acceptodds
Under review as a conference paper at ICLR 2027

Induction and Adaptation in Natural Language

Abstract

A general AI agent cannot anticipate every environment it will encounter and must therefore continually adapt its understanding of the world based on observations made at deployment. We investigate the benefits of representing and compressing such beliefs in natural language, which, paired with a pre-trained language model, forms a highly flexible world model that can map states and actions to predicted next states. Unlike the scalar signals that drive traditional neural world models, natural language is interpretable, editable, and portable. We treat learning such a world model as inference over a natural-language representation of the world, performed by searching directly in this space. To guide this search, we use a language model as a proposal that generates hypotheses from a strong prior, grades them against observations, and refines them in response to prediction errors. We first evaluate how well our method induces the underlying transition function across a range of induction tasks, where it substantially outperforms in-context learning baselines on next-state prediction. We then measure how these induced world models can support downstream interactive decision-making on ARC-AGI-3, a benchmark that requires agents to explore complex environments and infer their dynamics to solve novel tasks, and find that they serve as effective representations that drive strong downstream task performance.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
induction, world models, adaptive AI, natural language

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_MGmWOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces NL-LAN, a training-free method that represents an agent's world model as natural-language text. A frozen LLM proposes candidate world models, scores each by how well it predicts observed transitions, and revises them using the prediction errors in a beam search. On ARC-AGI-3, this beats in-context learning at next-state prediction (82.2% vs. 52.4%). In online play it solves 24 levels, against 12 for the same agent without the induced world model. The gains are smaller on other benchmarks and come at 4–13× the token cost.

Critical strengths
  • The method shows a large gain on ARC-AGI-3. It reaches 82.2% held-out next-state accuracy against 52.4% for the best in-context baseline, and 94.7% against at most 43.3% on transitions already in context (Table 1).
  • The ablations point to search as the mechanism. One-shot induction (NL-LAN-Low) gives limited gains, while refinement drives most of the improvement (ARC 30.5 → 82.2, Sequence Hard 15.0 → 73.0; Table 2), and Fig. 3 shows that more revisions also reduce variance. This matters because it ties the gains to a specific algorithmic component rather than to prompting a strong model.
Critical weaknesses
  • The control result rests largely on one game and one run. ar25 accounts for about 53% of NL-LAN’s total ARC-AGI-3 score (Table 12). Without it, the average falls from 3.10 to roughly 1.46, against 0.97 for the no-induction variant. The no-induction variant also wins on six games, and no seeds or error bars are reported.
  • I think the control baselines are not comparable. The ICL scores come from the public leaderboard, so they might use a different harness, prompts, and budget. NL-LAN’s own scaffold without induction already beats every one of them. This matters because the comparison against frontier agents mixes the effect of the scaffold with the effect of induction, and only the ablation within the paper is controlled.
  • There is no compute-matched baseline. NL-LAN spends about 4–13× more tokens than ICL (Table 11). Section 4.3 only takes compute away from NL-LAN and never gives ICL an equal budget, for example through self-consistency or larger reasoning budgets. This matters because the paper cannot yet rule out that much of the gain comes from extra inference compute rather than from the text latent.
Questions for authors
  1. Is the 1–4 cell pixel tolerance used for NL-LAN on ARC-AGI-3 also applied when scoring the in-context baselines in Table 1? If the baselines were scored by exact match, please report both methods under the same criterion.
  2. How does NL-LAN compare with an in-context baseline given a matched token budget, for example self-consistency or majority voting over multiple samples, or a larger reasoning budget? Without this, the gains cannot be separated from the extra inference compute.
2 comments

We sincerely thank the reviewer for their constructive feedback. We are glad they appreciate the strong predictive performance and the evidence supporting iterative refinement as an important component of NL-LAN. We address their concerns below, with manuscript changes highlighted in red (or violet where shared with Reviewer 2).

W1: Robustness of control results.

We agree that the strong performance on ar25 contributes substantially to the overall ARC-AGI-3 score and that repeated runs would provide a more robust estimate of performance variability.

Nevertheless, the improvements are not limited to this environment. Excluding ar25, NL-LAN still solves 19 levels compared to 12 for the no-induction variant, with a remaining score contribution of approximately 1.46 versus 0.97 (under the original 25-game normalisation). This suggests that the induced world models provide benefits beyond a single game, although the magnitude of the overall improvement is clearly influenced by ar25.

