FinBoardBench benchmarks LLMs on dynamic wealth management using Cashflow, Acquire, and Monopoly simulations, finding they struggle with liquidity and complex interactions despite static reasoning ability.
hub
Improving Language Model Negotiation with Self-Play and In-Context Learning from AI Feedback
23 Pith papers cite this work, alongside 35 external citations. Polarity classification is still indexing.
23
Pith papers citing it
35
external citations · Pith
abstract
We study whether multiple large language models (LLMs) can autonomously improve each other in a negotiation game by playing, reflecting, and criticizing. We are interested in this question because if LLMs were able to improve each other, it would imply the possibility of creating strong AI agents with minimal human intervention. We ask two LLMs to negotiate with each other, playing the roles of a buyer and a seller, respectively. They aim to reach a deal with the buyer targeting a lower price and the seller a higher one. A third language model, playing the critic, provides feedback to a player to improve the player's negotiation strategies. We let the two agents play multiple rounds, using previous negotiation history and AI feedback as in-context demonstrations to improve the model's negotiation strategy iteratively. We use different LLMs (GPT and Claude) for different roles and use the deal price as the evaluation metric. Our experiments reveal multiple intriguing findings: (1) Only a subset of the language models we consider can self-play and improve the deal price from AI feedback, weaker models either do not understand the game's rules or cannot incorporate AI feedback for further improvement. (2) Models' abilities to learn from the feedback differ when playing different roles. For example, it is harder for Claude-instant to improve as the buyer than as the seller. (3) When unrolling the game to multiple rounds, stronger agents can consistently improve their performance by meaningfully using previous experiences and iterative AI feedback, yet have a higher risk of breaking the deal. We hope our work provides insightful initial explorations of having models autonomously improve each other with game playing and AI feedback.
hub tools
citation-role summary
background 4
citation-polarity summary
roles
background 3polarities
background 3representative citing papers
A survey that unifies prior work on multi-agent LLM systems via the LIFE framework, mapping dependencies across collaboration, failure attribution, and autonomous self-evolution while identifying cross-stage challenges.
An AI-agent social platform generated mostly neutral content whose use in fine-tuning reduced model truthfulness comparably to human Reddit data, suggesting limited unique harm but flagging tail risks like secret leaks.
METRO induces both short-term actions and long-term planning from expert transcripts into a Strategy Forest, outperforming prior methods by 9-10% on two non-collaborative dialogue benchmarks.
The paper defines and measures 'problem drift' in multi-agent LLM debates across tasks and proposes DRIFTJudge and DRIFTPolicy as baselines to detect and reduce it.
An adaptive stochastic offer policy cuts adversary inference of private negotiation constraints by 43–50% on synthetic traces while keeping success and utility above 90%.
RLVR training teaches a 30B LLM to strategically explore a multi-buyer market and extract 70% of available surplus, outperforming frontier models up to 1T parameters in concurrent negotiation.
InduceKV is a retrieval-based continual adaptation method that uses bilevel selection to build a compact set of inducing KV memories for fixed-footprint updates to multimodal LLMs.
Introduces the Contagion Tensor and CAF metrics to measure output-distribution coupling in multi-agent LLM systems, with simulation ablations showing artifact removal and real-API tests on GPT-4o-mini and DeepSeek validating modality effects.
Centralized matching mechanisms outperform free negotiation in stability and efficiency with LLM agents, who also report preferences truthfully more often than humans, though not always in line with strategy-proofness predictions.
SEPO adds penalties for exploitability, collusion, and externalities to GRPO training of LM agents, yielding zero exploit advantage in Kuhn Poker and safety gains in four of five strategic domains.
Anchor-and-resume with spread-derived beta allows adaptive monotonic concessions in freight negotiations, achieving LLM-like performance with lower cost and higher transparency.
A small set of sparse autoencoder features in LLMs drives shifts between generous and selfish allocations in dictator games, with causal patching and steering confirming their role and generalization to other social games.
PROCHATIP is a chatbot system that learns to probe conversations for target business information at the right times, outperforming baselines in both data collection and user experience.
Frontier LLMs exhibit high scheming propensity in Cheap Talk signaling and Peer Evaluation games, achieving 95-100% success rates when choosing to deceive and 100% deception choice in one setup even without prompting.
Empirical analysis shows scaling inference compute via strategies like tree search can be more efficient than scaling model parameters, with 7B models plus novel search outperforming 34B models.
Creates a five-dimension taxonomy (counterparty, payload, interaction state, discovery mechanism, schema flexibility) from nine protocols and identifies architectural patterns plus convergence trends.
A method combining LLM-extracted qualitative cues with Bayesian belief tracking improves full agreement rates and preference estimation accuracy in multi-agent negotiations.
MAC framework selects Pareto-optimal LLM agents and masks low cross-consistency outputs for adaptive collaboration in medical decision-making.
TrustLLM defines eight trustworthiness principles, creates a six-dimension benchmark, and evaluates 16 LLMs showing proprietary models generally lead but some open-source ones are close while over-calibration can hurt utility.
Multi-agent debate with tit-for-tat arguments and a judge LLM improves reasoning by preventing LLMs from locking into incorrect initial solutions.
PRISMA augments self-training with direct preference optimization and an emotion-aware negotiation strategy chain-of-thought to produce more interpretable and effective negotiation dialogues on two new datasets.
The paper surveys the origins, frameworks, applications, and open challenges of AI agents built on large language models.
citing papers explorer
-
FinBoardBench: Benchmarking Dynamic Wealth Management and Strategic Financial Reasoning of LLMs via Board Game Simulations
FinBoardBench benchmarks LLMs on dynamic wealth management using Cashflow, Acquire, and Monopoly simulations, finding they struggle with liquidity and complex interactions despite static reasoning ability.
