🤖 AI

TERMS-Bench: Diagnosing LLM Negotiation Agents Beyond Deal Rate

This paper introduces Terms-Bench, a Bayesian-game framework that transforms negotiation evaluation from aggregate deal-rate rankings into actionable diagnostic analysis by using a known counterpart simulator to pinpoint specific agent failures in surplus extraction, belief calibration, and strategic compliance.

Original authors: Erica Zhang, Fangzhao Zhang, Aneesh Pappu, Batu El, Jose Blanchet, Susan Athey, Jiashuo Liu, James Zou

Published 2026-06-16

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Erica Zhang, Fangzhao Zhang, Aneesh Pappu, Batu El, Jose Blanchet, Susan Athey, Jiashuo Liu, James Zou

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a robot how to be a master negotiator. In the past, researchers tested these robots by having them haggle with other robots. They would just look at the final result: "Did they make a deal? Yes or No?"

The problem with this approach is like judging a chef only by whether the customer ate the meal, without tasting the food. If the robot made a deal, but gave away all the profit because it was too scared to ask for a fair price, the old test would still say, "Great job! Deal made!" It missed the fact that the robot was actually terrible at the strategy part of the job.

This paper introduces a new, smarter way to test these AI negotiators, called TERMS-BENCH. Here is how it works, using simple analogies:

1. The "Scripted Opponent" (The Environment as a Verifier)

In the old tests, the robot negotiated against another "black box" AI. You didn't know what that other AI was thinking, so you couldn't tell if your robot failed because it was bad, or because the other AI was weird.

In TERMS-BENCH, the researchers created a simulated opponent that is like a strict, predictable actor following a hidden script.

The Script: The opponent has a secret "reservation price" (the lowest it will sell for or highest it will buy for), a "urgency level" (how much time pressure it feels), and a "personality" (aggressive, friendly, or neutral).
The Twist: The AI agent doesn't know these secrets. It has to guess them based on what the opponent says and does.
The Verifier: The researchers do know the secrets. Because they know the script, they can look at the AI's performance and say exactly why it failed. Did it fail because it couldn't guess the opponent's price? Did it fail because it got confused by the opponent's friendly tone? Or did it just make a bad move even though it knew the right answer?

2. The "Six Personality Types" of Opponents

To really stress-test the AI, the researchers didn't just use one type of opponent. They created six different "families" of opponents, like different characters in a video game:

The Honest One (Candid): Says exactly what it means. If it's friendly, it acts friendly.
The Silent One (Taciturn): Has the same economic rules as the Honest one, but says very little. It tests if the AI can figure things out without hints.
The Reactive One (Expressive): Changes its behavior based on how the AI acts. If the AI is pushy, this opponent pushes back harder.
The Strategist (Strategic): Acts like the Reactive one but hides its true feelings in its words. It's a master of disguise.
The Chaotic One (Stochastic): Throws in random noise and confusing signals to see if the AI panics.
The Bully (Adversarial): Always tries to intimidate the AI.

By testing the AI against all these different "characters," the researchers can pinpoint exactly which skill the AI is missing.

3. The "Magic Glasses" (Oracle Interventions)

This is the most clever part of the paper. Sometimes an AI fails, but we don't know if it's because it's stupid (can't figure out the opponent) or clumsy (knows the answer but makes a bad move).

The researchers use a "magic glasses" trick:

Base Test: The AI negotiates normally, guessing the opponent's secrets.
The "Posterior" Glasses: They give the AI a cheat sheet that says, "Here is the exact probability of what the opponent is thinking." If the AI still fails, it means it's bad at acting on information, not bad at getting it.
The "Revealed Type" Glasses: They tell the AI the opponent's exact secret price and personality. If the AI still doesn't get a good deal, it means the AI is terrible at making the right moves even when it has perfect information.

This allows them to break down the failure into three parts:

Inference Failure: "You couldn't guess the opponent's price."
Uncertainty Failure: "Even with a good guess, you were too unsure to act."
Control Failure: "You knew the answer, but you made a clumsy move."