Unfortunately, repeated evaluations across all 25 interactive environments are prohibitively expensive due to the substantial LLM inference costs associated with induction, refinement, and control. We therefore follow a single-run evaluation protocol, while reporting per-environment results (Table 12) to make the variability across tasks transparent.

We acknowledge that these results do not establish statistical robustness and have clarified this limitation in the manuscript. Nevertheless, the aggregate improvement beyond ar25 provides encouraging evidence for the practical utility of our approach.

W2: Comparability of control baselines.

We agree that the public ARC-AGI-3 leaderboard results are not controlled comparisons, as agents differ in scaffolding, prompts, and inference budgets. Indeed, our no-induction variant already outperforms these baselines, indicating that the acting scaffold itself contributes substantially to performance. We therefore treat the leaderboard results as contextual reference points for overall system performance, rather than as evidence that isolates the effect of induction. The more informative comparison for this purpose is our within-framework no-induction ablation, although we acknowledge that inevitable differences in prompts due to the absence of a world model prevent perfect isolation. 

W3 & Q2: Compute-matched comparison.

We agree that inference compute is an important factor. NL-LAN deliberately allocates additional compute resources for model induction, whereas the ICL baseline already uses maximum reasoning effort and a 65,536-token limit. 

To investigate compute matching, we evaluated ICL with majority voting over independent samples on Sequence Hard (Gemini 3.1 Pro, high reasoning). At k=3 samples, ICL approximately matches NL-LAN's Pro token usage; we additionally test k=5 and k=10 to account for NL-LAN's Flash calls. Accuracy improves from 42.9% (k=1) to 45.9% (k=3), 48.6% (k=5), and 52.0% (k=10), but remains below NL-LAN's 73.0%, with diminishing returns. These results suggest that NL-LAN's advantage is not merely due to increased inference compute, but also to how that compute is allocated toward explicit world-model induction and refinement.

Q1: Pixel-error tolerance.

The environment-specific pixel-error tolerance (1–4 cells) is used both during NL-LAN's candidate scoring and held-out evaluation, where it is applied consistently to NL-LAN and all direct ICL baselines.

While comparing threshold-based and strict exact-match evaluation for Reviewer 2, we identified an inconsistency in how the thresholds were applied for sc25 (4 cells) and ka59 (1 cell). We have corrected this and recomputed the ARC-AGI-3 results under a consistent protocol. We apologise for the oversight and report the corrected threshold-based accuracies below for transparency.

SetupTrainTest
ICL Gemini 3.1 Pro23.113.3
ICL Opus 4.745.3 (+2.0)56.2 (+3.8)
NL-LAN Gemini 3.1 Pro → 3 Flash94.782.2
NL-LAN Opus 5 → Sonnet 590.9 (−1.1)70.2 (−1.0)

The main conclusions remain unchanged. We have clarified the evaluation protocol in Appendix A.2 and updated all affected results throughout the manuscript.

We thank the reviewer again for their feedback, which helped us strengthen our experimental analysis and better qualify our conclusions.

Thank you authors for your comments. I understanding there are substantial cost associated with additional runs. Other than that, I think most of my concerns are addressed and therefore, I would like to increase my assessment to weak acceptance.

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

This paper introduces NL-LAN, which represents environment dynamics through editable natural-language descriptions. Frozen language models propose descriptions from observed transitions, evaluate them through next-state predictions, and revise them using prediction errors. The resulting representations are updated as experience accumulates and used to guide action selection. The authors evaluate prediction on sequence transformations, AutumnBench, and ARC-AGI-3, and interactive control on ARC-AGI-3. They report improved predictive performance and control results of 3.10 points with 24 levels solved, compared with 0.97 points and 12 levels when the induced representation is masked.

Critical strengths
  1. The paper connects an interpretable representation of dynamics to an operational adaptation procedure. Natural-language descriptions are tested through decoded next-state predictions and revised using observed errors, giving the representations an explicit behavioral criterion. The resulting descriptions can then guide actions while the underlying language models remain frozen.

  2. The experiments demonstrate utility in both prediction and interactive control. Evaluation spans sequence transformations, AutumnBench, and ARC-AGI-3. In the control experiment, the full system achieves higher points when the induced representation is masked. This within-framework comparison supports the practical contribution of the induced information to action selection.

