acceptodds
Under review as a conference paper at ICLR 2027

Predicting Decisions of AI Agents from Limited Interaction through Text-Tabular Modeling

Abstract

AI agents increasingly interact with unfamiliar black-box counterparts whose behavior emerges from language models, prompts, memory, and control logic. We study whether a counterpart's next decision can be predicted from a small set of observed interactions, without access to its implementation. We cast this problem as target-adaptive text-tabular inference, combining behavioral examples from the specific target with the current strategic state, the dialogue, and a representation from a frozen language model Observer. Across bargaining and negotiation environments and a diverse population of independently built agents, adaptation to the target improves prediction in ways that additional population data cannot reproduce. Observer representations provide complementary information beyond explicit state and generic text features, but their value depends on what the structured interaction history already captures. These findings separate learning the behavior of a particular agent from representing the strategic situation, and provide a practical framework for predicting unfamiliar agents from limited interaction.

Primary area:
interpretability and explainable AI
Keywords:
opponent modeling, behavioral prediction, LLM representations, AI agents, text-tabular learning

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_bderOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper combines labeled source-agent decisions, a few target-agent games, structured states, dialogue, and frozen LLM Observer representations in TabPFN to predict black-box agents’ next decisions without task-specific gradient updates. Bargaining and negotiation experiments show that correctly labeled target games improve prediction beyond additional source data.

Critical strengths
  1. The study evaluates transfer from source LLM agents to independently scaffolded target agents with different prompts and control logic, providing a useful test of unfamiliar-agent prediction.

  2. The wrong-target, shuffled-label, and extra-source controls support the claim that prediction benefits from correctly labeled behavior of the particular target. These experiments help distinguish target-specific adaptation from simply increasing the amount of conditioning data.

  3. Comparisons with generic text encoders, Observer output probabilities, and controls for structured states and interaction trajectories help explain where hidden representations provide additional value.

Critical weaknesses
  1. Although Section 4 explicitly constructs inputs from the deciding agent’s observable view, this includes the target’s own private payoff parameters, which an external counterpart may not observe. The paper should clarify the intended forecasting setting and either evaluate performance using only counterpart-observable inputs or qualify its practical claims accordingly.

  2. The encoder and input-decomposition controls are useful, but generic encoders and the Observer still receive different inputs. Comparing them on the same serialized state and dialogue would better isolate the Observer representation’s value.

  3. TabPFN receives up to 3,000 labeled source decisions unavailable to the direct LLM baseline, so their performance gap does not isolate predictor choice. A target-only TabPFN control would help.

Questions for authors
  1. How do performance and Observer gains change when target-private parameters and their derived features are removed from both structured inputs and model prompts? Please report complete- and incomplete-information results separately.

  2. Table 2 reports median R², but Appendix M defines error bars using σ/√N over target–seed pairs. Please provide paired confidence intervals for Observer gains that account for repeated evaluations of each target and, for proposals, recompute the median R² difference during resampling.

  3. Please specify the information, history window, and pre-decision message cutoff available to each representation, and compare a generic encoder with the Observer on identical serialized inputs.

  4. Please compare against target-only TabPFN and give the direct LLM baseline a manageable set of relevant source examples to assess the effect of unequal source supervision.

6 comments
AuthorsOpenReview ↗

Response to Reviewer bder (1/3)

Thank you for the focused review. We ran all four requested analyses:

  • Q1: results with both players' private parameters removed from every input, by information regime.
  • Q2: paired cluster-bootstrap CIs, with the median-R² difference recomputed in each resample.
  • Q3: a generic encoder on the Observer's exact prompt, and a specification of what each representation sees.
  • Q4: a direct LLM with the same source examples as TabPFN, and target-only TabPFN.

W1 / Q1. "How do performance and Observer gains change when target-private parameters and their derived features are removed from both structured inputs and model prompts? Please report complete- and incomplete-information results separately."

Response. This is an important question. In bargaining, the Observer gain persists with public information alone. In incomplete-information negotiation, the estimated gain is smaller when the target’s own value is hidden. We will report both estimates in the abstract.

In §4.1 we will define the forecasting setting through two input views that differ in whether the forecaster knows the target’s own parameter:

  • Target view, as described in §4.1 of the submission: the target's own parameter is kept, and the opponent's private parameter is masked. This is the view of a forecaster who knows the target's own parameter, such as the platform that runs the interaction or the party that deployed the target.
  • Public view. Both players’ private parameters are removed from all inputs, and the Observer representation is extracted from the masked prompt. This is stricter than a counterpart’s view, since each player knows its own parameter. It captures the situation of a third party trying to predict one player’s actions.

Observer gain over game+text, by information regime: the range over K, and game+text → +Observer AUC at K=16.

