PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Eilam Shapira, Moshe Tennenholtz, Roi Reichart

arXiv:2605.12411v1cs.LGcs.AIcs.CLcs.MA

TL;DR

The paper asks whether an unfamiliar language-based agent’s next decision can be predicted from only a few prior interactions when its implementation is hidden. It formulates prediction as target-adaptive text-tabular learning with frozen-LLM Observer representations, outperforming direct prompting and simpler feature baselines while revealing task-dependent gains.

  • Problem

    The paper asks whether an unfamiliar counterpart’s next decision can be predicted from a few interactions despite hidden LLMs, prompts, control logic, and fallbacks.

  • Method

    The method combines structured game-state, offer-history, dialogue, and frozen-LLM hidden-state features in a tabular predictor adapted with K labeled target games.

  • Results

    The full model outperforms direct LLM-as-Predictor prompting and game+text baselines across transfer from 13 frontier-LLM agents to 91 held-out scaffolded agents.

  • Takeaways & Limitations

    Target-adaptive text-tabular learning is better framed for this problem than direct few-shot prediction, while Observer hidden states add decision-relevant signal beyond their committed answers.

  • Takeaways & Limitations

    The evaluation uses controlled abstractions rather than real markets, assumes access to a relevant source population, and finds Observer contributions vary across tasks.

Abstract

from arXiv · show

AI agents negotiate and transact in natural language with unfamiliar counterparts: a buyer bot facing an unknown seller, or a procurement assistant negotiating with a supplier. In such interactions, the counterpart's LLM, prompts, control logic, and rule-based fallbacks are hidden, while each decision can have monetary consequences. We ask whether an agent can predict an unfamiliar counterpart's next decision from a few interactions. To avoid real-world logging confounds, we study this problem in controlled bargaining and negotiation games, formulating it as target-adaptive text-tabular prediction: each decision point is a table row combining structured game state, offer history, and dialogue, while $K$ previous games of the same target agent, i.e., the counterpart being modeled, are provided in the prompt as labeled adaptation examples. Our model is built on a tabular foundation model that represents rows using game-state features and LLM-based text representations, and adds LLM-as-Observer as an additional representation: a small frozen LLM reads the decision-time state and dialogue; its answer is discarded, and its hidden state becomes a decision-oriented feature, making the LLM an encoder rather than a direct few-shot predictor. Training on 13 frontier-LLM agents and testing on 91 held-out scaffolded agents, the full model outperforms direct LLM-as-Predictor prompting and game+text features baselines. Within this tabular model, Observer features contribute beyond the other feature schemes: at $K=16$, they improve response-prediction AUC by about 4 points across both tasks and reduce bargaining offer-prediction error by 14%. These results show that formulating counterpart prediction as a target-adaptive text-tabular task enables effective adaptation, and that hidden LLM representations expose decision-relevant signals that direct prompting does not surface.

1 Introduction

The paper asks whether an unfamiliar language-based agent’s next decision can be predicted from a few prior interactions, without access to its hidden implementation. It formulates this as target-adaptive text-tabular prediction and finds complementary value in frozen-LLM Observer representations.

  • The task predicts an unfamiliar counterpart’s next decision despite hidden LLMs, prompts, control logic, and rule-based fallbacks.
  • Controlled bargaining and negotiation games provide matched strategic conditions with known payoffs, free-text dialogue, and observable decisions.
  • Each decision point combines public game state, offer history, and dialogue, while K complete prior games from the same target provide labeled adaptation examples.
  • LLM-as-Observer adds hidden states from a small frozen LLM to game-state and dialogue features, using the LLM as an encoder rather than a final predictor.
  • The full model, trained on 13 frontier-LLM agents and tested on 91 held-out scaffolded agents, outperforms direct prompting and game+text baselines.

2 Related work

Prior work predicts behavior from limited histories, studies LLMs as strategic populations, and develops text-tabular and frozen-language-model representations. This paper combines these directions for rapid adaptation to previously unseen strategic agents.

  • Prior opponent modeling, negotiation, teamwork, and Theory-of-Mind methods show that short histories can support prediction but often assume known or matched agent types.
  • LLM-as-Predictor uses a large API-based LLM prompted in context to predict decisions directly from few-shot examples.
  • Text-tabular learning combines structured game fields with free-form dialogue, matching decision points that contain both strategic state and language.
  • This setting requires rapid adaptation to a newly observed strategic agent from only K games, using source-population rows without gradient-based retraining.
  • Frozen language-model hidden states can encode task-relevant information not always captured by final outputs, motivating their use as transferable features.

3 Data

The study uses controlled, language-based bargaining and negotiation games with two complementary agent populations for cross-population transfer. Each decision is represented as a text-tabular prediction problem involving game state, offer history, dialogue, and the target agent’s next move.

  • GLEE provides sequential two-player economic games in which agents make offers or accept/reject decisions through public interaction histories and free-text messages.
  • Bargaining divides a fixed sum over alternating offers, with configurations varying in horizons, discount factors, and information about the other agent’s discount factor.
  • Negotiation concerns a single indivisible good with private seller reserves and buyer valuations, alternating price offers, acceptance, rejection, or an outside option.
  • The source population contains 13 frontier LLMs varying in underlying model, while the held-out target population contains 91 agents varying mainly in scaffolding around a shared underlying LLM.
  • The source–target split tests transfer between agents differing mainly in underlying LLM and agents differing mainly in scaffolding.

4 Method

The method treats predicting an unseen agent’s next move as target-adaptive tabular inference over multimodal decision-point rows. A tabular predictor combines public game information, dialogue, Observer representations, and labeled examples from the target’s prior games.

  • Feature modalities: Each decision point becomes a multimodal row containing public game state, dialogue, and a decision-oriented Observer representation.The Observer is a small frozen LLM whose hidden state is used as a feature rather than whose answer is treated as the prediction.
  • Prediction tasks: The model predicts response acceptance as binary classification and the next proposal as regression over normalized offer values.These tasks cover the two observable decision types studied in the bargaining and negotiation games.
  • Target adaptation: An agent-identity indicator separates population-level regularities from target-specific behavioral deviations.Without this marker, source and target rows would be treated as exchangeable despite coming from different decision-makers.
  • Target adaptation: For each target, the predictor combines source-population rows with K observed target games while receiving only public information at inference time.Private values and the target’s prompt, code, and control logic are not supplied.
  • Baselines: The evaluation compares the full target-adaptive model with a game+text baseline and direct LLM-as-Predictor prompting.The baseline removes Observer features, while direct prompting does not produce a reusable representation combined with source-population rows.

5 Experimental setup

The experiments test cross-population transfer from frontier-LLM agents to held-out scaffolded agents under varying amounts of target adaptation. Performance is measured separately for binary response decisions and continuous proposal prediction.

  • Evaluation protocol: The evaluation trains on the frontier-LLM tournament population and tests on held-out hackathon agents one target at a time.Source variation changes the underlying LLM, whereas target variation changes prompts, control logic, and rule-based fallbacks under fixed scaffolding.
  • Evaluation protocol: K takes values 0, 2, 4, 8, and 16 observed games, with evaluation performed on each target’s remaining games.The tabular predictor uses up to 3,000 balanced source decisions together with the target’s K-game decisions.
  • Metrics: Response prediction uses AUC, while proposal prediction uses R2 over normalized offers.TabPFN v2.6 instantiates the tabular prediction module for both tasks.

6 Results

Observer hidden states outperform direct prompting and the game+text baseline for response prediction across both game families. For proposals, the Observer helps most in bargaining, where structured history alone is insufficient, while direct prompting is poorly calibrated for numerical offers.

  • Response prediction: +4.0pp in bargaining and +4.9pp in negotiation are the best Observer gains over game+text features for response prediction at K=16.The Observer also exceeds LLM-as-Predictor by +6.1pp in bargaining and +6.7pp in negotiation at K=16.
  • Response prediction: Observer hidden states improve response prediction over the tabular baseline even at K=0, without target-specific examples.This indicates that the representation contributes signal independently of target adaptation examples.
  • Direct prompting: LLM-as-Predictor trails LLM-as-Observer at every K in both response-prediction game families.The comparison uses a large frontier API model with the current game and the target’s observed games directly in context.
  • Proposal prediction: +0.05 median R2 is the approximate bargaining proposal gain over game+text features at K=16.Using Gemma-2-2B, one-offer prediction error falls from $552 to $473, a 14% reduction, on a nominally $10,000 split.
  • Proposal prediction: In negotiation, the game+text baseline is already strong at K=16, so Observer variants provide no clear additional proposal improvement.The bargaining-specific gain occurs where interpreting dialogue alongside numerical offers is especially important.
  • Direct prompting: In bargaining, LLM-as-Predictor has negative median R2 even at K=16, while in negotiation it remains below game+text features.The results characterize direct prompting as poorly calibrated for numerical offer regression.

7 Robustness and ablation

Ablations identify structured game features as the backbone and Observer states as a complementary situational layer. The Observer effect remains consistent across encoder providers and mid-to-late layers rather than depending on one tuned representation.

  • Feature ablations: Removing structured game features causes the largest performance collapse, particularly in negotiation.This establishes game-state features as the essential backbone of the feature hierarchy.
  • Feature ablations: Once Observer states are included, generic sentence embeddings provide no meaningful marginal gains.The Observer supplies situational information that generic dialogue embeddings do not capture.
  • Latent versus direct prediction: Tabular prediction from Observer hidden states consistently outperforms the Observer’s direct accept/reject logits across all conditions.The tabular learner decodes strategic signals from the latent space more effectively than the LLM’s output head.
  • Layer and provider stability: Observer gains remain stable across relative depth 0.6–0.9 for response and proposal tasks in bargaining and negotiation.Figure 3 organizes these gains by task, game family, and number of K-shot examples.
  • Layer and provider stability: Gemma-2-2B, Qwen3-1.7B, and Llama-3.2-1B hidden states all consistently improve over the baseline.The stability across providers, tasks, and layers supports the robustness of the Observer signal.

8 Discussion and conclusion

The paper frames unfamiliar-agent decision prediction as target-adaptive text-tabular learning, combining structured strategy signals, dialogue, and target-specific observations. Its results support separating representation from adaptation, while remaining bounded by controlled games, source-population access, and task variation.

  • Discussion and conclusion: Target-adaptive text-tabular learning combines structured game-state and offer-history features, dialogue features, and K observed target games.The tabular learner conditions on source-population and target-specific rows to adapt without observing the target’s prompt, code, or control logic.
  • Discussion and conclusion: The Observer adds a complementary decision-oriented signal beyond game-state and generic dialogue representations.Its contribution is strongest when language and strategic positioning matter beyond offer history.
  • Discussion and conclusion: Observer gains vary by task: they improve response prediction and bargaining proposal prediction but add little in negotiation proposal prediction.In negotiation, structured state, offer history, and target examples already predict the next offer well.
  • Discussion and conclusion: The evaluation transfers from controlled frontier-LLM agents to held-out scaffolded agents with different prompts, control logic, and rule-based fallbacks.This tests adaptation to newly encountered engineered agents from a reusable controlled source population.
  • Discussion and conclusion: The study is limited because controlled games abstract language-mediated commerce, the method assumes a relevant source population, and Observer benefits vary across tasks.These boundaries qualify how broadly the approach should be generalized.
  • Discussion and conclusion: The proposed recipe separates representation from adaptation: language models construct decision-relevant representations, while supervised learning combines incentives and observed decisions.This design assigns final prediction to a model that can combine structured and population-level evidence with target-specific observations.

B Frontier-LLM tournament model list

This section documents the model populations, configurations, cohort construction, and proposal-prediction protocol used in the experiments. It also specifies the task-matched Observer prompts and the primary Observer-versus-baseline comparison.

  • Model population: The frontier-LLM source tournament contains 13 models, all receiving the same game-facing system prompt with only the underlying LLM varying.The model list spans providers including Anthropic, Google, Meta, Mistral, OpenAI, and xAI.
  • Game configurations: Table 4 enumerates hackathon configurations by bargaining and negotiation parameters, with all configurations enabling free-text messages.Bargaining varies money, horizon, information, communication, and discount factors; negotiation varies valuations, price order, horizon, and information.
  • Target population: Hackathon evaluation uses 91 team-stage agents and 11,341 accept/reject decisions across 4,921 bargaining and negotiation games.The predictive cohort includes every team-stage agent appearing at any competition stage, not only the publicly released final-round submissions.
  • Proposal prediction: Proposal prediction estimates a proposer’s next offer from round r≥2 using a scale-normalised target and closed-form inverse recovery of the dollar amount.The regression protocol uses K∈{0, 2, 4, 8, 16}, game-level splits, five seeds, and TabPFN in regressor mode.
  • Observer prompts: Task-matched prompt suffixes orient the Observer toward accept/reject decisions, bargaining proposer gains, or negotiation price amounts.The suffix is changed only for Observer extraction in proposal prediction; the response-prediction suffix remains unchanged.
  • Feature comparison: The proposal comparison contrasts a game+text baseline with the same feature stack augmented by the Observer hidden-state representation.The primary comparison measures the Observer’s marginal contribution on top of game and text features.
  • Cohort construction: Proposal-prediction cohorts require at least 30 round-≥2 proposer decisions and target standard deviation ≥0.02, yielding 78 bargaining and 20 negotiation agents.These filters define the included hackathon agents for the regression analysis.

E Provider replication of the logits-vs-hidden-states gap

The provider-replication analysis tests whether the Observer’s hidden-state advantage over direct logits persists across evaluation setups and Observer providers. The reported comparisons isolate readout type under cross-population transfer.

  • Matched comparison: The analysis compares direct Observer logits with hidden-state representations under cross-population transfer at K=16 using TabPFN.Table 5 reports mean AUC over five seeds for 72 bargaining and 39 negotiation targets.
  • Logits versus hidden states: For Gemma-2-2B, adding the logit scalar yields only a marginal improvement over game+text features, whereas adding hidden states yields a clearly larger improvement.Combining logits and hidden states is a wash relative to hidden states alone.
  • Interpretation: The predictive signal lives in the Observer representation rather than its direct readout.The same conclusion is stated for the matched logits-versus-hidden-states comparison.
  • Provider replication: The logits-versus-hidden-states comparison is replicated with Qwen3-1.7B and Llama-3.2-1B Observers from unrelated providers.Neither Observer shares a training pipeline or parent company with the hackathon agents’ underlying LLM.
  • Provider replication: Table 6 reports direct p(accept) AUC for logits and TabPFN results using game-state features plus averaged Observer hidden states.Hidden states are averaged over each Observer’s upper-stack layer band at relative depth 0.6–0.9.

F Additional experimental details

The additional details specify Observer extraction, tabular evaluation, prompting, splitting, hardware, compute, and uncertainty reporting. Together they define the reproducibility settings behind the reported experiments.

  • Observer extraction: Observer hidden states for Gemma-2-2B, Qwen3-1.7B, and Llama-3.2-1B were extracted over approximately 67K games and 200K decisions.Upper-stack layers were cached in one forward pass per game, with the main representation averaging each model’s upper-stack layer band.
  • Tabular evaluation: TabPFN evaluations used up to 3,000 source-balanced training rows, target K-game decisions, and at most 500 test rows per cell.Each cell ran with default settings on a single RTX A6000 in approximately 10–30 seconds.
  • Direct prompting baseline: The direct LLM-as-Predictor baseline used Gemini 2.5 Flash with thinking_budget=0 across bargaining and negotiation targets, K values, seeds, and test decisions.A separate thinking-on pilot used thinking_budget=2000 but was not scaled to the full table because of cost.
  • Compute: Experiments consumed approximately 140 A6000 GPU-hours, while preliminary and discarded runs brought broader project usage to roughly 400 GPU-hours.The reported total includes extraction and tabular evaluation, plus the Gemini API budget described separately.
  • Hyperparameters: Observer representations average upper-stack layers at relative depth 0.6–0.9, using default TabPFN settings and five seeds.The Observer layer counts are 26 for Gemma-2-2B, 28 for Qwen3-1.7B, and 16 for Llama-3.2-1B.
  • Data splitting: At each K, the split samples K whole games for adaptation and reserves the remaining games for testing.Game-level splitting prevents rounds from the same game appearing on both sides and makes K count games rather than decisions.
  • Uncertainty: Reported ± values are standard errors of per-agent, per-seed AUC or R2 values for each evaluation cell.The uncertainty unit is the agent-seed observation rather than an individual decision.

G Game+text features baseline: feature specifications

The baseline represents each decision as a multimodal tabular row combining structured game state, dialogue embeddings, and an agent-identity indicator; the Observer representation is an additional model component. Features cover configuration, current offers, recent history, and family or domain compatibility across bargaining and negotiation tasks.

  • Feature blocks: Each decision row combines game-state features, dialogue representation, and an agent-identity indicator in the game+text baseline.The Observer hidden-state representation is appended only in the Observer-augmented model.
  • Bargaining features: Bargaining game-state inputs contain 24 columns covering configuration, current offer, five prior rounds, and a family indicator.Configuration includes round and discounting variables; offer features encode the split and inflation losses.
  • Negotiation features: Negotiation game-state inputs contain 25 columns covering configuration and absolute seller and buyer outside-option payoffs.The configuration includes reservation valuations, product-price scale, messaging, and information regime.
  • Text representation: Dialogue is encoded by concatenating all messages exchanged within a responder’s round and applying all-MiniLM-L6-v2 to produce a 384-dimensional vector.Rounds without messages use a placeholder string, and the resulting representation belongs to the dialogue feature block.
  • Identity feature: The agent-identity indicator uses one-hot dimensions for source agents and the held-out target, separating train and test at K=0 but adding within-test information when K>0.The target column is always zero in training at K=0 and active for test rows.
  • Preprocessing: PCA projections for dialogue and Observer representations are fit on each evaluation cell’s training pool and then applied unchanged to the test pool.No test-row information enters PCA fitting.
Loading 2605.12411v1…