Critical weaknesses
  1. The distinct contribution over existing induction methods and the fairness of the comparisons remain insufficiently established. Section 2 discusses directly relevant work on textual hypothesis refinement and WorldCoder, but matched comparisons with these methods are missing. In Table 11, ARC token usage is 296.3M versus 38.8M for ICL, approximately 7.6 \times higher, while the Low variant changes both search and the decoder. The results therefore do not adequately separate the effects of natural-language representation, iterative search, and additional model computation, limiting conclusions about the method's unique advantages.

  2. I think the guarantee that updates preserve previously correct predictions lacks algorithmic and experimental support. Section 3 bases this claim on full-history scoring, but Appendix A.2 actually selects at most 10/20 transitions for scoring. Even over the full history, a higher total number of correct predictions permits some previously correct examples to become incorrect. The guarantee therefore neither follows from the current selection rule nor receives dedicated regression analysis. Retaining this claimed advantage of adaptation would require a revised mechanism or direct evidence; otherwise, the conclusion should be narrowed.

Questions for authors

1&2 see weaknesses

  1. Section 4.5 states that the no-induction variant masks the world-model input to the policy, but the goal-belief update in Appendix B.1 also reads that representation. Is this information path masked as well? After clarifying both components' information access, could the authors report repeated runs under a common action budget and full control costs?

  2. Appendix A.2 and Table 4 allow a 1–4-pixel error tolerance for ARC. Is this used only for candidate scoring during induction, or also for final test accuracy, and is it applied consistently to every baseline? Could the authors report both strict-match and tolerance-based results to establish whether the main gains depend on this scoring choice?

4 comments

We sincerely thank the reviewer for their detailed feedback. We are glad they appreciate the interpretability of our approach and its demonstrated utility. Below, we address their concerns and describe the corresponding manuscript revisions, highlighted in blue (or violet where shared with Reviewer 3).

W1: Distinct contribution and fairness of comparisons.

We agree that our approach builds on existing work on textual hypothesis induction and refinement. Our contribution is not hypothesis search in isolation, but its use for continual world-model induction, where natural-language hypotheses about environment dynamics are explicitly evaluated through decoded transitions, refined against prediction errors, and subsequently used for interactive decision-making. Unlike programmatic approaches such as WorldCoder, our representation does not require an executable specification of the dynamics; instead, it leverages the pretrained LLM as a decoder.

Compute matched baseline. We agree that inference compute is an important factor. NL-LAN deliberately allocates additional compute resources for model induction, whereas the ICL baseline already uses maximum reasoning effort and a 65,536-token limit. 

To investigate compute matching, we evaluated ICL with majority voting over independent samples on Sequence Hard (Gemini 3.1 Pro, high reasoning). At k=3 samples, ICL approximately matches NL-LAN's Pro token usage; we additionally test k=5 and k=10 to account for NL-LAN's Flash calls. Accuracy improves from 42.9% (k=1) to 45.9% (k=3), 48.6% (k=5), and 52.0% (k=10), but remains below NL-LAN's 73.0%, with diminishing returns. These results suggest that NL-LAN's advantage is not merely due to increased inference compute, but also to how that compute is allocated toward explicit world-model induction and refinement.

Decoder-controlled ablation. In our original experiments, NL-LAN-Low used Gemini 3.1 Pro as the decoder to align the model choice with the ICL baseline, which uses Gemini 3.1 Pro throughout. However, we agree that changing both factors complicates the interpretation. We therefore reran NL-LAN-Low with the same Gemini 3 Flash decoder as in full NL-LAN, keeping the remaining configuration unchanged.

Test accuracy changes from 30.5% to 28.5% on ARC-AGI-3 and from 84.7% to 77.7% on AutumnBench; Sequence Easy and Hard remain unchanged at 100.0% and 15.0%. Importantly, the substantial improvement from NL-LAN-Low to full NL-LAN persists. This strengthens our conclusion that iterative search and refinement, rather than the choice of decoder, drive much of the improvement. Since we were convinced that keeping the same decoder is the cleaner ablation, we updated the manuscript accordingly. 

W2: Preservation of previously correct predictions.

We thank the reviewer for identifying this overstatement. We agree that scoring candidates on previous observations does not guarantee that every previously correct prediction remains correct, even when using the full trajectory. The approach encourages consistency with earlier observations while keeping evaluation costs manageable, but does not provide a strict non-regression guarantee. Section 3.4 and Appendix B.1 now clarify this and explicitly describe subset-based scoring.

Q3: Information access in the no-induction ablation.

For the no-induction ablation, no world model is induced or supplied to either the policy or goal-belief updater. The agent instead uses its current observation, running goal belief, and recent action history. Its prompts are matched as closely as possible to the NL-LAN-based run, except for removed world-model references and exploration instructions tied to an incomplete world model.

We acknowledge that prompt differences may contribute to the observed improvement, preventing attribution exclusively to induction. This is now clarified in Section 4.5 and Appendix B.1.

Repeated control evaluations across all 25 environments were prohibitively expensive. We therefore report single-run results and per-environment outcomes (Table 12). Although exact costs were not logged, approximate token usage estimated from provider-side data is:

RunFlashProTotal
Full NL-LAN~3.8B~0.65B~4.45B
No induction~0.2B~0.35B~0.55B

Q4: Pixel-error tolerance in ARC-AGI-3.

The environment-specific pixel-error tolerance (1–4 cells) applies during induction and held-out evaluation for all methods, including ICL, accounting for environment-specific effects such as non-deterministic fuel-bar changes.

While re-examining the evaluation pipeline, we identified an inconsistency in how the threshold was applied for sc25 (pixel threshold 4) and ka59 (threshold 1). We corrected this and recomputed the results. We apologise for the oversight and report both tolerance-based and strict exact-match accuracies:

SetupTrainTrain exactTestTest exact
ICL Gemini 3.1 Pro23.121.013.310.8
ICL Opus 4.745.3 (+2.0)41.456.2 (+3.8)35.7
NL-LAN Gemini 3.1 Pro / 3 Flash94.763.082.260.5
NL-LAN Opus 5 / Sonnet 590.9 (−1.1)70.770.2 (−1.0)49.7

The main conclusion remains unchanged: NL-LAN outperforms direct ICL under both evaluation criteria. Appendix A.2 now clarifies the protocol, and all corrected results are reflected throughout the manuscript (violet).

We thank the reviewer again for their valuable feedback, which helped us strengthen the methodological clarity and transparency of our evaluation.

Thank you for correcting the ARC evaluation and confirming that the prediction gains persist under exact-match scoring.

I maintain my rating. A matched comparison with prior hypothesis-refinement methods is still missing, and differences in prompts and compute make the control gains difficult to attribute to the induced world model.

Thank you for your follow-up. We understand your remaining concerns but would like to clarify two points:

  • We agree that a comparison with prior hypothesis-refinement methods would be interesting. However, our contribution is not a claim of superiority over existing hypothesis-search algorithms, but the use of explicit natural-language hypotheses as adaptable world models for prediction and downstream interactive decision-making.
  • The no-induction ablation uses the same acting framework, removing the world model from both the policy and goal-belief updater. Prompt differences are limited to instructions that depend on the world model's availability and are therefore a natural consequence of the ablation rather than independent changes to the framework. We acknowledge that this does not perfectly isolate the world model's contribution and have updated our claims accordingly.

We hope this clarifies the scope of our contributions and the interpretation of the control results. Thank you again for your thoughtful feedback.

Reviewer_bBS6OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces a technique of using a pre-trained LLM, appropriately prompted, to perform the role of a world model. The authors report a number of experiments showing the promise of their approach compared to other in-context learning techniques, including in such challenging tasks as ARC AGI 3.

Critical strengths

The paper is clearly written and is a pleasure to read. I especially appreciated the literature review section that actually engages with existing research and places this new work in the broader context in non-trivial ways.

The idea of using an LLM as an “engine” for a world model is interesting and promising.

Critical weaknesses
  1. I felt like some of the paper’s claims were overstated, especially when it comes to contrasting regular world models and the proposed method.

Early on, the authors argue against static traditional world models and in favor of LLM-based ones.

E.g. “Despite this, most learned world models in modern machine learning remain static.”

I think this might be a bit of an oversimplification/overstatement of the difference. Indeed, on some level, most world models remain static due to fixed weights, but the same can be said about the proposed method, driven by a fixed-weight LLM. I resonate with the general statement, but fixed weights don’t necessarily mean rigidity in the way the statement seems to imply. For example, it is conceivable that the evolution of a distributed hidden state in some recurrent world model might hide similar semantics as the method the authors are proposing, just without explicit verbalization.

I believe it’d greatly benefit the paper to go through the first few sections and clarify the distinction, and perhaps to tone down the juxtaposition.

  1. While the authors acknowledge the fact that their choice of tasks is particularly suited to LLMs, I believe it would be more appropriate to mention it earlier, and, ideally, have at least some simple experiments in poorly LLM-suited domains (timescales, levels of abstraction, etc.). I believe it’d be most valuable for the community to know and understand when it’s promising to use an LLM as a world model and when it’s not.
Questions for authors
  1. Continuing the critical section above,

An LLM as a world model is also frozen. What, in your view, makes LLMs more universal as the “world model engine”? It is tempting to think that it’s due to properties of language itself, but it could also just be a consequence of the tremendous amounts of data that modern LLMs are trained on, that modern world models can only dream of (pun intended). What is your take on this?

  1. (Minor suggestion) I think the way the authors use the term in-context learning is potentially confusing. The newly proposed method can also be seen as ICL. It might be better to reframe it, otherwise some readers might assume that the proposed approach involves some prompting.
2 comments

We sincerely thank the reviewer for their thoughtful and constructive feedback. We are glad they appreciated the writing and literature review, and found the paper's idea “interesting and promising”. We address the concerns below, and any changes related to the review are marked in green in the manuscript.

W1: Flexibility of NL-LAN compared to conventional world models.

We agree that our original wording was too strong. Both conventional world models and our approach typically rely on fixed model parameters, and we agree with the reviewer that other models, such as recurrent models or other forms of latent adaptation models, are inherently capable of modelling an evolving world at test time.

Our intended distinction is therefore not that conventional world model architectures are inherently rigid, but rather a comment on the objectives of such models. Most world models are trained to target a certain distribution, and learning/updating the world model given new data does not occur explicitly at test time, whether or not such an adaptation mechanism exists in the architecture.

In contrast, NL-LAN uses an explicit space designed to adapt the model of the environment at test time (which can also be thought of as a parameter space), with natural-language latent representations used for this space. This provides an interpretable and directly configurable interface for adapting the model to different settings without updating its core weights. However, this flexibility, of course, remains bounded by the capabilities of the underlying LLM and the expressiveness of the natural-language representation.

We revised the introduction accordingly, avoiding overly conflating the ability to adapt or maintain an evolving state at test time with the objective of modelling an evolving representation of the environment at test time. Instead, we target the tendency of world models to maintain a fixed train-test divide, where they are not explicitly trained to update their world model at test time.

W2: Applicability to domains less suited to LLMs.

We agree that identifying when LLM-based world models are beneficial, and where their limitations lie, is important. Our experiments deliberately focus on settings where relevant states and dynamics can be meaningfully expressed in natural language, which we acknowledge limits the generality of our conclusions. We clarified this scope earlier in the paper.

To further investigate these limitations, we added Appendix E (last pages), introducing three controlled environments that probe compact numerical, complex state-dependent numerical, and stochastic dynamics. We also discuss challenges posed by fine-grained temporal dynamics. This section highlights settings in which specialised models or extensions to our framework may be more appropriate than NL-LAN alone, and is a valuable addition to the paper.

Q1: What makes LLMs more universal as world model engines?

We believe the versatility of LLMs as world model engines stems from a combination of broad prior knowledge acquired during pretraining and the flexibility of natural language as a shared, compositional interface. We agree that the scale and diversity of pretraining data likely play a major role, rather than this advantage arising from language alone.

An additional advantage of natural language is its ability to combine different levels of abstraction within a single representation, ranging from high-level semantic descriptions (e.g., "the background changes colour") to precise spatial rules (e.g., "the object moves three pixels to the right"). While neural world models can also capture such abstractions implicitly, natural language makes them explicit, interpretable, and directly editable.

Our claim is thus primarily about practical reusability and interpretability, rather than inherently superior expressivity or universal generalisation. As conventional world models benefit from increasingly large and diverse training datasets, it will be interesting to investigate how the respective strengths of explicit natural-language representations and implicit learned dynamics evolve.

In-context learning terminology.

We thank the reviewer for this suggestion and agree that NL-LAN can itself be viewed as leveraging in-context learning. Our intended distinction is between direct in-context prediction and prediction through an explicit natural-language latent representation. We clarified this terminology when introducing the baseline.

We thank the reviewer again for their constructive suggestions, which helped us clarify the scope and positioning of our contributions.

Thank you again for your constructive review. We wanted to briefly follow up on our response and the revised manuscript, particularly the clarified positioning relative to conventional world models and the new experiments in Appendix E. We would be happy to address any remaining questions or concerns.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.