RegimeTarget viewK=16Public viewK=16
Bargaining, incomplete information+.027 to +.035.821 → .848+.025 to +.035.823 → .848
Bargaining, complete information+.022 to +.041.787 → .828+.018 to +.037.793 → .830
Negotiation, incomplete information+.021 to +.030.762 → .792+.007 to +.011.781 → .793
Negotiation, complete information−.013 to −.004.837 → .833−.024 to +.005.817 → .822

Direct LLM under the public view. We reran Gemini 3.5 Flash-Lite without thinking on the same test decisions. In negotiation, hiding both valuations in incomplete-information games changed the LLM’s AUC from .502 to .464 at K=0 and from .734 to .718 at K=16. Public-view TabPFN with the Observer scored .723 and .835, respectively. In bargaining, the public-view and original LLM prompts were identical. The LLM scored .626 at K=0 and .737 at K=16, and public-view TabPFN with the Observer scored .691 and .854.

We will define both views in §4.1, add the public-view table to §6, and report the K=16 negotiation estimates for both views in the abstract: “In incomplete-information negotiation, after 16 target games, its estimated gain was .029 AUC when the forecaster knew the target’s valuation and .011 with public information alone.”

AuthorsOpenReview ↗

Response to Reviewer bder (2/3)

Q2. "Please provide paired confidence intervals for Observer gains that account for repeated evaluations of each target and, for proposals, recompute the median R² difference during resampling."

Response. We now report paired cluster-bootstrap CIs for all Observer gains. In bargaining, every CI excludes 0, for responses and for proposals. In negotiation, response gains are significant at K≥2 and proposal gains are not.

Responses. We use a cluster bootstrap over targets (5,000 resamples), in the target view, with Gemma-2-2B layers 16–23. Team-clustered intervals are similar. AUC, game+text → +Observer:

KBargainingGain [95% CI]NegotiationGain [95% CI]
0.640 → .668+.027 [.018, .037].812 → .820+.008 [−.002, .019]
2.693 → .724+.031 [.022, .041].809 → .828+.019 [.005, .036]
4.732 → .764+.032 [.023, .042].813 → .831+.018 [.006, .031]
8.765 → .801+.036 [.026, .047].807 → .828+.021 [.007, .036]
16.808 → .840+.033 [.024, .042].847 → .866+.019 [.005, .038]

Proposals (Table 2, proposal panel). The submitted runs split at the decision level. We recomputed them with game-level splits, so no conditioning game appears in the test set in any of the 2,395 cells. We used:

  • updated features, in the target view;
  • PCA fit on conditioning rows only;
  • the median-R² difference recomputed in each resample, as you requested.

All runs use Gemma-2-2B, layer 20. Median R², game+text → +Observer:

KBargainingGain [95% CI]NegotiationGain [95% CI]
0.307 → .501+.194 [.087, .313].542 → .557+.015 [−.094, .117]
2.410 → .586+.176 [.055, .264].600 → .654+.054 [−.039, .132]
4.534 → .634+.100 [.043, .208].627 → .651+.024 [−.041, .094]
8.612 → .695+.083 [.039, .152].728 → .760+.031 [−.043, .114]
16.658 → .720+.061 [.032, .145].789 → .794+.005 [−.006, .065]

The bargaining proposal gain holds at every K under the stricter protocol. The negotiation gain is not distinguishable from 0, as the paper states.

We also compared against the paper's direct LLM on identical test decisions, with both methods selecting context games using the same split function and seed. At K=16 in bargaining, TabPFN+Observer reaches a median R² of .800, against −.303 for the LLM. All CIs exclude 0, in both games and at every K.

We will add the paired confidence intervals for Observer gains to the paper and update the proposal panel of Table 2.

AuthorsOpenReview ↗

Response to Reviewer bder (3/3)

W2 / Q3. "Please specify the information, history window, and pre-decision message cutoff available to each representation, and compare a generic encoder with the Observer on identical serialized inputs."

Response. We added a representation from a stronger generic encoder (bge-m3) on exactly the same prompt. It adds nothing to the base configuration, and the Observer retains its gain. The table gives results after K=16 target games for bargaining / negotiation.

Added toNone+ bge-m3+ Observer
Game features + identity.815 / .852.818 / .850.848 / .869
Game features + text + identity.807 / .849.816 / .848.840 / .864

Information available to each representation:

  • Game features: configuration, the target's own parameter, the current offer, and the last five rounds.
  • Dialogue embedding: current-round messages sent before the target's decision.
  • Observer and bge-m3: the reconstructed decision-time prompt, covering the full game up to the decision.

We will add the comparison and the description of the information available to each representation to the paper.

W3 / Q4. "Please compare against target-only TabPFN and give the direct LLM baseline a manageable set of relevant source examples to assess the effect of unequal source supervision."

Response. A practical advantage of tabular models is that they can use many context rows at low cost. Providing the same amount of context to an LLM is expensive. We therefore tested the tabular model with the Observer when both methods receive the same small source sample. Our method leads when both systems receive one labeled decision from each of 16 source games and observe 16 target games.

We ran Gemini 3.5 Flash-Lite with K=0 and K=16 target games on the full target pools.

Matched source supervision. We gave both systems the same 16 labeled source decisions and the same K target games and evaluated them on the same test decisions. With no target games, the LLM is more accurate. After 16 target games, the tabular model with the Observer performs significantly better. The number of source labels alone therefore does not explain the advantage.

AUCBargaining K=0Bargaining K=16Negotiation K=0Negotiation K=16
Gemini 3.5 Flash-Lite.692.735.657.732
TabPFN, no Observer.624.821.582.804
TabPFN + Observer.608.837.560.806

In a separate comparison using the submitted input view, we gave TabPFN 16 target games and no source decisions. It reached .799 AUC in bargaining and .803 in negotiation without the Observer, and .817 and .808 with the Gemma-2-2B layer-20 Observer. Adding the full source pool brought the scores with the Observer to .840 and .864.

Why does the LLM not receive all source rows? With 16 labeled source decisions, the prompts averaged 5,904 input tokens. TabPFN can use up to 3,000 labeled source decisions per prediction. Extrapolating from the measured number of tokens added by 16 examples, the 3,000 examples alone would take about 590,000 input tokens per prediction in bargaining and 635,000 in negotiation. At 0.40 USD per million input tokens, we estimate about 4,700 USD in input costs for the 19,294 test predictions in this comparison (K=0 and K=16).

We will add the comparison with matched source supervision to the paper.

We hope these analyses address your concerns. If any point remains open, we are happy to clarify or run further checks during the discussion.

Reviewer_yi9GOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the problem of predicting the next decision of an unfamiliar black-box AI agent from a limited number of observed interactions. The authors formulate the task as target-adaptive text-tabular inference and propose a framework combining structured game-state features, dialogue representations, and hidden-state representations extracted from a frozen LLM Observer. A pretrained TabPFN model is then used to predict the target agent's acceptance decisions or future proposals. Experiments are conducted on bargaining and negotiation environments, with cross-population transfer from 13 frontier-LLM agents to 91 independently constructed hackathon agents. The results demonstrate that target-specific interaction histories improve prediction accuracy and that Observer hidden states provide complementary information beyond structured features and direct LLM outputs.

Critical strengths
  1. Predicting the behavior of unfamiliar black-box agents from limited interaction is an important problem in multi-agent systems, especially as LLM-based agents increasingly interact with independently developed counterparts.

  2. The cross-population evaluation and the controls involving shuffled labels, incorrect target identities, and additional source data provide convincing evidence that target-specific behavioral observations contribute meaningful predictive information.

Critical weaknesses
  1. The Observer evaluation is restricted to three relatively small LLMs (1B–2B), while target agents primarily rely on the Gemini 2.5 Flash family. Broader evaluations across model scales and architectures are needed to establish the robustness and generality of the findings.

  2. The main experiments focus on bargaining and negotiation, which share similar alternating-offer protocols and structured state representations. The preliminary persuasion experiment is insufficient to demonstrate broader applicability. Evaluation in more diverse social simulation and agent decision-making settings would strengthen the conclusions.

Questions for authors
  1. How well does the proposed method generalize to larger LLMs and more diverse agent architectures?

  2. Can the framework generalize beyond structured bargaining and negotiation tasks to broader social or behavioral decision-making scenarios?

  3. What is the main practical or scientific value of the framework? Does it reveal meaningful agent decision-making mechanisms beyond nonlinear feature representations?

1 comment
AuthorsOpenReview ↗

Response to Reviewer yi9G

Thank you for recognizing the importance of the problem and the value of the cross-population setting. We ran five larger Observers in response to W1, and we answer Q2 and Q3 directly below.

W1 / Q1. "The Observer evaluation is restricted to three relatively small LLMs (1B–2B) […] How well does the proposed method generalize to larger LLMs and more diverse agent architectures?"

Response. We added five larger Observers, up to 27B parameters. In bargaining, four of the five give a significantly larger gain than their smaller counterpart at every K. The fifth performs the same. In negotiation, the improvement is limited to small K.

All Observers read the same prompt, at layers of the same relative depth, next to the same game+text features. AUC:

FeaturesBargaining K=0Bargaining K=16Negotiation K=0Negotiation K=16
Game+text.625.807.782.849
+ Gemma-2 2B.649.839.791.865
+ Gemma-2 9B.674.864.817.870
+ Gemma-2 27B.697.875.805.866
+ Qwen3 1.7B.637.835.811.875
+ Qwen3 8B.636.835.821.875
+ Qwen3 14B.717.877.830.881
+ Llama-3.2 1B.631.824.778.855
+ Llama-3.1 8B.648.847.793.858

We will add the results for every K and the paired confidence intervals to the paper.

W1 (cont.). "…while target agents primarily rely on the Gemini 2.5 Flash family."

Response. Yes. We will note in §8 that the targets vary in the scaffolding they build around one LLM family. The dataset of decisions by 91 agents developed for the same tasks is itself a distinct contribution. Additional axes of variation would strengthen the evaluation. We are unaware of such data for any task and would gladly test on it if available.

W2 / Q2. "Can the framework generalize beyond structured bargaining and negotiation tasks to broader social or behavioral decision-making scenarios?"

Response. The framework can be applied when we can represent what an agent has observed up to a decision and specify the action to predict. Our evaluation covers structured two-player economic games. As you point out, the persuasion experiment shows that the same pipeline can predict decisions in a game without alternating offers, but its five agents come from a single population.

Whether the framework can predict an unfamiliar agent’s decisions in open-ended or multi-party interactions remains an open question. We will make that scope explicit in §8.

Q3. "What is the main practical or scientific value of the framework? Does it reveal meaningful agent decision-making mechanisms beyond nonlinear feature representations?"

Scientifically, the question is whether the way a language model represents a game situation can help predict how another agent will act in that situation. For example, an Observer that sees a negotiation offer and the conversation leading up to it might itself reject the offer. We set aside the action it would choose and use the internal representation it formed of the situation. A tabular model combines that representation with game features and generalizes from decisions agents made in past situations to an agent’s decision in a new one. In our experiments, the Observer representation improved prediction even relative to a generic text encoder given exactly the same input.

Practically, this lets us draw on experience from interactions with many agents to predict how a new agent will behave after observing only a few of its games. We could evaluate new agents more quickly and use these forecasts to choose a more accurate best response to an opponent, helping us develop stronger agents or more stable economic mechanisms.

We hope these results address your concerns. If any point remains open, we are happy to provide further analyses during the discussion.

Reviewer_p2KmOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
3(+1)
Read review
Paper summary

This paper studies the prediction of unfamiliar black-box AI agents' decisions from limited interaction histories. The authors propose a target-adaptive text-tabular framework that combines structured game-state features, dialogue representations, and hidden states extracted from a frozen LLM Observer, using TabPFN for in-context prediction. Experiments on bargaining and negotiation tasks demonstrate that incorporating a few target-specific interaction histories substantially improves prediction accuracy, while Observer representations provide additional benefits beyond explicit state features and direct LLM predictions. The paper further investigates the contributions of target-specific adaptation and Observer representations through controlled experiments.

Critical strengths
  1. Meaningful problem setting. The paper investigates an interesting and practically relevant problem of predicting unfamiliar black-box agents from limited interactions. The cross-population evaluation involving independently developed agents provides a useful setting for studying target-specific behavioral adaptation.

  2. Relatively comprehensive experimental analysis. The paper includes well-designed control experiments that distinguish target-specific adaptation from simply adding more data, and investigate the contributions of Observer hidden states beyond explicit features and model outputs. The findings provide interesting insights into the roles of behavioral adaptation and state representation.

Critical weaknesses
  1. Insufficient comparisons with alternative prediction methods. The proposed framework relies heavily on TabPFN, but lacks comparisons with conventional tabular predictors such as XGBoost or CatBoost, as well as alternative representation learning approaches. Additionally, the main direct LLM baseline uses a zero-thinking configuration, which may limit the strength of this comparison.

  2. Limited generalizability of the evaluation. The main experiments focus on bargaining and negotiation tasks that share similar alternating-offer interaction structures. Moreover, the target agents are developed within a common competition setting. It remains unclear whether the findings generalize to more diverse agent architectures, interaction protocols, and decision-making environments.

Questions for authors
  1. Alternative prediction baselines. Could the authors compare TabPFN with alternative tabular predictors, such as XGBoost or CatBoost, under the same target-adaptation setting to better establish the necessity of the chosen prediction framework?

  2. Generalization beyond alternating-offer games. Could the authors provide additional evidence of generalization to environments with different interaction protocols, given that both main evaluation tasks share a similar alternating-offer structure?

3 comments
AuthorsOpenReview ↗

Response to Reviewer p2Km (1/2)

Thank you for the positive assessment, and for highlighting the cross-population setting. We ran the comparisons you asked for: seven alternative predictors, a same-input representation comparison, and a direct LLM with thinking.

W1 / Q1. "Could the authors compare TabPFN with alternative tabular predictors, such as XGBoost or CatBoost, under the same target-adaptation setting to better establish the necessity of the chosen prediction framework?"

Response. We ran seven alternative predictors under the same target adaptation. The framework works with other tabular predictors. Kumo Tabular, released in October 2026 and currently the strongest tabular model on the TabArena leaderboard, and TabICLv2 perform comparably to or better than TabPFN. In bargaining, the tree ensembles trail the tabular foundation models once target games are available.

All predictors receive the same rows and the same K target games, and are scored on the same test decisions. AUC after 16 target games, game+text → +Observer (Observer layer 20):

PredictorBargainingNegotiation
TabPFN (paper).807 → .840.849 → .864
Kumo Tabular (large).827 → .857.846 → .860
TabICLv2.805 → .851.846 → .851
TabDPT.765 → .826.810 → .804
CatBoost.769 → .801.837 → .843
XGBoost.764 → .791.844 → .832
Random forest.779 → .814.835 → .842
Logistic regression.599 → .709.813 → .810

Across all K:

  • Kumo Tabular is higher than TabPFN in bargaining at every K (K=16: .827 vs .807 without the Observer, .857 vs .840 with it). In negotiation the two do not differ significantly.
  • TabICLv2 stays within ±.02 of TabPFN.
  • CatBoost, XGBoost and random forest (default hyperparameters) also improve with target games. In bargaining, from K=0 to K=16: XGBoost .620 → .764, random forest .604 → .779, TabPFN .625 → .807. They are lower than TabPFN by .02 to .07 at K≥2, most at K=4–8 (K=8: XGBoost .684 vs .746). In negotiation without the Observer they are close to TabPFN at K=16 (XGBoost .844 vs .849).
  • Observer across predictors. In bargaining it improves every predictor at every K. At K=16 the gain ranges from +.027 (XGBoost, .764 → .791) to +.110 (logistic regression, .599 → .709). In negotiation the gain is significant for Kumo Tabular at every K (K=16: .846 → .860) and for TabPFN at K=2–8 (K=8: .793 → .812).

