on tasks that did not require any additional environment dynamic only required a single goal-translation process, using approximately 100 LLM output tokens. We find it appropriate to discuss the amortisation costs of learning a task, based on deployment runs costing 2-3 orders of magnitude less than task-learning processes, being essentially negligible.

Q3

We believe that there is an inherent variability in LLM outputs that was a factor in no-refutation to accomplish a task one more time that its full counterpart in one of the model and environment configurations. This could have been mitigated by a larger study on the agent's variability. Apart from that, the no-refutation variation of HYPER will encounter less issues the more aligned that the environment dynamics are with the LLM's prior (the more aligned, the least likely it is for no-refutation to accept an incorrect belief).

Reviewer_4q3ZOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes HYPER, a neuro-symbolic agent that uses LLMs to generate hypotheses about environment dynamics, designs experiments to falsify them, and integrates validated knowledge into a PDDL world model for classical planning. Experiments in four grid-world environments suggest improvements over direct LLM planning and a variant without hypothesis refutation.

Critical strengths
  • The separation of LLM-based hypothesis generation from symbolic planning is a reasonable design choice.

  • The explicit provenance mechanism for tracing learned beliefs to their supporting experiments is potentially useful.

Critical weaknesses
  • Limited novelty. The core idea of generating world models with LLMs and revising them through environmental feedback has been explored in prior work, such as WorldCoder (Tang et al., 2024), the authors mentioned WorldCoder but did not establish any discussion in their differences?

  • Weak baselines and baseline models. The no-refutation baseline directly accepts every hypothesis, making it unsurprising that validation improves performance. The direct LLM planning baseline is also relatively simplistic compared with modern interactive coding or agentic approaches. Moreover, the evaluation relies partly on weak LLMs (e.g., open-weights models or GPT-luna), rather than comparing against stronger agents/harnesses built on frontier models. These choices limit the significance of the reported improvements.

  • Timeout-dominated failures. The authors report that 100% of GPT failures, 97% of Kimi failures, and 95% of Gemma failures are caused by timeouts. Furthermore, HYPER is allowed 200 processes, compared with 100 planning processes for the direct LLM baseline, without a matched computational budget. It is therefore unclear whether the differences reflect genuine architectural advantages or arbitrary resource limits.

  • Limited empirical evaluation. The experiments involve only four hand-designed, deterministic, fully observable grid worlds (and most of them Fire crossing variants), with just 3–5 runs per condition. Even in this restricted setting, performance is inconsistent, and several results are marked invalid. There is also no systematic evaluation of generalization to new tasks, initial states, or environments.

  • Insufficient justification of the falsification mechanism. The same LLM generates hypotheses and experiments, potentially introducing confirmation bias. Experiments are randomly selected rather than optimized for informativeness. Comparisons against random exploration, information-gain-based experiment selection, or simpler transition-based model learning are missing.

Questions for authors
  1. How novel is this relative to existing model-based LLM agents?
  2. Is the method genuinely learning from scratch? The agent starts with an empty action model, but it receives a structured representation of states and available actions. The hypothesis space is consequently much more constrained than that of an agent discovering useful state variables, object categories, and actions from raw observations.
2 comments

We thank the reviewer for their questions and extensive comments. We will proceed to address them, one by one.

Q1

HYPER differentiates itself from agents such as WorldCoder by presenting an explicit conjecture and refutation mechanism that explicitly validates the dynamics it hypothesises. It also maintains a trace that serves as epistemic support of everything it learns.

Q2

It could be considered that any agent architecture, even the most basic ones, have some notion of prior knowledge arising from its internal architecture (even if it were a neural network, it is parametrised in some meaningful way) and defined interfaces for perception and action. The claim that the agent learns from scratch can be relaxed to learning the world dynamics and objectives from scratch, conditioned on prior LLM-knowledge associating natural language terms in the input.

We acknowledge, however, that less processed inputs (eg. image data) could be desirable in some circumstances, but this remains future work. We also thank the reviewer for the suggestion to consider letting the agent discover useful properties that help categorise and reason about unequal but similar objects, which we believe is an interesting direction for future work.

Thanks for the authors' response. I still think explicit comparisons to existing methods (WorldCoder, TheoryCoder) are needed. And my other concerns remain unaddressed. My score remains the same.

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

The paper proposes HYPER, a neuro-symbolic agent that learns environment dynamics through LLM-based conjecture and refutation. The system generates hypotheses in response to planning failures, novelties (divergences between predictions and observations), or failed experiments. It revises hypotheses based on experimental outcomes and incorporates accepted hypotheses as beliefs into an action model used for classical planning. A provenance ontology records the evidence supporting each belief. The authors evaluate the framework in four custom gridworld environments, comparing it with a variant without refutation and an LLM-based planning baseline.