4. What They Found

When they tested 13 of the smartest AI models available (like Claude, GPT-5, and Gemini), they found some surprising things:

The "Deal Rate" Lie: Almost all the AIs were great at making a deal (95%+ success rate). But the old tests stopped there.
The Real Gap: When they looked at how much profit the AI made, the results were all over the place. Some models made great deals; others made deals but gave away almost all the money.
The "Cue" Trap: Many AIs got tricked by the opponent's tone. If the opponent sounded friendly or pressured, the AI would often give up too much money, even if the math said it shouldn't.
Different Bottlenecks: Some AIs were bad at guessing the opponent's mind. Others were great at guessing but bad at making the final move.

The Bottom Line

The paper argues that we need to stop just counting "deals made" and start diagnosing how the deal was made. TERMS-BENCH is like a medical scan for AI negotiators. Instead of just saying "Patient is healthy" (Deal Made), it tells you exactly which organ is failing (e.g., "The AI is bad at reading emotional cues" or "The AI can't handle time pressure").

This helps developers know exactly what to fix: Do they need to train the AI to be less emotional? Do they need to teach it to ignore pressure? Or do they need to teach it to make bolder moves? It turns a simple ranking list into a detailed instruction manual for building better AI.

Technical Summary: TERMS-BENCH

Problem Statement

Negotiation serves as a central mechanism for economic exchange and a canonical testbed for agentic language models (LLMs), requiring multi-turn interaction under hidden preferences, strategic communication, and binding constraints. However, evaluating LLM negotiation agents presents a unique methodological challenge: unlike math or code, negotiation lacks an intrinsic verifier. Existing evaluation paradigms rely on LLM-vs.-LLM interactions or aggregate outcome metrics such as deal rate. This approach suffers from two structural limits:

Opacity of Attribution: Because the counterpart is a black-box policy with unspecified latent preferences, outcomes conflate agent competence with counterpart variability, preventing the attribution of failures to specific reasoning errors.
Loss of Diagnostic Granularity: Even metrics validated by human preference or multi-component scoring collapse performance into scalar rankings, obscuring where in the negotiation workflow an agent succeeds or fails (e.g., inference vs. control).

Consequently, current benchmarks fail to diagnose whether an agent's failure stems from unstable reasoning, prompt sensitivity, or latent behavioral variability, leaving failures opaque to practitioners.

Methodology: TERMS-BENCH

The authors introduce TERMS-BENCH (Testbed for Economic Reasoning in Multi-turn Strategy), a framework that constructs the environment itself as the verifier. Rather than using another LLM as a judge, TERMS-BENCH specifies the counterpart's latent type, policy, and payoff structure explicitly, making the environment observable to the evaluator while remaining hidden from the agent.

1. Framework Formulation

TERMS-BENCH models negotiation as an extensive-form, incomplete-information Bayesian game:

Environment ( $\Gamma$ ): Defined by a set of feasible outcomes, private type spaces for the agent ( $T_A$ ) and counterpart ( $T_B$ ), utility functions, and a prior over latent types ( $\mu$ ). A type $t_i$ encodes a reservation value ( $r_i$ ), outside option ( $e_i$ ), urgency ( $\kappa_i$ ), and strategic stance ( $\eta_i$ ).
Protocol ( $P$ ): Fixes move ordering, action space (economic acts paired with natural language), information revelation rules, and termination conditions.
Agent Policy ( $\pi$ ): Maps information states (history, private context, beliefs) to actions.

2. Instantiation: Bilateral Price Negotiation

The paper instantiates this framework in a bilateral price negotiation setting where an agent negotiates a scalar price with a simulator-driven counterpart.

Counterpart as Diagnostic Instrument: The counterpart is governed by a fixed stochastic kernel $\pi_B$ parameterized by a sampled latent type $t_B$ . The agent must infer $t_B$ from observed prices and messages.
Counterpart Families: To isolate specific capabilities, the authors define six counterpart behavior families varying along two axes:
- Economic Reactivity: How much behavior depends on the agent's recent offers vs. latent type (e.g., Type-Instrumental vs. High-Reactivity).
- Cue Reliability: Whether language cues (sentiment, posture) accurately reflect the latent stance (e.g., Accurate vs. Uninformative vs. Noisy vs. Pressuring).
- The core families are Candid, Taciturn, Expressive, and Strategic, with stress tests Stochastic and Adversarial.

3. Evaluation Metrics and Attribution

TERMS-BENCH moves beyond deal rate to a multi-axis diagnostic framework:

Terminal Value: Surplus efficiency ( $SE^+_\pi$ ), normalized by the Zone of Possible Agreement (ZOPA).
Agreement Calibration: Rates of agreement in feasible episodes ( $AGR^+_\pi$ ) and false agreements in infeasible episodes ( $FAGR^-_\pi$ ).
Opponent Modeling: Belief error regarding reservation value, urgency, and stance.
Protocol Compliance: Rates of critical violations (price bounds, individual rationality).

