Passive learning of active causal strategies in agents and language models

What can be learned about causality and experimentation from passive data? This question is salient given recent successes of passively-trained language models in interactive domains such as tool use. Passive learning is inherently limited. However, we show that purely passive learning can in fact allow an agent to learn generalizable strategies for determining and using causal structures, as long as the agent can intervene at test time. We formally illustrate that learning a strategy of first experimenting, then seeking goals, can allow generalization from passive learning in principle. We then show empirically that agents trained via imitation on expert data can indeed generalize at test time to infer and use causal links which are never present in the training data; these agents can also generalize experimentation strategies to novel variable sets never observed in training. We then show that strategies for causal intervention and exploitation can be generalized from passive data even in a more complex environment with high-dimensional observations, with the support of natural language explanations. Explanations can even allow passive learners to generalize out-of-distribution from perfectly-confounded training data. Finally, we show that language models, trained only on passive next-word prediction, can generalize causal intervention strategies from a few-shot prompt containing examples of experimentation, together with explanations and reasoning. These results highlight the surprising power of passive learning of active causal strategies, and may help to understand the behaviors and capabilities of language models.

Paper

Similar papers

Peer review

Reviewer U5ax7/10 · confidence 3/52023-07-03

Summary

This paper provides a thorough analysis of whether agents trained to imitate passive data can indeed learn active strategies for intervening during evaluation. In particular, this paper demonstrates that agents can "learn" causal structure from passive data in that they correctly intervene on the appropriate variables during testing, even if such interventions were not encountered during training. A formal "existence proof" for active causal learning is given. Experiments are demonstrated on a causal DAG environment, a pixel-based shapes environment, and few-shot prompting of a language model.

Strengths

*Originality:* The paper provides an interesting demonstration that refutes some claims from the causality literature. In particular, the paper shows that "it is possible to learn generalizable strategies for causal experimentation and intervention from passive data alone" given that the data was generated from an expert's actions. *Quality:* the experiments appear technically sound and are well-executed. *Clarity:* Overall the paper is well-written, with understandable diagrams + captions. Moreover, the flow of the paper is impressive, with the authors smoothly anticipating and responding to potential concerns. *Significance:* The paper qualifies some claims surrounding agents' abilities to learn active causal strategies from passive data. In particular, this extends our understanding of the limits of agent learning. Moreover, this may amplify the concerns for safety of such agents, as they may appear to learn generalizable reasoning strategies even through passive data.

Weaknesses

While overall the paper is technically sound, the first claim made in the contribution seems a bit too strong when compared against the setting given by the proof. In particular, the authors claim "We formally explain why it is possible for an agent to learn a generalizable strategy for exploiting causal structure from passive data, as long as the agent can intervene at test time". However, their proof relies on an assumption that the system is accurately represented by a finite DAG, which may be unrepresentative of real-world scenarios, which may include cycles due to feedback loops. It would be more accurate to characterize their contribution as an "intuitive explanation of why agents may be able to learn generalizable strategies for exploiting causal structure under certain settings, namely that the system can be represented as a finite DAG."

Questions

Were any ablations done to study how the coverage of the training distribution affects the ability of the agent to learn causal strategies? For example, in the adaptive experimentation setting in the DAG environment the expert policy only explores all the relevant nodes. What happens if the expert policy explores more nodes than the causally relevant nodes? The agent still has enough information to identify the causal structure, but it would be interesting to see if it is able to behave optimally in this case. More generally, do the authors expect the learning of active causal strategies from passive data to be efficient when presented with more complex environments?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

4 excellent

Contribution

3 good

Limitations

The limitations are adequately addressed.

Reviewer KP4B8/10 · confidence 5/52023-07-06

Summary

The paper studies the possibility of learning to identify new casual structures through purely passive learning, which entails the feasibility of acquiring robust task performance in novel conditions through offline learning. The authors hypothesize that it is possible to learn generalizeable causality exploration strategies from offline data that demonstrate such strategies. An agent can then apply these strategies to explore the casual structures of new environments and effectively achieve goals in them. Experiments on several synthetic environments show that explanations play a crucial role in enabling learning generalizeable exploration strategies. These results provide further insights on the ability of LLMs to quickly acquire interactive skills, even when trained exclusively on offline data.

Strengths

- The paper studies a fundamental question whose answer can have profound impact on the understanding and development of AI agents. It challenges a long-standing assumption about the limitations of learning from observational data. It suggests that observational data are not all the same; certain forms observations can be more useful than others. The new understanding is powerful. It can help explain the recent successes of LLMs, which are mostly trained through observations of the world. More general, it offer valuable insights to better leverage offline data to reduce the cost of collecting interactive data. - The experiments are well-designed and the analyses are insightful. The results can lead to further theoretical and empirical investigation, especially on the role of the explanations in learning. - The paper is very clear and well-written. The arguments are nicely presented.