Critical strengths
  1. The experiments provide preliminary evidence that incorporating refutation benefits performance on a challenging task. In fire crossing, the full system succeeds in 2/3 runs with GPT and 1/5 runs with Kimi, while the corresponding variants without refutation and LLM-based planning baselines achieve no successes. These results offer initial support for combining conjecture with refutation to improve task performance.
Critical weaknesses
  1. The assumption underlying gradual correction of the action model is not directly evaluated. The paper assumes that the system validates more correct than incorrect beliefs and uses this assumption to motivate the gradual improvement of its action model. However, the experiments report task success and hypothesis acceptance rates without distinguishing correct from incorrect accepted beliefs or quantifying how erroneous beliefs are corrected. This assumption therefore lacks direct empirical support.

  2. The experiments do not sufficiently isolate the contribution of refutation. Removing refutation also removes the exploration and observations associated with those experiments. Moreover, the termination limits count goal-directed and/or conjecture–refutation processes, which do not establish a common environment interaction budget across methods. Consequently, the current comparisons do not clearly distinguish gains from rejecting incorrect hypotheses from gains due to additional exploration and observations.

Questions for authors
  1. Does refutation improve the action model? Could you provide a complete trace of an actual conjecture–refutation process and show whether revision improves predictions on states not used in the experiments? Can the existing logs quantify how often incorrect hypotheses are accepted as beliefs and how often these erroneous beliefs are subsequently corrected?

  2. Can the benefits of refutation be distinguished from those of additional exploration? Does the full system still outperform the variant without refutation under the same environment interaction budget? Do the gains persist when the baseline is allowed additional exploration to collect more observations?

  3. What are the specific advances over closely related work? WorldCoder already separates LLM-driven world model learning from independent planning, while POPPER uses LLMs to design and execute falsification experiments. What key design advances does HYPER introduce in experiment selection, belief integration, or action model revision? Which experimental results establish the contributions of these advances? WorldCoder, NeurIPS 2024, POPPER, ICML 2025

  4. Do the gains extend to a broader range of settings? The current evaluation includes only four environments. Have you evaluated the method in more diverse environments or on benchmarks used in prior work, such as MiniGrid in WorldCoder? Do these results support the method’s effectiveness across a broader range of settings?

1 comment

We thank the reviewer for their questions and extensive comments. We will proceed to address them, one by one.

Q1

There is much difficulty in evaluating objective correctness of beliefs from the perspective of the agent. Refutation exists as a way to pre-emptively validate these hypotheses and to serve as a first filter.

In the case that these beliefs are wrong in a meaningful way, one could expect them to trigger a novelty once again: using the logs, it is possible to see how often this kind of erroneous beliefs are subsequently corrected. However, finding the kind of erroneous beliefs that do not cause new novelties (eg because they take no part in the tasks executed, or exist in a context the agent cannot experience) will require a human expert doing qualitative validation, which also introduces a subjective factor. There are several approaches to take for a human evaluation, including but not limited to a) detecting incorrect hypotheses being validated by inspecting belief integration processes, b) detecting incorrect belief corrections by inspecting novelty-conjecture-refutation-integration chains of processes, or c) evaluating the extensiveness of the experimental support of a validated learnt belief.

Q2

We acknowledge the limitations in our experimental setup. A stronger, more focused comparison between full and no-refutation HYPER, with fixed environment interaction budget, could be made.

By examining our experimental runs, we have observed a repeated phenomenon in which the Hypothesiser module proposes a hypothesis too restrictive compared with the real dynamics. Without refutation, it produces a world model that makes it impossible to generate a plan to the agent's goal, needing the intervention of the Hypothesiser in further conjecture processes to correct this failure. Instead, a refutation process can catch this before it can affect the learning process.

Q3

HYPER differentiates itself from agents such as WorldCoder by presenting an explicit conjecture and refutation mechanism that explicitly validates the dynamics it hypothesises. It also maintains a trace that serves as epistemic support of everything it learns.

As opposed to POPPER (which needs hypotheses to be externally given), HYPER is an end-to-end architecture that is autonomously able to detect unpredicted phenomena, hypothesise a plausible cause, and design a series of experiments to falsify it.

Q4

We acknowledge that a broader evaluation could be made by including tasks of different nature, but we considered that there was a risk in taking broadly-used benchmarks, as these are likely to appear in LLM training data and thus result in strongly biased results.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.