Oracle Interventions: To attribute failures, the framework compares the agent's performance against three conditions:

Base: The agent infers the type from scratch.
Oracle-Posterior: The agent receives the exact Bayesian posterior of the counterpart's type.
Revealed-Type: The agent receives the true latent type.
Model-Based Oracle: A dynamic-programming reference policy ( $\pi^*$ ) that is Bayes-optimal for the specified environment.

This allows the decomposition of the performance gap into Inference ( $\Delta_{inf}$ ), Uncertainty ( $\Delta_{unc}$ ), and Control ( $\Delta_{ctrl}$ ) bottlenecks.

Key Results

The authors evaluated 13 frontier LLM agents (including Claude Opus 4.6/4.7, GPT-5.5, Gemini 3.1 Pro, etc.) and three fixed-concession baselines across 1,800 seeded episodes.

High Deal Rate Masks Value Extraction: While most frontier models achieve near-saturation deal rates ( $>97\%$ in feasible episodes), they diverge significantly in surplus extraction. For instance, Claude Opus 4.6 achieved a surplus efficiency of 0.694, while Doubao-Seed-2.0-Pro achieved 0.522 despite a higher deal rate (99.9%).
Cue-Use Failure: Agents consistently extracted less surplus when informative cues were present compared to when they were muted ( $\alpha_{cue} < 0$ ). This suggests agents may over-concede to "warm" cues or exhibit brittle behavior under "pressure" cues, failing to decouple linguistic signals from economic strategy.
Information-Action Gap: Frontier models failed to convert latent-type inference into surplus.
- Posterior Injection: Providing the exact Bayesian posterior to agents yielded weak or inconsistent returns, closing only ~1% of the gap for top models.
- Belief Calibration: Belief errors did not improve over time; in some cases, they increased, with stance inference being the dominant source of error.
- Inference Penalty: Agents often performed worse in families where latent-type inference was economically instrumental compared to high-reactivity families.
Distinct Bottlenecks via Oracle Decomposition:
- GPT-4o-mini was primarily control-limited (98% of the gap), struggling to act on revealed information.
- Claude Opus 4.7 and Grok 4.20 were uncertainty-limited, with large gaps remaining even after type revelation.
- Claude Opus 4.6 showed a mixed profile across inference, uncertainty, and control.
Stable Bargaining Fingerprints: Agents exhibited stable behavioral typologies (e.g., "anchor-and-hold," "mid/balanced") and self-constraint profiles (IR violation rates vs. exit discipline) that were consistent across counterpart families, suggesting these are intrinsic model properties rather than environmental artifacts.
Data-Grounded Validation: When tested on a data-grounded variant using Amazon catalog statistics, the rank ordering of agents was largely preserved (Spearman $\rho=0.90$ ), but the "difficulty multiplier" effect amplified the performance gap between strong and weak models. Stronger models leveraged public product anchors effectively, while weaker models struggled with wider price ranges.
Commercial Impact: In a "Bankroll Mode" simulation chaining multiple episodes, small per-period errors compounded into significant solvency failures. The terminal balance spread between the strongest and weakest LLMs was ~21x, far exceeding the ~14x spread in single-episode surplus efficiency.

Significance and Claims

The paper claims that TERMS-BENCH reframes negotiation evaluation from a black-box ranking exercise to an agent-attributable diagnostic tool. By making the environment the verifier, it enables:

Scalable, Human-Annotation-Free Diagnosis: Identifying specific failure modes (inference, uncertainty, control) without relying on subjective human judges.
Unmasking Hidden Failures: Revealing that high deal rates often mask critical deficiencies in surplus extraction, cue sensitivity, and self-constraint.
Actionable Insights for Practitioners: Providing a controlled foundation to show where agents break down and what to strengthen (e.g., improving belief calibration vs. strategic control).

The authors position this work as a necessary step toward deployable negotiation agents, noting that while the current instantiation focuses on bilateral price negotiation, the framework is generalizable to multi-party and multi-product settings. They acknowledge that human studies and richer multi-issue settings remain important next steps for validating transfer beyond the current scope.