The bargaining results therefore hold across predictors. In negotiation, the Observer gain appears mainly with TabPFN and Kumo Tabular.

We will add the results for every predictor at every K to the paper.

W1 (cont.). "…as well as alternative representation learning approaches."

Response. We added two comparisons. A generic encoder on the Observer's exact input adds ≈0, and sequence models that learn their own representation trail the tabular learners after 16 target games.

  • Generic encoder on the identical input. A strong generic encoder (bge-m3) reads the same prompt as the Observer, at the same dimensionality. Next to game features and agent identity at K=16, AUC is .815 / .852 with no encoder, .818 / .850 with bge-m3 and .848 / .869 with the Observer. Across K, the Observer exceeds bge-m3 by +.031 to +.038 in bargaining and +.013 to +.023 in negotiation.
  • Sequence models that learn their own representation. An adapted ToMnet and a TAO-inspired Trajectory Transformer, trained end to end on game trajectories, reach at most .700 in bargaining and .722 in negotiation at K=16 across the two input conditions (with and without the Observer). We selected hyperparameters on held-out source agents for each condition. Even without the Observer, every tree ensemble and tabular foundation model in the table above outperforms both trajectory models in both games.

We will add the comparisons with the encoder and the trajectory models to the paper.

AuthorsOpenReview ↗

Response to Reviewer p2Km (2/2)

W1 (cont.). "Additionally, the main direct LLM baseline uses a zero-thinking configuration, which may limit the strength of this comparison."

Response. We reran the direct LLM at the highest thinking level. Thinking helps the LLM, and the full method still leads after 16 target games.

The Gemini 2.5 Flash-Lite model we used for the paper had been retired, so we used its newer-generation counterpart, Gemini 3.5 Flash-Lite. We compared no thinking with the highest thinking setting at K=0 and K=16 on the full target pools. Every row in the table uses the public view. In incomplete-information games, both predictors see only public payoff information.

AUCBargaining K=0Bargaining K=16Negotiation K=0Negotiation K=16
Direct LLM, no thinking.626.737.464.718
Direct LLM, highest thinking.676.793.564.779
TabPFN, game+text.654.826.743.833
Ours (TabPFN + Observer).691.854.723.835
Ours − thinking LLM [95% CI]+.015 [−.013, .041]+.061 [.044, .080]+.159 [.088, .230]+.056 [.026, .088]

We will add the comparison with the direct LLM at its highest thinking setting to the paper.

W2 / Q2. "Could the authors provide additional evidence of generalization to environments with different interaction protocols, given that both main evaluation tasks share a similar alternating-offer structure?"

Response. In §7.4 and Appendix J, we report an experiment on a persuasion game with no alternating offers. We applied the same response-prediction pipeline to the five GLEE agents with sufficient data, holding out one agent at a time. Adding the Observer gave a small positive gain at every KK.

For bargaining and negotiation, the hackathon gave us a separate target population of 91 agent versions developed by 34 teams. For persuasion, we could test prediction on held-out agents within GLEE, but we had no independent target population on which to repeat the cross-population evaluation. We will make this distinction clearer in the paper.

We hope this addresses your concerns. If anything remains unclear, we are happy to clarify further during the discussion.

Thank you for the thoughtful response and the additional experiments. The new comparisons have addressed my main concerns, and I will raise my soundness score accordingly. I appreciate the clarification of the scope of the generalization results and remain supportive of acceptance.

Reviewer_UYa3OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
3(+1)
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies whether a counterpart’s next decision can be predicted from a small set of observed interactions, without access to its implementation. The author casts this problem as target-adaptive text-tabular inference, combining behavioral examples from the specific target with the current strategic state, the dialogue, and a representation from a frozen language model Observer.

The author claims its framing novelty from the conjunction: per-agent next-decision prediction, from limited interactions, in language-based bargaining and negotiation, under source-target population shift.

Critical strengths
  1. The motivation of this work is very clear and interesting. Overall, the writing is easy to understand and follow.
  2. The adaptation controls are well designed (Table 3). Source rows and the test set are held fixed while only the adaptation rows change.
Critical weaknesses
  1. No explicit reason why the problem should be formulated as a text-tabular prediction task. No in-context learner that keeps the trajectory structure in any other form is tried or explained. I only see justifications for why we need an in-context learner, but not why a tabular learner.
  2. The “language-based” element of the claimed novelty is thin. Language enters only as input (“free-text observations,” p. 2), where it contributes little (Table 8; §7.2). The predicted “decision” is restricted to numeric actions defined by the chosen games ( §4.1), so the agent’s language output is excluded without justification.
  3. The cited related works are not enough. Most recent works of in-context opponent modeling (e.g., TAO, ICLR 2024, OMIS, NeurIPS 2024) are missing, where they used transformers over trajectories to make predictions. Also, a more recent field of Behavior prediction of LLM/agent is missing as well (e.g., Centaur, Nature 2025).
  4. Why only compare with direct LLM prompting? The paper already cited predictors such as ToMnet (Rabinowitz et al. 2018) and Chawla et al. 2022 (negotiation-dialogue opponent modeling) (p. 2, §2).
Questions for authors

Q1 (W1). What motivates the tabular formulation specifically, beyond the need for an in-context learner? Did you compare against a learner that keeps the trajectory structure?

Q2 (W2). “Language-based” is part of the claimed novelty (p. 2), yet language enters only as input, and “decision” is restricted to numeric actions (p. 5, §4.1). Is this restriction a property of the problem or of the chosen games, and how does it limit the framing claim?

Q3 (W3, W4). Why is no published method (e.g., ToMnet, TAO, OMIS) compared? If they do not apply to this setting, please explain why.

Q4 (W4). LLM-as-Predictor does not see the up to 3,000 source rows available to the tabular pipeline (p. 5 vs. p. 6), and it is evaluated on at most 30 test decisions per cell versus 500 (p. 20). How does this mismatch affect the comparison?

I am happy to raise my scores if you can provide enough evidence regarding these questions.

4 comments
AuthorsOpenReview ↗

Response to Reviewer UYa3 (1/2)

Thank you for the constructive review and for stating what evidence would change your assessment. For each of Q1–Q4, we either ran a new experiment or propose a change to the paper.

  • Q1: two trajectory-preserving learners are now baselines; after 16 target games both trail every tree ensemble and tabular foundation model we tested, with or without the Observer.
  • Q2: We will narrow the contribution statement to action prediction.
  • Q3: We added ToMnet and a TAO-style model as baselines. We will discuss TAO, OMIS and Centaur in the related work.
  • Q4: the LLM comparison is now matched on source supervision and on test decisions.

Q1 (W1). "What motivates the tabular formulation specifically, beyond the need for an in-context learner? Did you compare against a learner that keeps the trajectory structure?"

Response. We added two baselines that retain the trajectory structure. At K=16, the Observer improves both trajectory models in negotiation. In bargaining, both score slightly lower with the Observer. In both games, the trajectory models trail every tree ensemble and tabular foundation model we tested, with and without the Observer.

Motivation. We have a large labeled source population and a few target games. Each decision is described by the same heterogeneous fields: game state, the last five rounds, text and Observer embeddings, and agent identity. Representing source and target decisions as rows of one table lets a pretrained in-context tabular predictor condition on both at low cost, without training a new model for each target. A trajectory model can also adapt in a single forward pass, so the choice is an empirical question. We tested it by implementing two trajectory models:

  • an adapted ToMnet (Rabinowitz et al., 2018), in which an LSTM encodes each past game;
  • a Trajectory Transformer inspired by TAO (Jing et al., 2024), which reads all games as one sequence with separators.

Both were meta-trained episodically on 11 source agents. We used two held-out source agents to select hyperparameters separately for each game and each model, with and without the Observer. We also ran tabular learners that have no pretraining, and gave the trajectory models the Observer state as an additional input. All predictors are scored on the same test decisions.

AUC after 16 target games, game+text → +Observer (Gemma-2-2B, layer 20):

LearnerBargainingNegotiation
Adapted ToMnet (trajectory).700 → .687.679 → .722
Trajectory Transformer (trajectory).691 → .682.691 → .717
XGBoost (tabular).764 → .791.844 → .832
CatBoost (tabular).769 → .801.837 → .843
Random forest (tabular).779 → .814.835 → .842
TabPFN (tabular).807 → .840.849 → .864

A separate experiment shows what the source population contributes. We held TabPFN with the Observer fixed and grew its source pool from 16 decisions to the full source population. With no target games, AUC rose from .608 to .691 in bargaining and from .560 to .723 in negotiation.

We will add the rationale for the tabular formulation and the comparison across learners at each K, with and without the Observer, to the paper.

AuthorsOpenReview ↗

Response to Reviewer UYa3 (2/2)

Q2 (W2). "“Language-based” is part of the claimed novelty (p. 2), yet language enters only as input, and “decision” is restricted to numeric actions (p. 5, §4.1). Is this restriction a property of the problem or of the chosen games, and how does it limit the framing claim?"

Response. The restriction comes from the games. Their outcomes are determined by accept/reject decisions and offers, so these are the actions we predict. We will narrow the claim in the paper accordingly.

Humans negotiate in natural language, and AI agents do the same when negotiating with people or with one another. The same numerical offer can be presented as final or as an invitation to negotiate further. We therefore use the dialogue alongside the game state to predict agents’ acceptance decisions and proposals. In our experiments, the game state and history account for most of the Observer’s predictive gain.

We will revise the contribution statement in §1 to read “few-game prediction of the actions of black-box agents that interact through natural language, cast as target-adaptive text-tabular inference.”

Q3 (W3, W4). "Why is no published method (e.g., ToMnet, TAO, OMIS) compared? If they do not apply to this setting, please explain why."

Response. We now compare with ToMnet and a TAO-style model (Q1). OMIS and Chawla et al. do not map directly onto next-decision prediction for the reasons below. In §2 we will discuss TAO, OMIS and Centaur alongside Chawla et al., which we already cite.

  • OMIS: its opponent imitator is a component of a search-based policy. Our trajectory baselines test the in-context conditioning idea; they do not reproduce OMIS.
  • Chawla et al. (2022): they infer a negotiator's issue priorities in multi-issue negotiation. That is a latent preference, not the next decision, and our games have a single issue.

We will add the following paragraph to §2: “In-context opponent modeling trains transformers on interaction trajectories to identify and respond to opponents (TAO and OMIS). Action prediction there is a component of a learned policy. We compare against trajectory models in this spirit. Population-level models such as Centaur predict human choices from large training corpora. We adapt to one black-box agent from a few games.”

Q4 (W4). "LLM-as-Predictor does not see the up to 3,000 source rows available to the tabular pipeline (p. 5 vs. p. 6), and it is evaluated on at most 30 test decisions per cell versus 500 (p. 20). How does this mismatch affect the comparison?"

Response. A practical advantage of tabular models is that they can use many context rows at low cost. Providing the same amount of context to an LLM is expensive. We therefore tested the tabular model with the Observer when both methods receive the same small source sample. Our method leads when both systems receive one labeled decision from each of 16 source games and observe 16 target games.

Why 30 test decisions per seed for the LLM versus 500 for the tabular model? The cost of the direct LLM prevented us from evaluating it on all test decisions for the submission. We re-scored TabPFN on exactly the test decisions that the submitted LLM answered. At K=16 our method with the Observer reaches .853 versus .770 in bargaining and .864 versus .785 in negotiation. It leads significantly at every K in both tasks.

Matched source supervision. We gave both systems the same 16 labeled source decisions and the same K target games and evaluated them on the same test decisions. With no target games, the LLM is more accurate. After 16 target games, the tabular model with the Observer performs significantly better. The number of source labels alone therefore does not explain the advantage.

AUCBargaining K=0Bargaining K=16Negotiation K=0Negotiation K=16
Gemini 3.5 Flash-Lite.692.735.657.732
TabPFN, no Observer.624.821.582.804
TabPFN + Observer.608.837.560.806

Why does the LLM not receive all source rows? With 16 labeled source decisions, the prompts averaged 5,904 input tokens. TabPFN can use up to 3,000 labeled source decisions per prediction. Extrapolating from the measured number of tokens added by 16 examples, the 3,000 examples alone would take about 590,000 input tokens per prediction in bargaining and 635,000 in negotiation. At 0.40 USD per million input tokens, we estimate about 4,700 USD in input costs for the 19,294 test predictions in this comparison (K=0 and K=16).

You wrote that you would be happy to raise your score given enough evidence on these questions. We hope the experiments above provide it, and we are glad to run further analyses during the discussion.

Thank you for the additional experiments. The trajectory baselines address the question directly, and I appreciate the effort. A few clarifications would help me interpret the comparison:

  1. For the Inputs to the trajectory models, did the adapted ToMnet and the Trajectory Transformer receive the same engineered per-step features as the tabular rows, or raw offers and messages? And at inference, did they see labeled source-population decisions in context, as TabPFN does with up to 3,000 rows, or only through their weights? Your source-pool experiment shows source rows raise zero-shot AUC substantially (.608 → .691 and .560 → .723). If only TabPFN conditions on them, the comparison mixes the learner with access to source data.
  2. Both trajectory models were meta-trained on 11 source agents. Could the gap reflect too few tasks for episodic meta-training rather than the trajectory representation itself?
  3. Please specify how the tree baselines (Random forest, XGBoost, and CatBoost) were trained: on which rows (source, target K-game, or both), and whether the identity column was included. Random forest reaches .779/.835, close to TabPFN and well above both trajectory models. If the tree models are fit on the same source and target rows as TabPFN while the trajectory models are not, the gap may reflect access to source data rather than the formulation. If all learners had equal access, the result would suggest the advantage comes mainly from the engineered row features rather than the in-context tabular learner. Which interpretation do you think holds?

Finally, can you summarize the takeaway on why the tabular formulation is a best fit for this problem? Beyond the empirical comparison, could you explain the design rationale for the tabular formulation? Specifically:

  1. which properties of this problem make a fixed-width row an adequate representation of a decision;
  2. what this representation gives up, given that history is truncated to five rounds, dialogue is reduced to five dimensions, and sequential structure within and across games is flattened;
  3. under what conditions you would expect the formulation to stop being appropriate.

Stating these assumptions in the paper would clarify when the approach applies.

Thank you for the quick follow-up and for the additional questions.

1. What information did each model receive? At each decision step, ToMnet and the Trajectory Transformer received exactly the same game-state features as the tabular models. Both trajectory models learned from the source agents during training, so information about those agents was encoded in their weights, as it was for the tree-based models. The trajectory models received the target-agent data as input at prediction time. The tabular foundation models (TFMs) received both source- and target-agent data in context, without any fine-tuning.

2. Are 11 source agents enough? Your concern that 11 source agents may be too few is reasonable. To test this, we trained predictors on the hackathon agents without the Observer. This expanded the source pool from 11 agents to 55–59 in bargaining and 34–35 in negotiation. We compared the expanded pool with a pool of 11 hackathon agents. Each cell reports AUC at K=16 for bargaining / negotiation.

Model11 hackathon agentsExpanded hackathon pool
ToMnet.715 / .770.829 / .807
Trajectory Transformer.718 / .787.826 / .809
Random Forest.832 / .852.878 / .866
TabPFN.871 / .880.907 / .897

The results support your concern. Expanding the training pool improved both trajectory models, especially in bargaining. At K=16, when all models had access to the same hackathon agents and used the same number of dialogue dimensions, Random Forest and TabPFN outperformed both trajectory models in both game families and at both pool sizes.

3. How were the tree-based models trained? We trained Random Forest, XGBoost, and CatBoost on both source- and target-agent rows, including the agent identity column. TabPFN received the same rows and identity column in context. The tree-based models’ advantage over the trajectory models supports our decision to formulate the task as a tabular prediction problem, even without the in-context learning capabilities of TFMs. Tree-based models are strong predictors and were state of the art on tabular data until about three years ago. More recently, TFMs have outperformed tree-based models on tabular tasks, so their stronger performance in our experiments is consistent with results on TabArena.

4. Why represent each decision as a row? At each decision point, the question is whether the agent will accept the offer on the table. The features describing the game, the current offer, and previous rounds have consistent meanings across games. Putting them in the same columns lets the predictor use the target agent’s past responses to similar offers directly. The same row can also include a representation of the dialogue. We therefore need a model that can use both text representations and numerical features, which makes a tabular model that supports text representations a natural choice. The Observer also produces numerical features that the tabular model can use directly.

5. What does the row representation give up? A five-round window can miss a promise made earlier in the game, and compressing the dialogue can obscure wording that affects the decision. The representation also discards the order in which the adaptation games were played. This makes it less suitable for agents that change strategy from one game to the next, such as a human player who learns from experience or an AI agent that is updated between runs.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response: new experiments and revisions to the paper

We thank all four reviewers for their careful and constructive feedback. We are encouraged that the reviewers found the problem important and practically relevant (p2Km, yi9G) and the motivation clear (UYa3).

The reviews asked for specific baselines, fairness controls and statistics. We ran these experiments during the discussion period. While preparing them, we fixed a bug in the tabular pipeline shared by the baseline and the model with the Observer. The pipeline matched game-state features by column position, but 13 target agent versions stored the columns in a different order. Matching columns by name alone raised the negotiation baseline from .719 to .800 AUC with no target games.

The table lists each concern, the experiment we added and its outcome. Full tables are in our individual replies to the reviewers. We are working on a revised PDF of the paper that will incorporate the rebuttal results. Pairs of numbers are AUCs for bargaining / negotiation.

Concern (reviewer)New experimentResult
Why formulate this as a tabular problem? Trajectory baselines are needed (UYa3)Adapted ToMnet and a TAO-inspired Trajectory TransformerAt K=16, even without the Observer, every tree ensemble and tabular foundation model we tested outperforms both trajectory models in both games.
Other tabular predictors (p2Km)Seven alternatives, including CatBoost, XGBoost, Kumo Tabular and TabICLv2The Observer’s contribution extends beyond TabPFN. We reproduce the result with other learners.
Stronger direct LLM (p2Km)A newer model with and without thinking, using the public view for both predictorsAt K=16 thinking raises the direct LLM from .737 / .718 to .793 / .779. The tabular model with the Observer reaches .854 / .835.
Unequal source supervision and test-set size (UYa3, bder)The same 16 source decisions, target games and test decisions for both systems.Both models receive the same 16 labeled source decisions, one per game. With no target games, the direct LLM performs better on both tasks. After 16 target games, TabPFN with the Observer substantially outperforms the LLM on both.
Forecasting from public information alone (bder)Both players’ private parameters removed from all inputs (tabular features, LLM, text encoder and Observer)At K=16, the Observer improves bargaining AUC by .031 with public information alone. In incomplete-information negotiation, its gain is .011 with public information and .029 when the forecaster knows the target’s valuation.
An alternative encoder given the same prompt as the Observerbge-m3 on the Observer’s exact promptFeatures from a stronger generic encoder do not improve the tabular models. The Observer retains its gain.
Paired confidence intervals and median R² (bder)Paired cluster bootstrap. Proposals were recomputed with game-level splits.The confidence intervals support the same conclusions.
Larger Observers (yi9G)Larger versions of our small Observers: Gemma-2 9B/27B, Qwen3 8B/14B and Llama-3.1 8BLarger Observers contribute greater improvements in performance.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.