-
Beyond Individual Intelligence: Surveying Collaboration, Failure Attribution, and Self-Evolution in LLM-based Multi-Agent Systems
A survey that unifies prior work on multi-agent LLM systems via the LIFE framework, mapping dependencies across collaboration, failure attribution, and autonomous self-evolution while identifying cross-stage challenges.
-
The Moltbook Files: A Harmless Slopocalypse or Humanity's Last Experiment
An AI-agent social platform generated mostly neutral content whose use in fine-tuning reduced model truthfulness comparably to human Reddit data, suggesting limited unique harm but flagging tail risks like secret leaks.
-
METRO: Towards Strategy Induction from Expert Dialogue Transcripts for Non-collaborative Dialogues
METRO induces both short-term actions and long-term planning from expert transcripts into a Strategy Forest, outperforming prior methods by 9-10% on two non-collaborative dialogue benchmarks.
-
Stay Focused: Problem Drift in Multi-Agent Debate
The paper defines and measures 'problem drift' in multi-agent LLM debates across tasks and proposes DRIFTJudge and DRIFTPolicy as baselines to detect and reduce it.
-
Behavioral Privacy Leakage in Agentic Negotiation: Formalizing and Mitigating Inference Attacks via Randomized Policies
An adaptive stochastic offer policy cuts adversary inference of private negotiation constraints by 43–50% on synthetic traces while keeping success and utility above 90%.
-
Strategic Bargaining in Multi-Buyer Markets: Reinforcement Learning from Verifiable Rewards for LLM Negotiations
RLVR training teaches a 30B LLM to strategically explore a multi-buyer market and extract 70% of available surplus, outperforming frontier models up to 1T parameters in concurrent negotiation.
-
InduceKV: Fixed-Footprint Continual Adaptation of Multimodal LLMs via Inducing KV Memories
InduceKV is a retrieval-based continual adaptation method that uses bilevel selection to build a compact set of inducing KV memories for fixed-footprint updates to multimodal LLMs.
-
The Contagion Tensor: A Framework for Measuring Output-Distribution Coupling in Multi-Agent LLM Systems -- and Auditing the Claims It Enables
Introduces the Contagion Tensor and CAF metrics to measure output-distribution coupling in multi-agent LLM systems, with simulation ablations showing artifact removal and real-API tests on GPT-4o-mini and DeepSeek validating modality effects.
-
Do Matching Mechanisms Work with LLM Agents?
Centralized matching mechanisms outperform free negotiation in stability and efficiency with LLM agents, who also report preferences truthfully more often than humans, though not always in line with strategy-proofness predictions.
-
Safe Equilibrium Policy Optimization for Strategic Agent Policies
SEPO adds penalties for exploitability, collusion, and externalities to GRPO training of LM agents, yielding zero exploit advantage in Kuhn Poker and safety gains in four of five strategic domains.
-
Anchor-and-Resume Concession Under Dynamic Pricing for LLM-Augmented Freight Negotiation
Anchor-and-resume with spread-derived beta allows adaptive monotonic concessions in freight negotiations, achieving LLM-like performance with lower cost and higher transparency.
-
Understanding the Mechanism of Altruism in Large Language Models
A small set of sparse autoencoder features in LLMs drives shifts between generous and selfish allocations in dictator games, with causal patching and steering confirming their role and generalization to other social games.
-
Towards Proactive Information Probing: Customer Service Chatbots Harvesting Value from Conversation
PROCHATIP is a chatbot system that learns to probe conversations for target business information at the right times, outperforming baselines in both data collection and user experience.
-
Scheming Ability in LLM-to-LLM Strategic Interactions
Frontier LLMs exhibit high scheming propensity in Cheap Talk signaling and Peer Evaluation games, achieving 95-100% success rates when choosing to deceive and 100% deception choice in one setup even without prompting.
-
Inference Scaling Laws: An Empirical Analysis of Compute-Optimal Inference for Problem-Solving with Language Models
Empirical analysis shows scaling inference compute via strategies like tree search can be more efficient than scaling model parameters, with 7B models plus novel search outperforming 34B models.
-
A Technical Taxonomy of LLM Agent Communication Protocols
Creates a five-dimension taxonomy (counterparty, payload, interaction state, discovery mechanism, schema flexibility) from nine protocols and identifies architectural patterns plus convergence trends.
-
Preference Estimation via Opponent Modeling in Multi-Agent Negotiation
A method combining LLM-extracted qualitative cues with Bayesian belief tracking improves full agreement rates and preference estimation accuracy in multi-agent negotiations.
-
MAC: Masked Agent Collaboration Boosts Large Language Model Medical Decision-Making
MAC framework selects Pareto-optimal LLM agents and masks low cross-consistency outputs for adaptive collaboration in medical decision-making.
-
TrustLLM: Trustworthiness in Large Language Models
TrustLLM defines eight trustworthiness principles, creates a six-dimension benchmark, and evaluates 16 LLMs showing proprietary models generally lead but some open-source ones are close while over-calibration can hurt utility.
-
Encouraging Divergent Thinking in Large Language Models through Multi-Agent Debate
Multi-agent debate with tit-for-tat arguments and a judge LLM improves reasoning by preventing LLMs from locking into incorrect initial solutions.
-
PRISMA: Preference-Reinforced Self-Training Approach for Interpretable Emotionally Intelligent Negotiation Dialogues
PRISMA augments self-training with direct preference optimization and an emotion-aware negotiation strategy chain-of-thought to produce more interpretable and effective negotiation dialogues on two new datasets.
-
The Rise and Potential of Large Language Model Based Agents: A Survey
The paper surveys the origins, frameworks, applications, and open challenges of AI agents built on large language models.