Weaknesses

A major weakness of the paper is that it is mostly empirical. While the authors provides intuition on why their hypothesis can be true, the paper is still missing a satisfactory theoretical explanation. The environments studied, while having interesting structures, are artificially constructed. The paper may further benefit from having experiments on real-world tasks. The artificial environment may have odd properties that allow generalization to occur. The main hypothesis rests on the assumption that an agent can learn generalizeable exploration strategies. While the empirical results support this assumption, it is unclear why it should hold in theory, because the strategy demonstrations are also observational data. Why is it easier to learn generalizeable exploration strategies than directly learning generalizeable exploitation strategies? I'd strongly suggest the authors add ablation experiments that dissect the importance of the exploration phase (i.e. results without exploration). This would strengthen the claim that the exploration phase is needed. The authors emphasize the role of explanations in acquiring robust exploration strategies, but they neither clearly define "explanations" nor show examples in the main paper. This term is quite general and can mean many different things. While some examples are shown in the appendix, I think it should be more explicit in the main paper what type of explanations are used, what causal structures/properties they describe. I would also like to see more discussions or experiments on the effect of the explanation medium on the performance. Do explanations have to be in natural language? Or do they just need to have certain properties? Related work: [1] Interactive learning from activity description: provides descriptions of activities as feedback to agents, leading to more efficient and effective learning than RL. [2] Distilling Step-by-Step! Outperforming Larger Language Models with Less Training Data and Smaller Model Sizes: shows that learning to predict explanations can enhance the pre-training of LLMs.

Questions

See my "weakness" section.

Rating

8: Strong Accept: Technically strong paper, with novel ideas, excellent impact on at least one area, or high-to-excellent impact on multiple areas, with excellent evaluation, resources, and reproducibility, and no unaddressed ethical considerations.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

4 excellent

Presentation

4 excellent

Contribution

4 excellent

Limitations

The authors compose a very thoughtful and thorough discussion section. I like that the authors anticipate and warn readers to avoid mistakenly interpret the claims made in the paper as passive learning can be as good as active learning.

Reviewer X6cj7/10 · confidence 4/52023-07-06

Summary

This paper investigates whether it is possible for an agent to learn how to perform and make use of active causal interventions, despite being trained on purely passive / observational data (e.g. supervised learning, behavioral cloning, etc). The authors first argue that this is possible in principle, if the observed data contains behavioral traces of an expert who performs causal interventions in an environment to learn about its causal structure, and then exploits the learned information to achieve a goal. They then perform experiments in a causal DAG environment, a visual "odd-one-out" environment, and linguistic version of the "odd-one-out" environment (in order to test pretrained LLMs). In the casual DAG environment, they find that trained agents can learn to generalize causal exploration and exploitation strategies to unseen DAGs, just by passive learning of expert strategies on the training set. They find similar results in the odd-one-out environment, with stronger performance achieved when the training data includes natural language explanations of the interventions performed by the expert. Finally, they find that LLMs prompted with few-shot examples of causal experimentation can generalize to new settings where the latent cause of reward is absent from the few-shot examples, provided that explanation and reasoning traces are also provided in the few-shot examples.

Strengths

This was a well-written paper that addresses an important topic of debate that has arisen about the capabilities of models trained purely on observational data, including pretrained (large) language models. I have several questions about the details of the experiments, and about the generality of causal experimentation strategies that are learned from observations of an expert, but overall I think the paper presents clear theoretical arguments and solid empirical demonstrations for its claims, showing that learning causal intervention strategies from offline/data is possible in principle, and is exhibited in practice to at least some degree of generality in actual ML models. The DAG experiments are especially well designed, showing that passively learned behavior is not well explained by approximate heuristics, and better explained by optimal causal reasoning. The odd-one-out and LLM experiments are more open to interpretation due to the absence of heuristic or correlational baselines, and not knowing whether the LLM was trained on similar data, but still provide evidence that passive learners may acquire generalizable causal intervention strategies in non-trivial settings (insofar as it is unlikely for the high performance to be achieved by a learned heuristic).

Weaknesses

While the experiments have been designed carefully enough to show that observed casual strategies are generalized to some extent, I think they could be extended in the number of ways to probe the limits of this generalization. In the DAG experiments for example, it is clear that the agent must have learned some kind of permutation invariant strategy for experimentation and intervention, given that the training data entirely lacks causal links that are present in the test episode. However, I think the experiments still leave open exactly how abstract the learned strategy is -- does it generalize to DAGs with 6 nodes, or exhibit size generalization more generally? (I expect not, due to the network architecture.) Can it generalize to DAGs with non-linear causal relationships? And can it adapt its experimentation strategy when it is *not* cued about which variables are relevant, perhaps by learning to imitate an expert that uses an optimal experimental design that minimizes the number of exploration queries before exploiting? (I found the explicit cues about which variables are relevant to be a fairly strong simplifying assumption -- it would be good to at least vary the number of variables which are cued as relevant to a wider range of numbers.) Similar size generalization tests would be interesting to test in the "odd-one-out" and LLM experiments -- e.g. having a test trial with 4 distinct objects, not just 3. For the LLM experiments in particular, it would also be good to say something about whether similar prompts are likely or unlikely to have appeared in Chinchilla 70B training data. Generalizability is related to a more general question I have about what exactly "counts" as having learned an active causal strategy. In its strongest form, this might require verifying that a learned model actually implements a causal discovery algorithm and causal exploitation algorithm that is optimal, or at least sound and complete. Of course, this is very hard to verify without strong mechanistic interpretability techniques, and in the absence of that, heuristic baselines are helpful to compare against. Because heuristic baselines were included in the DAG experiments, I'm willing to believe that the model has successfully learned a close approximation to the optimal exploitation strategy in the n = 5 case, which presumably requires the agent to simulate the effect of a causal intervention on an internal representation of the DAG structure that was identified from the experimentation phase. But because similar baselines aren't included in the "odd-one-out" or LLM experiments, it could be that the learned strategy in those cases is just some "clever trick" that happens to do well most of the time without "true" causal reasoning. As such, it'd strengthen the paper to show some heuristic or correlational baselines for those experiments, if only to demonstrate that there are no obvious heuristics that can easily achieve the ~90% success rates in the odd-one-out and LLM experiments. For example, one baseline might be to always transform the color of the first/nearest object during experimentation (and to pick the unique colored object in the test trial if that transformation is successful). Another heuristic might be to always pick the object based on the feature that was most often unique in the training data. In addition, I think it'd good to note the limitations of what we can conclude about a model's learned strategies based on observational equivalence alone. Regarding the need to predict/include linguistic explanations in the "odd-one-out" experiments, it was unclear to me whether the results shown in Figure 3 (and in the corrected version, Figure 13) are from a model that was also trained to predict natural language explanations. My guess is that they are, but I think the presentation clarity of this aspect could be improved, e.g. by stating in the caption of Figure 3 that the results shown involve predicting the explanation, by giving an example in the Appendix of what a full training episode (including the natural language explanations) looks like, and by explaining how the natural language explanations are generated (I had to consult the Lampinen et al (2022) paper in order to determine this). Finally, I think it's somewhat misleading to describe the passive data in section 5.2.1 as "perfectly confounded" -- after all, the explanations constitute observational data that makes the setting no longer confounded! After all, it is only in worlds where some feature X causes reward R that choosing a object unique in feature X would result in the explanation E = e_X+ ("In this game, I was rewarded for unique X", and choosing an object unique in some other feature would result in the explanation E = e_X- ("The rewarding dimension must not be Y." where Y != X). Hence, the relevant part of the causal DAG is: R <- X -> E Once the agent observes E = e_X, it can reliably infer that there is causal link from X -> R in the true causal graph, since X is a common cause of both R and e_X. So the fact that explanations "deconfound" is no surprise at all.

Questions

See above for various questions and suggestions I have about additional baselines, additional generalization tests, clarifying the role of explanations in the odd-one-out experiments, whether these models can be said to have "truly learned" causal reasoning, and whether the term "perfectly confounded" is apt. In addition, I'm curious whether the authors still endorse the following analysis, given that their post-submission corrections show much better performance of passive learners in the odd-one-out experiment: > Indeed, although our passive agents performed far above chance in the odd-one-out environments, they did not achieve as robust of performance as the online RL agents in earlier work [30]. This gap may be due to irreversible choices in these settings, which are difficult to learn from purely positive examples, without some active correction of off-optimal-policy behavior [e.g. 48, 11]. Post-Rebuttal Response: The authors have addressed my questions and suggestions, and I continue to recommend acceptance.

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

4 excellent

Limitations

Some limitations of their findings are discussed (see "What this paper does not imply:"), and the authors fairly summarize their findings as showing that passive learning of causal strategies is restricted to the case where expert intervention policies are actually present in the data. I think it'd be valuable to add discussion of exactly how much we can expect the learned causal strategies to generalize, and whether it is possible to establish if they actually constitute sound and complete causal discovery algorithms.

Reviewer c7Wr7/10 · confidence 4/52023-07-07

Summary

The paper deals with the problem of understanding whether causal structures can be learned from passive observational data. The study shows formally that it is possible for an agent to learn a generalizable strategy from passive data, as long as the agent can intervene at test time. Then, the experimental results show empirically that agents trained on expert data can indeed generalize at test time to infer and use causal links which are never present in the training data.

Strengths

* The paper addresses an interesting and important problem * The paper shows that natural language explanations support generalizing causal strategies from passive data * The findings of the paper have several applications, including using tools with LLMs * The paper is well written and easy to follow

Weaknesses

* It would be interesting to explore more complex environments

Questions

None

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

The paper should describe limitations more explicitly in a dedicated paragraph

Reviewer KP4B2023-08-11

Thanks for you response

I have read the response and decide to keep my current score. I advocate for the acceptance of the paper

Reviewer X6cj2023-08-12

Thanks for the additional experiments.

This is great. It's nice to see the additional DAG and LLM experiments. I appreciate the explanation about how heuristics would perform in the odd-one-out experiments, and I think future readers will benefit from it as well. I continue to recommend acceptance.

Reviewer U5ax2023-08-13

Thank you for the response.

I appreciate the changes and would like to keep my score.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC