Abstraction Agent

Boning Li; Longbo Huang

Source-linked AI summary

Abstraction Agent

Boning Li, Longbo Huang

arXiv:2609.04303v1Sep 3, 2026cs.MAcs.AIcs.CLcs.GT

TL;DR

Large-game abstraction traditionally depends on game-specific evaluators that are unavailable for many games. The Abstraction Agent elicits continuous strategic features from natural-language rules, clusters scored private states, and reduces exploitability by up to 62% versus EHS while transferring across games.

Problem
Effective information abstraction traditionally requires game-specific evaluators, training data, or game-tree traversal, limiting applicability to less-studied games.
Method
The Abstraction Agent discovers calibrated strategic features from natural-language rules, scores private states, filters redundant features, and clusters them without game-specific evaluators or traversal.
Results
The pipeline reduces HUNL turn exploitability relative to EHS by up to 62%, beats scalar rank at every ROVER Trials granularity, and transfers unchanged across PLO4, HUNL, and Riichi Mahjong.
Takeaways & Limitations
LLMs can serve as structured interfaces that convert implicit strategic knowledge into explicit numerical features for downstream game-solving algorithms.
Takeaways & Limitations
Potential-aware abstractions achieve lower exploitability because they use exact equity calculators; the agent closes about a third of that gap without such evaluators.

Abstract

from arXiv · show

Information abstraction, which groups strategically similar private states into a tractable number of buckets, is essential for scaling game-solving algorithms to large imperfect-information games. Constructing effective abstractions, however, has traditionally required domain-specific evaluators such as hand-strength calculators or equity estimators, which demand expert knowledge and engineering effort and are unavailable for most less-studied games. We propose the Abstraction Agent, a zero-shot pipeline that uses a large language model (LLM) to discover continuous strategic features from a natural-language game description, score private states on these features, and cluster them into abstraction buckets, without any game-specific evaluator, training data, or game-tree traversal during abstraction construction. The pipeline runs in four phases: feature discovery with calibration anchors, batched private-state scoring, correlation-based feature selection, and $k$-means clustering. The resulting abstractions reduce lifted-strategy exploitability by up to 62% relative to an expected-hand-strength baseline on heads-up no-limit Texas hold'em (HUNL) turn endgames, and beat a scalar rank baseline at every granularity on ROVER Trials, an original game absent from any pretraining corpus. Beyond these quantitative benchmarks, the pipeline transfers with unchanged prompts to four-card Pot-Limit Omaha, HUNL preflop and flop, and Riichi Mahjong, where the discovered features track each game's recognized strategic concepts. This is structured knowledge elicitation: converting implicit strategic knowledge in LLM parameters into explicit numerical features for downstream algorithmic computation. The code is available at https://github.com/lbn187/AbstractionAgent.

1. Introduction

The paper introduces a zero-shot LLM pipeline that constructs multi-dimensional game abstractions from natural-language rules without evaluators, training data, or game-tree traversal. It improves over standard baselines on HUNL and ROVER Trials and transfers unchanged across several games.

The approach uses LLMs to elicit qualitative strategic structure while delegating quantitative equilibrium computation to an exact game solver.
The Abstraction Agent converts natural-language game descriptions into multi-dimensional information abstractions without hand evaluators, rollout simulators, or game-tree traversal.
62%: HUNL turn abstractions reduce exploitability relative to EHS at K=100, beating EHS in every K=100 run.The advantage grows monotonically with bucket count and requires a frontier backbone.
On ROVER Trials, the rules-only pipeline beats the scalar rank baseline at every granularity in a game absent from pretraining corpora.This result is presented as evidence of feature construction rather than retrieval of memorized poker knowledge.
The pipeline transfers unchanged to PLO4, HUNL preflop and flop, and Riichi Mahjong, producing strategically coherent abstractions.These settings lack practical evaluator-based abstraction signals or make them expensive to compute.

2. Related Work

Prior abstraction methods use hand ranks, equity, potential-aware signals, learned features, or game-tree traversal. This paper instead uses pretrained LLM knowledge as a domain-knowledge-free source of explicit strategic features.

Traditional poker abstractions cluster hands by rank, expected hand strength, or equity, while potential-aware methods incorporate future-card distributions.
EVPA learns abstraction features from CFR traversal data, whereas WEVA avoids neural training but still requires game-tree traversal.
The Abstraction Agent uses pretrained LLM knowledge instead of game-specific training or traversal to provide abstraction signals.
Unlike LLM game-playing and online evaluation work, this approach uses LLMs as offline feature engineers for game-theoretic abstraction.
Calibration anchors and constrained output formatting address the technical challenge of eliciting consistent numerical outputs from LLMs.

3. Preliminaries

The paper formalizes two-player zero-sum extensive-form games, hidden-information information sets, strategies and equilibrium, exploitability, and abstraction-induced strategy lifting.

An extensive-form game represents histories, terminal outcomes, available actions, chance events, players, and utilities, with zero-sum utilities summing to zero.
Information sets group histories indistinguishable to a player, requiring identical available actions and behavior across those histories.
A strategy assigns action distributions to information sets, while a Nash equilibrium leaves neither player able to gain by deviating.
Exploitability measures the average gain available to an optimal adversary; equilibrium has exploitability zero, and larger values indicate less optimal strategies.
Information abstraction maps N private types to K much smaller buckets, forcing types in each bucket to share an action distribution after strategy lifting.

4. Method

The Abstraction Agent uses a four-phase, game-agnostic pipeline: discover calibrated continuous features, score holdings, remove redundant dimensions, and cluster feature vectors into K buckets.

The game-agnostic agent receives a declarative GameConfig and avoids hand-strength calculators, equity estimators, and rollout simulators.
Scoring against fixed features keeps calls local, makes batching consistent through anchors, and leaves global partitioning to K-means.
Feature discovery proposes continuous [0,1] strategic features with descriptions and 0.0, 0.5, and 1.0 calibration anchors.
Private-state scoring prompts the LLM to assign each holding a feature-score vector using batched holdings and shared anchors.
Feature selection removes near-constant and highly correlated dimensions, then rescales the surviving columns before clustering.
K-means partitions the holdings into K buckets using the rescaled feature matrix.

5. Experimental Setup

The study evaluates the Abstraction Agent against evaluator-based and embedding baselines on HUNL and tests portability across ROVER Trials, poker variants, and Riichi Mahjong.

Abstraction Agent Pipeline: The pipeline maps a game configuration and private holdings to K abstraction buckets through feature discovery, batched scoring, feature selection, and multi-restart k-means.Feature selection removes near-constant and highly correlated dimensions before rescaling the survivors.
ROVER Trials: ROVER Trials is an original two-player zero-sum hidden-information game whose private states are defined by three five-valued rover dials.It tests whether the agent can construct strategic features from rules rather than rely on memorized poker-specific structure.
Baselines: Baselines include EHS, potential-aware abstraction, LLM-Embed, random assignment, and full abstraction; evaluator-based methods require domain-specific equity computation.EHS uses expected equity and Monte Carlo rollout, while potential-aware abstraction additionally uses future-card variance or enumeration.
Implementation and cost: The implementation uses GPT-5.5 for main results, deterministic batched private-state scoring, retries, and approximately ten scoring batches per seed.The study also tests Llama-3.3-70B and Qwen-2.5-32B under the same prompts and reports parse rate and clustering stability.

6. Results

The Abstraction Agent improves over EHS on HUNL at finer granularities, beats scalar rank on ROVER Trials, and transfers coherently across additional games.

HUNL turn endgames: 62%: at K=100, the agent reduces HUNL turn exploitability relative to EHS, with all 20 runs beating EHS and confidence intervals below 1.0.The mean exploitability ratio declines from 1.28 at K=20 to 0.78 at K=50 and 0.38 at K=100.
Backbone dependence: GPT-5.5 is the only tested backbone that reliably beats EHS; Llama-3.3-70B is near parity and Qwen-2.5-32B is consistently worse.This comparison is reported as mean Agent/EHS exploitability ratios over ten seeds.
Comparison with Other Methods: At K=50, the agent beats random assignment, LLM-Embed, and EHS, while potential-aware abstraction remains strongest and uses exact equity computation.The agent closes about one third of the gap to potential-aware abstraction without a domain-specific evaluator.
ROVER Trials: On ROVER Trials, the agent beats scalar Rank at every granularity, with the margin widening from 0.92 at K=5 to 0.66 at K=30.Against potential-aware abstraction, ratios are 0.91–0.94 at K∈{10,20,30}, but confidence intervals overlap 1.0.
Constructed versus retrieved features: ROVER results support construction from supplied rules because the game has no published strategy literature or per-holding bucketing recipe to copy.The paper frames this as evidence for useful feature construction rather than retrieval of a memorized poker abstraction.
Cross-game portability: The unchanged pipeline produces strategically coherent features across PLO4, HUNL preflop and flop, and Riichi Mahjong.The features track suit coordination and connectivity, draw-versus-made separation, premium-pair grouping, and Mahjong shanten-speed, acceptance, and draw-swing concepts.

7. Analysis

The analysis attributes the agent’s gains to multi-dimensional features that retain strategic information lost by scalar EHS, while showing construction and portability beyond memorized poker knowledge.

A multi-feature abstraction preserves vulnerability, nut potential, and blocker effects that scalar EHS cannot represent.Made strength correlates strongly with EHS, whereas improvement and river-shift features capture complementary dimensions.
Exploitability ratios fall from 1.28 at K=20 to 0.78 at K=50 and 0.38 at K=100 as granularity increases.The corresponding clustering is stable across seeds, with mean pairwise ARI of 0.77 overall.
Vulnerability_To_Outdraws and Blocker_Value carry the most unique clustering signal in leave-one-out ablations.
Against exact potential-aware methods, the agent’s value is evaluator-free abstraction that closes about a third of the gap to PA rather than matching it.PA achieves lower exploitability because it uses exact equity calculators; the zero-shot pipeline requires neither evaluators nor game-tree traversal.
The pipeline beats scalar rank on original ROVER Trials and transfers to PLO4, HUNL, and Riichi Mahjong, supporting construction beyond memorized poker features.ROVER is absent from pretraining, while the unchanged pipeline discovers domain-specific features across other games.

8. Conclusion

The paper concludes that the Abstraction Agent can convert natural-language game rules into strategic features and abstractions without game-specific evaluators or game-tree traversal. It achieves strong HUNL and ROVER results and produces coherent abstractions across additional games and settings.

The Abstraction Agent discovers and scores strategic features from natural-language game descriptions without game-specific evaluators or game-tree traversal.Its prompts and downstream scoring support formal information abstraction from textual rules and private-state renderings.
HUNL turn exploitability decreases by up to 62% relative to EHS, with all K=100 runs beating EHS and gains increasing with bucket count.
On ROVER Trials, the rules-only pipeline beats scalar rank at every granularity, while unchanged prompts produce coherent abstractions in PLO4, HUNL preflop/flop, and Riichi Mahjong.
The appendix supplies game descriptions as the system’s only domain knowledge, excluding game-specific evaluation code, hand evaluators, and equity calculators.
The pipeline’s prompts are game-agnostic, with domain knowledge entering through GameConfig fields and feature-scoring instructions.

D. Experimental Details

The experiments evaluate discovered features and clusters across HUNL, PLO4, and HUNL flop and preflop settings, using reproducible solver configurations and exact exploitability measurements where applicable. The resulting clusters separate strategically meaningful hand types without poker-specific code.

The evaluation covers HUNL turn subgames, HUNL preflop, PLO4 preflop, and HUNL flop, with exploitability measured for HUNL turn and ROVER.
HUNL Preflop: HUNL preflop clustering separates premium pairs, suited connectors, offsuit broadways, and junk hands across 169 canonical hands in K=10 clusters.All seven discovered features were retained and the parse rate was 100%.
HUNL Preflop: HUNL clusters separate strongest starting hands, premium suited broadways, small-to-medium pairs, and suited from offsuit holdings according to strategic coherence.
HUNL Turn: Made-strength features correlate with EHS at r≈0.87–0.93, while improvement and river-shift features show low or negative correlation.The correlation pattern indicates that the latter features encode strategic dimensions beyond scalar EHS.
PLO4 Preflop: PLO4 preflop clusters separate rank strength, suitedness, connectivity, and coordination across 16,432 canonical hands in K=30 clusters.The clusters distinguish premium rainbow holdings from low-card suited rundowns and separate top from low rundowns despite shared straight connectivity.
HUNL Flop: On the A♠K♥Q♦ flop, K=20 clusters separate sets, the nut Broadway straight, top two pair, kicker tiers, draws, and no-interaction air.

J. Prompt-Variant and Rules-Only Robustness

Prompt-variant results indicate that the reported abstraction gains are robust to alternative feature-discovery instructions and do not depend on the public range summary.

At K∈{20,50}, rules-only, range-aware, and feature-discovery prompt variants agree within per-seed standard deviation of 0.10–0.35.
The rules-only variant matches the range-aware variant, indicating that the public range summary is not the source of the advantage.

K. Feature Ablation Results

Feature ablations identify which discovered signals most influence clustering, while Riichi Mahjong clusters separate private states by distinct strategic roles.

Feature Importance: Vulnerability_To_Outdraws and Blocker_Value carry the most unique clustering signal, whereas River_Equity is largely redundant.Leave-one-out importance is measured by mean adjusted Rand index, with lower values indicating greater impact from feature removal.
Riichi Mahjong Analysis: Riichi Mahjong clustering discovers six features and retains five after correlation filtering.The reported clusters use a mid-game setting and separate hands by offensive speed, defensive resilience, and value potential.
Riichi Mahjong Analysis: The resulting Riichi Mahjong clusters separate hands by offensive speed, defensive resilience, and value potential.Representative clusters were produced with a 100% parse rate.

M. Additional Implementation and Analysis Details

The pipeline is designed for game-agnostic deployment through compact rule and rendering specifications, reliable prompt formatting, and cross-game feature patterns. Its analyses show high parsing robustness, while future work targets model dependence, oracle-free selection, online refinement, and unparsable-score handling.

Game-Agnostic Design: Applying the agent to a new game requires a GameConfig and two rendering functions, without evaluation code, rollout simulators, or domain-specific feature engineering.The GameConfig typically contains 20–50 lines of natural language and supplies rules, scoring context, and rendering callbacks.
Prompt Design: Calibration anchors, explicit spread and independence constraints, and fixed batch formatting improve feature-discovery consistency and parsing reliability.Anchors map 0.0, 0.5, and 1.0 to concrete situations, while numbered outputs use a fixed feature-value format.
Cross-Game Feature Analysis: Recurring feature families span current made-hand strength, nut potential, blocker effects, and river volatility across games and runs.Current made-hand strength is the family most correlated with EHS, with reported correlations of approximately r≈0.87–0.93.
Robustness: Parse rates exceed 99% overall across evaluated games, reaching 100% on turn, preflop, PLO4, and Mahjong and 97.9% on flop.Three discovery-prompt variants produce mean ratios agreeing within per-seed standard deviation at K∈{20, 50} for turn results.
Future Work: Future work includes understanding dependence on the LLM backbone, developing oracle-free abstraction-quality proxies, refining features in the solver loop, and replacing midpoint imputation.The authors state that a frontier model is currently needed to beat EHS.
Broader Implications: The rules-to-features-to-clusters-to-exploitability chain supports broader uses in structured scoring, LLM evaluation, automated feature engineering, and other strategic domains.The paper specifically identifies security games, auctions, and negotiation as potential transfer settings.