Active Reasoning in an Open-World Environment

Recent advances in vision-language learning have achieved notable success on complete-information question-answering datasets through the integration of extensive world knowledge. Yet, most models operate passively, responding to questions based on pre-stored knowledge. In stark contrast, humans possess the ability to actively explore, accumulate, and reason using both newfound and existing information to tackle incomplete-information questions. In response to this gap, we introduce $Conan$, an interactive open-world environment devised for the assessment of active reasoning. $Conan$ facilitates active exploration and promotes multi-round abductive inference, reminiscent of rich, open-world settings like Minecraft. Diverging from previous works that lean primarily on single-round deduction via instruction following, $Conan$ compels agents to actively interact with their surroundings, amalgamating new evidence with prior knowledge to elucidate events from incomplete observations. Our analysis on $Conan$ underscores the shortcomings of contemporary state-of-the-art models in active exploration and understanding complex scenarios. Additionally, we explore Abduction from Deduction, where agents harness Bayesian rules to recast the challenge of abduction as a deductive process. Through $Conan$, we aim to galvanize advancements in active reasoning and set the stage for the next generation of artificial intelligence agents adept at dynamically engaging in environments.

Paper

Similar papers

Peer review

Reviewer yGBc3/10 · confidence 4/52023-07-06

Summary

This paper introduces Conan, a new benchmark for evaluating active reasoning in an embodied/open-world environment where an agent acts as a detective and must answer questions about the actions and intentions with the traces of another agent (a vandal). The experiments benchmark a RL + vision-and-language baseline and a more structured model that uses Bayesian inference on the Conan task.

Strengths

- ambitious benchmark trying to tackle an exciting and interesting problem that highlights a limitation in the current focus of the community: lack of benchmarks testing reasoning/question answering with active information gathering - problem and setting are well-motivated - interesting setup for the active reasoning problem with a vandal/detective

Weaknesses

- many important details that are not clear. One of the main things that needs more clarity is what the format of the benchmark is - it would be very helpful to have more examples and earlier in the paper of what the traces look like / what context is available to models, or maybe an example of a series of reasoning steps a model (or human) would go through to answer a question in Conan. See questions below for specific points of confusion - I wasn't totally convinced/clear that the questions in Conan are focused on testing abductive reasoning. Although eventually we would want to solve this with end-to-end methods like the RL baselines presented here, I feel like some heuristic exploration baselines would be important to understand what it takes to solve the benchmark and contextualize the scores of the RL agents (what scores do these agents get?): - the ideal explorer (specifically wondering: is the ideal explorer the ideal policy?) - an agent that explores the whole 64x64 grid (how much better does the ideal explorer do compared to this?) - using an oracle extractor that extracts the trace-relevant frames instead of the every-k key-frame extractor - instead of using RecurrentPPO, some heuristic extractor that allows agents to condition on key frames in the past (e.g. the frames with traces that are closely associated with the question) - Related to previous point, the benchmark also conflates *memory* and *exploration* with the abductive reasoning challenges -- both of these are already huge challenges for RL/current models. I worry that these may make the benchmark too difficult, and additionally not actually test for the reasoning abilities that it claims to test. Overall, I just don't have a good understanding of what solving the benchmark actually entails - if we have agents with good memory + exploration, is the reasoning involved actually quite simple? I think heuristic baselines/agents with privileged information would help the case here, as well as simplifying aspects of the task that are orthogonal to abductive reasoning

Questions

- table 1: what's the difference between embodied and open-world? - l146: what does a trace look like? are objects left behind, does the state change, etc.?. l151: when you say "32 distinctive traces" do you actually mean "32 object interactions" and traces can consists of multiple object interactions? - l153: "60 abductive reasoning tasks" - does this mean 60 question templates? - l164-168: what's an example of a path + alternative and why exactly does this require the detective to gather additional information? - why is the detective allowed to take the same actions as the vandal instead of just exploring the scene and collecting information? (feels like potentially unnecessary complexity) - l226-227: "masks are first employed to encode questions..." - what does this sentence mean? - How does the explorer condition on the question (e.g. some sentence embedding of the question or something else?) - Is there a step limit on the explorer's episodes? - l174-177: this made it sound like the detective and the agent were separate. My understanding is that the agent *is* the detective? - l255: what is the representation of symbolic map? - l307: to clarify, does the ideal explorer get the frames of the ground truth vandal trajectory, and then we extract every 30 frames? how much of the gap in performance is due to the keyframe extractor dropping important frames vs. the inaccuracy of the reasoning model? - sec 4.4: does this assume access to the vandal policy pi? Is that realistic? - l345: why would vanilla-trans be more susceptible to insufficient evidence? Isn't a more likely explanation that it's not pretrained to do question answering, in contrast to the other models? - AfD doesn't improve on standard (small improvements at best / mixed results across the board) - doesn't this suggest that this kind of reasoning is not super important in the benchmark?

Rating

3: Reject: For instance, a paper with technical flaws, weak evaluation, inadequate reproducibility and incompletely addressed 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

2 fair

Presentation

2 fair

Contribution

2 fair

Limitations

Yes

Reviewer 2rwT5/10 · confidence 4/52023-07-06

Summary

The paper introduces Conan, an interactive open-world environment for evaluating the active reasoning abilities of agents. In Conan, agents need to answer questions by actively seeking for evidence and acquiring new knowledge in a setting of incomplete information. Conan is formulated as a detective game. First, during the initialization of the game, a vandal agent completes a task, leaving traces in the environment. Next, questions are generated for a detective agent to answer. The detective has to answer these questions by actively interacting with the environment to reconstruct the actions of the vandal.

Strengths

* The environment differs from existing benchmarks for its active reasoning and interactive multi-round setting * The paper is very well written and easy to follow * Conan is an interesting environment that could spur further research on active reasoning in interactive contexts

Weaknesses

* It is not clear how challenging the task is and how good the performance of the models reported in the paper is. A human study would help the reader understand this critical point. * A qualitative analysis of the games and the behavior of the different agents would also be needed in the main paper to understand when the agents make mistakes and to what extent the tasks are solvable * As mentioned in the paper, it looks like the explorer does not supply particularly relevant information to the reasoning model, substantiating the intuition that the bottleneck is there, rather than in the reasoning abilities of the model

Questions

Is a human study or an error analysis feasible?

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

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

4 excellent

Contribution

2 fair

Limitations

The authors described the limitations of the paper

Reviewer nDEB7/10 · confidence 4/52023-07-07

Summary

This paper proposes a benchmark for "active reasoning" titled Conan, where, instead of passively answering questions from provided context, an agent must interactively explore its environment to discover information. The authors differentiate such a task from so-called "passive reasoning" tasks such as video-language understanding where the visual input is directly fed into the model. Here, a model itself is responsible for exploring and attaining input that will help a model answer a question. Conan is implemented in a Minecraft-like gridworld where a "detective" agent must identify activities that a rule-based, goal-oriented "vandal" agent completed, by using traces of the vandals' behavior in the environment. E.g. if a vandal's goal was to make a wooden pickaxe, it may have cut some trees down, and the task for the model will be to answer why the vandal decided to make a wooden pickaxe. Two alternative methods are adopted to approach this task: one which uses an explorer agent, trained with an exploration reward, to generate evidence that is fed into a standard VLM, and another, "Abduction from Deduction", where one directly learns to predict the goal of the vandal from an inferred state trajectory. Overall this is an interesting paper, and I think the dataset and benchmark will be valuable for the community. The models tested for this environment are relatively simple, not really making full use of the "joint exploration and reasoning" abilities that Conan purportedly tests, however. Moreover I do have some outstanding questions, and some issues with the experiments (specifically a missing baseline) that prevent me from assigning a higher score. However I'm open to changing my score after the author response and discussion period.

Strengths

- To my knowledge, the "active reasoning" component of Conan is an important area in vision/lang/RL, and is certainly underexplored in current embodied QA settings (although it is probably implicitly present in instruction following benchmarks and work like SayCan). - The dataset seems high quality and should be a useful contribution for the field. - Fairly sound experimental evaluation, comparing a variety of RL explorer agents with a variety of vision-language models.

Weaknesses

- **Missing negative control baselines.** The experimental results should compare evidence gathered by the trained explorer agents to a weaker negative baseline, e.g. a random untrained explorer policy or no explorer input at all. This is needed to convincingly show that the explorer agents are actually learning to explore in a way that is more beneficial than chance, at least for the downstream QA task. - In general I'm definitely more interested in to what extent performance is bottlenecked by good exploration on this task, rather than fixing the explorer evidence and trying subtly-differently-trained large VLMs. - As mentioned by the in the conclusion (L376-377), one of the key promises of Conan is that strongest performance on this benchmark should intuitively be achieved by models that jointly learn to reason and explore, but right now a decoupled two-stage process is adopted—first training an exploration model with an oracle "find trace" reward, extracting relevant keyframes, then training a VLM on top of the frozen policy. As a result, models trained for this task (for now) look like typical embodied QA models doing passive reasoning (just with the extra step of training an explorer to get the vision input). Of course, it's not strictly necessary for a benchmark paper to include an strong novel model for the benchmark, but without it we aren't able to evaluate whether Conan performance numbers have **headroom** for more sophisticated approaches to joint exploration and reasoning. (It seems like this should be the case, but we can't really be sure). - Some more outstanding questions (see next section)

Questions

- The idea of answering questions via "traces" left behind in the environment is a little hard to grasp and could use more explaining. There are some example traces in Figure 1b but these are small figures and aren't explained that much. The authors could spend some more time walking through example "traces" and more fully convince readers that the traces indeed provide answers to some of the more subtle questions in the benchmark (e.g. how do traces indicate that the vandals goal is to "get the diamond", couldn't the vandal have alternative goals like get iron/coal?). - It might be valuable to have an "oracle" setting to demonstrate that with perfect trace evidence and enough training models can 100% this task—or else explain why such an oracle doesn't exist (e.g. what causes the lagging performance even for the "Ideal Explorer" setting in Table 3? Is it an amount of training data issue? Are there still limitations as to the quality of the trajectory attained even by the ideal explorer, and is there an even more suitable oracle for this?) - Related to this, can authors further explain L308 ideal explorer: "visible to the ground-truth vandal’s trajectory"? What does "visible to" mean—shouldn't the ideal explorer literally *be* the ground-truth vandal's trajectory? - More details could be provided on the specific kinds of generalization splits tested here. At test time, is there a subset of abductive reasoning tasks held out that the model has never seen? Or does the model see the same questions it has seen during training, just in unseen environments? Do the test environments differ systematically from those during train (or is it possible to induce such a ) - Evaluating (and supporting) such compositional/systematic generalization splits would greatly increase the appeal of the benchmark. - (Related to the above) Do authors have an intuition as to whether test-time performance drop is more due to the VLM's inability to generalize, or the explorer's inability to generalize and generate good evidence for unseen enviroments (if test does indeed have unseen environments)? Is it possible to disentangle the two using different evaluation splits in the environment? - In Figure 4, does a score of near 100 as attained by eg TRPO indicate that the explorer indeeds find all traces (+100) reward all of the time? If not, is there an explicit measure of that?

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

4 excellent

Contribution

4 excellent

Limitations

Yes

Reviewer i6nS5/10 · confidence 2/52023-07-22

Summary

To address the gap in handling incomplete-information questions, this paper introduces an interactive open-world environment called "Conan." The purpose of Conan is to motivate and evaluate agents' active reasoning ability by requiring them to explore, gather evidence, and combine knowledge to solve complex scenarios. The ultimate goal is to improve AI agents' ability to interact with and learn from the world.

Strengths

This paper introduces a new interactive open-world environment called "Conan." The purpose of Conan is to motivate and evaluate agents' active reasoning ability by requiring them to explore, gather evidence, and combine knowledge to solve complex scenarios. The ultimate goal is to improve AI agents' ability to interact with and learn from the world.

Weaknesses

The author claims that the detective spawned in the environment needs to answer all questions through exploration, connecting key frames, and reaching conclusions. Yet, some questions presented in the Figures/Tables can be answered without exploring the environment. The process of answering questions through exploration requires more clarity and elaboration. Regarding the reward function used in the task, the author mentions that it incentivizes the agent to search for clues and traces related to the given question. However, the paper lacks detailed information about how the reward is designed and the specific motivation behind it. The writing appears to be cumbersome and illogical, making it difficult to grasp the actual meaning of the task. The purpose of introducing this new task is not clearly stated, leaving readers with questions about its significance and potential applications. In conclusion, while the paper introduces an intriguing concept of an interactive open-world environment for AI agents, there are aspects that require further clarification and refinement. Addressing the limitations and providing more detailed explanations about the task and reward design would enhance the overall understanding and impact of this research.

Questions

See Weaknesses

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

2 fair

Presentation

2 fair

Contribution

3 good

Limitations

See Weaknesses

Reviewer u9zC6/10 · confidence 5/52023-07-26

Summary

This paper introduces Conan, an interactive environment as a benchmark to evaluate agent’s active abductive reasoning abilities to answer questions in an incomplete (or partial) information scenario. Because of partial information, the model requires further exploration in the scene to answer the questions which is posed as a detective game. The paper also proposes the Abduction from Deduction (AfD) approach which relies on Bayesian statistics. The experimental evaluation showcases the strengths and weaknesses of various machine learning models in different settings.

Strengths

- The paper is well-motivated and clearly written in most parts. - The authors promise the availability of corresponding code and a detailed description of hyperparameters which would be essential for reproducibility if open-sourced. - The paper introduces the AfD approach using Bayesian rules which is interesting to the research community. - Relevant baselines have been selected to comprehensively evaluate various RL and multi-modal models.

Weaknesses

- The proposed environment looks like a toy benchmark with limited real-life applications. - Dataset statistics is missing in the work which might expose data biases related to goal, intents or survival questions. - The question templates are also limited and it is not clear if the model learns to take advantage of any data biases to answer the questions. It is also not clear if the questions are asked in a particular order. - This is framed as a multi-choice QA task with limited generalizability. - The proposed approach lacks a close integration between exploration and reasoning processes where the training and evaluation for different models is independent.

Questions

- Can you elaborate on how this research can be improved to integrate exploration and reasoning processes in future work? - What was the main motivation behind maintaining the detective set invincible, ie. not maintaining the survival status on line 178? - On line 363, the authors mention that the proposed approach is akin to “human-like reasoning”. It is not clear as to how solving this toy task with an independent explorer and reasoning model brings us closer to “human intelligence” Suggestions: - It would help to make the best results in bold in Table 3.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, 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

3 good

Presentation

4 excellent

Contribution

3 good

Limitations

See weaknesses for more information about the limitations

Reviewer 2rwT2023-08-15

Thanks

I appreciate the response of the author to address my comment. I increased my score by 1 point.

Authorsrebuttal2023-08-16

Response to Reviewer 2rwT

We greatly appreciate your follow-up. We would also like to share some preliminary results from our human study. We recruited 10 participants from our subject pool and trained them within the proposed Conan environment. For the "goal" split, they averaged 110.0 exploration steps per question and answered 90.9% of the questions correctly. These findings indicate that humans exhibit more efficient exploration and better reasoning capabilities. We will include further details about this aspect in the revised version.

Reviewer nDEB2023-08-15

Thanks

Thanks to authors for their detailed rebuttal. I appreciate the follow-up experiments and negative control baselines which verify that the models are indeed improving by using the image input. However I do think the random explorer baseline is also important, as it's possible to get some evidence by just randomly exploring (apologies if I did not sufficiently emphasize this in my initial review). Regardless, I think the other responses to the rebuttal also increase my confidence in the paper—I'll increase my score to a 7.

Authorsrebuttal2023-08-16

Response to Reviewer nDEB

Thank you very much for your follow-up! We are pleased to provide some preliminary results while training our models with the random explorer. Also on the "goal" split, we have observed the following accuracy rates: Vanilla-Trans 25.9%, F-BiLM-BERT 35.8%, F-BiLM-DeBERTa 41.6%, and Flamingo-mini 38.0%. These results show that the performance with the random explorer is notably improved compared to scenarios with empty visual inputs. Interestingly, we also observe that while the random explorer greatly falls short of the ideal explorer, its performance is only slightly lower than that of the TRPO explorer. Notably, our agents are initialized at the starting point of the traces, which indicates that they can collect some traces around them through random exploration. This observation suggests that agents do indeed benefit from the presence of traces, although they may be collected randomly. Besides, while the trained TRPO explorer exhibits certain effectiveness, it still has a long way to go.

Reviewer nDEB2023-08-16

Thanks

It's great to see these follow up experiments! It does seem like random exploration is doing well and that there's more headroom to grow from TRPO. These results increase my confidence in the paper and I would encourage their inclusion in the paper.

Authorsrebuttal2023-08-17

Thank you for your follow-up! We will include these results in the paper. We are glad to have more discussions in case you have any further concerns.

Reviewer u9zC2023-08-17

I have read the author's response and fellow reviewer's feedback. I believe that the authors have addressed most of my concerns. Particularly, I also liked the point raised by Reviewer nDEB and the author’s experiments related to the random baselines.

Authorsrebuttal2023-08-17

Dear Reviewer yGBc, We wanted to confirm whether the concerns and questions you raised in your initial review have been successfully addressed in our responses and revisions. We truly value your insights and feedback. Please feel free to post additional questions and comments about our work during the author-reviewer discussion period.

Area Chair QC9B2023-08-20

Thanks for your response -- AC Comment

Thanks for providing your responses to the reviewers' comments. They are comprehensive and answer many of the concerns raised during the initial review phase. The discussion with other reviewers also provides additional context for the questions raised during the initial review. -- Your AC

Authorsrebuttal2023-08-20

We would like to express our gratitude to the AC and all reviewers for their efforts in evaluating our work. Your feedback and suggestions have played a pivotal role in improving our work. We are glad to address any future concerns or questions that may arise.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC