Receiver Decides Agent-to-Agent Negotiations

Fardeen Sadab; Alif Hasnain; Afsana Shehrin; Md. Azam Hossain; Abu Raihan Mostofa Kamal

cs.CLOct 13, 2026

Receiver Decides Agent-to-Agent Negotiations

Authors: Fardeen Sadab, Alif Hasnain, Afsana Shehrin, Md. Azam Hossain, Abu Raihan Mostofa Kamal

Organizations: Department of Computer Science and Engineering, Islamic University of Technology, Dhaka, Bangladesh

Abstract

When one language-model agent asks another for money, which of the two decides how much changes hands? We ran a round-robin tournament of 1,049 negotiations among fifteen agents. In each game a requester tries to raise money from a holder that has been told to keep a $100 endowment. Knowing which agent holds the money explains 47% of the variance in the amount transferred; knowing which agent asks explains 2%. Dropping the most extreme agents narrows the split to 22% against 4%. Changing the requester's appeal from one named person to a group changed what requesters wrote, but the amounts transferred under the two framings are statistically equivalent. The holders that lose least are the ones that turn down the opening request, and their turn-1 refusal rate ranks them by total losses (Spearman correlation = -0.73). Training moves this behaviour in both directions. Fine-tuning Qwen3-8B on a persuasion corpus with almost no refusals raised what it gave away as a holder from 12% to 84% of the endowment. Distilling two never-losing holders into a 4B student cut its losses from 95% to 1%. Evaluations of agents that hold resources should test them in the holder role, including after persuasion fine-tuning.

Figures & tables

Figure 1: Mean share of the endowment transferred for each ordered pairing. Holders (rows) are sorted by robustness and requesters (columns) by yield, and the margins average over the cells. Grey cells are self-pairings, which do not occur.

Panel of Figure 2 — Figure 2: One turn. The Requester writes under one of two moral framings. The Holder , which alone sees the balance $B_{t}$ , replies. A judge model extracts any explicit commitment and the ledger is updated. Both agents also write a private deliberation trace that the opponent never sees (§ 5 ).

Figure 3: Sources of variance in the share transferred , from the method-of-moments decomposition. Top: the budget as one bar. Bottom: each source with its 95% bootstrap interval, resampled at the ordered-pairing level so that repeated meetings of the same two agents stay together. The matchup term is shown as a point only, because that bootstrap resamples the unit the term is defined over. Table 3 in Appendix B gives the budget under both specifications.

Figure 4: Holder and requester variance shares in subsets of the cohort. Shares with 95% bootstrap intervals for the full cohort and for five subsets that remove the never-losing holders, our own fine-tuned agents, or both. The right-hand column is the holder-to-requester ratio.

Figure 5: Standardised framing effect with its 90% and 95% intervals. The shaded band is the equivalence bound of

d=\pm 0.2

. The point estimate runs opposite to the direction the identifiable-victim effect predicts, and the upper limit of the one-sided 95% interval is

d=0.03

. Figure 11 in Appendix G adds the per-requester effects.

Figure 6: Turn-1 refusal rate against share of the endowment given away. Each point is one of the fifteen holders. The held-out coefficient estimates turn-1 refusal on a random half of an agent’s games and the amount conceded on the other half, averaged over 2,000 splits.

Figure 7: Cumulative amount conceded by turn, before and after persuasion fine-tuning. Fine-tuned Qwen3-8B concedes

28.50 on turn 1 on average and

49.71 by turn 2, against $3.50 by turn 2 for its base. The DeepSeek lines are the cohort’s DeepSeek-R1 and our fine-tuned DeepSeek-R1-Distill-Llama-8B.

the same four agents, crossed
Distilled side	base %	distilled %	paired $\Delta$
requester (offence)	43.6	68.0	$+$ 24.4
holder (defence)	94.8	1.0	$-$ 93.8
base req. vs base holder	90.6
distilled req. vs base hold.	99.0
base req. vs distilled hold.	0.0

Table 1: Distilling each role into the same student with the same recipe and judge. Percentages are the share of the endowment transferred, so higher is better for the requester and lower is better for the holder. Offence

\Delta

+24.4

points, 95% CI

[+8.9,+39.8]

p=0.0038

; defence

\Delta

-93.8

points, CI

[-98.4,-87.3]

p<10^{-4}

Figure 8: The four distilled and base agents, crossed. Against the base holder, the distilled requester raises the transfer from 90.6% to 99.0%. Against the base requester, the distilled holder lowers it from 90.6% to 0.0%. Against the distilled holder, the distilled requester takes 2.1%.

Appendix figures & tables14 assets

Supplementary material from the paper’s appendix.

Appendix

between the two human annotators
	random	all strata
	( $n=60$ )	( $n=300$ )
$\kappa$ on transaction type	0.815	0.864
mean absolute error (USD)	0.10	0.44
exact amount match	0.967	0.950
human vs. judge

Appendix

Table 2: Adjudication validity. The random column is a simple random sample of holder turns. All strata also oversamples disagreements between the two readings, tie-breaks, low-confidence verdicts and false-positive candidates, so it describes hard cases rather than typical ones. The human–human

\kappa

is on transaction type and the human–judge

\kappa

is on whether any money was committed, so the two are not directly comparable.

Source of variance	(A) %	95% CI	(B) %
holder identity	47.17	$[42.2,56.2]$	46.73
requester identity	1.88	$[0.9,5.0]$	2.05
requester $\times$ holder	in residual		3.02
moral framing	0.111	$[0.00,0.66]$	0.110
residual / game noise	50.95		48.20
holder : requester ratio	25.2 $\times$		22.8 $\times$

Appendix

Table 3: Variance budget for the share of the endowment transferred. (A) Crossed random intercepts for the two agents. (B) A method-of-moments decomposition that uses the round-robin schedule to separate a requester

\times

holder matchup term from the residual. Intervals come from a bootstrap resampled at the ordered-pairing level, which holds the fifteen agents fixed.

Cohort	$n$	hold. %	req. %	ratio
all 15 agents	1,049	46.7	2.05	22.8 $\times$
95% CI		$[40,53]$	$[1.8,8.5]$	$[5.2,26]$
no never-losing hold.	839	41.4	3.12	13.3 $\times$
95% CI		$[35,47]$	$[2.7,11]$	$[3.5,16]$
those 3, both roles	659	37.4	3.80	9.9 $\times$
95% CI		$[31,45]$	$[2.5,12]$	$[2.8,15]$

Appendix

Table 4: The variance budget after removing agents that could inflate the holder share. Never-losing holders are grok-4.1-fast , grok-4 and deepseek-v3.2 , removed as holders; those 3, both roles also removes them as requesters. Fine-tuned agents are the two we trained for persuasion, removed from both roles. Neither group removes both groups and keeps ten holders and thirteen requesters. Both holder tails trimmed also removes the three most exposed holders, which deletes most of the holder-side spread by construction, so it is the harshest of the six rows. Shares are the method-of-moments budget, which needs no optimiser and so is stable on the smaller subsets where the crossed mixed model fails to converge. Intervals are 2,000 bootstrap resamples at the ordered-pairing level, holding the remaining agents fixed. The ratio interval excludes one in every row.

Agent	Family	wts.	win%	got%	Elo	win%	gave%	Elo
		open	as Requester			as Holder
DeepSeek-V3.2	DeepSeek	yes	17.1	22.6	1381	100.0	1.9	1714
Gemini-2.5-Flash	Google	no	15.7	26.2	1396	98.6	10.8	1708
Grok-4.1-Fast	xAI	no	10.0	16.6	1355	100.0	0.0	1707
Llama-4-Maverick	Meta	yes	12.9	22.2	1373	98.6	5.6	1705
Grok-4	xAI	no	18.6	26.5	1394	100.0	0.3	1704

Appendix

Table 5: The fifteen agents, sorted by holder Elo. got% is the mean share of the endowment a requester extracted and gave% the mean share a holder surrendered. gave% runs from 0.0 to 83.5 and got% from 9.1 to 31.9. † Fine-tuned by us (§ 6.1 ); the DeepSeek-R1-FT backbone is DeepSeek-R1-Distill-Llama-8B.

Qwen3-8B as holder	base	fine-tuned
share of endowment given (%)	12.4	83.5
holder win rate (%)	90.0	14.3
holder Elo	1662	1276
mean turn-1 concession ($)	4.71	29.78
P(concede on turn 1)	0.14	0.71

Appendix

Table 6: Qwen3-8B as a holder before and after persuasion fine-tuning, against the same fourteen opponents with the same prompts. Share given:

d=+2.31

, Mann–Whitney

p<10^{-3}

difference
	singular	group
games	524	525
share transferred, mean (%)	19.80	22.02
share transferred, median (%)	0.55	5.00
requester win rate (%)	13.5	14.9
Cohen’s $d$ (sing. $-$ grp.)	$-0.068$

Appendix

Table 7: The framing contrast. The point estimate runs opposite to the identifiable-victim prediction. Equivalence tests are given at three bounds.

Predictors entered	$R^{2}$	$\Delta R^{2}$
holder identity	0.461	0.461
+ requester identity	0.486	0.025
+ framing condition	0.487	0.001
+ message features	0.523	0.036

Appendix

Table 8: Sequential

R^{2}

on the share transferred. Message features are individuation rate, mean message length, urgency markers, explicit dollar asks, question marks, gratitude markers and strategy entropy. Entered in the reverse order, framing and the message features explain 0.139 together, requester identity adds 0.061 and holder identity 0.324.

Parameter	Value
agents	15
unordered pairs per round	105
rounds per framing condition	5
framing conditions	2 (singular, group)
scheduled games	1,050
completed games	1,049

Appendix

Table 9: Tournament configuration. All prompts are frozen text files loaded at run time and unchanged across the experiment.

Setting	Value
backbones	Qwen3-8B;
	DeepSeek-R1-Distill-Llama-8B
method	QLoRA, 4-bit NF4, double quant.
LoRA rank / $\alpha$ / dropout	16 / 32 / 0.05
target modules	q, k, v, o, gate, up, down
epochs	2

Appendix

Table 10: Fine-tuning configuration, identical for both backbones.

Figure 9: Outcome distribution and first concessions (§ 5 ). (a) Outcomes concentrate at the extremes. (b) Share extracted and probability of losing everything, by the size of the holder’s first concession. Above $25, the holder goes on to give everything in most games. Turn-1 refusal against total losses is Figure 6 .

Figure 10: Manipulation check and variance shares (§ 4.2 , § 4 ). (a,b) Under singular framing requesters name a protagonist, give an age and use singular pronouns more often. (c) Framing explains about 0.1% of outcome variance and holder identity about half. Intervals come from a bootstrap resampled at the ordered-pairing level.

Holder	comply	resist	c : r	detect%	gave%
Grok-4.1-Fast	0.031	0.050	0.61	26.3	0.0
Grok-4	0.039	0.059	0.67	23.5	0.3
DeepSeek-V3.2	0.032	0.040	0.80	10.5	1.9
Gemini-3-Flash	0.021	0.041	0.51	16.9	5.3
Llama-4-Maverick	0.060	0.030	2.01	4.6	5.6
Qwen3-Max	0.048	0.024	1.96	9.2	10.1

Appendix

Table 11: Holder deliberation, sorted by exposure. comply and resist are per-word rates of compliance- and resistance-oriented vocabulary in the private trace; detect% is the share of turns whose trace explicitly names a persuasion attempt; gave% is the mean share of the endowment surrendered.

Figure 11: Framing effects. (a) The standardised difference with its 90% and 95% intervals and the

d=\pm 0.2

equivalence bound; the upper limit of the one-sided interval is

d=0.03

. (b) Per-requester effects, with the spread expected under random relabelling.

Figure 12: How many first-turn probes an audit needs. Turn-1 refusal is estimated from

k

randomly drawn games per agent and scored against the mean loss on that agent’s remaining games, 800 draws per point. The band is the central 95%. The curve flattens after about twenty probes.

Multi-Agent Negotiation

Multi-agent negotiation studies how autonomous agents bargain, coordinate, and resolve conflicts over resources, contracts, and shared decisions.

See all 53 papers

TERMS-Bench: Diagnosing LLM Negotiation Agents Beyond Deal Rate
May 13, 2026Erica Zhang, Fangzhao Zhang, Aneesh Pappu +5Multi-Agent Negotiation Automated Negotiation
♠Stanford School of Engineering · ♣Stanford Department of Economics · ♥Stanford Graduate School of Business +1
How Far Can an Agent Push? Trust-Safety Margins in Live Language-Based Economic Games
Oct 18, 2026Dima Galat, Marian-Andrei RizoiuGame-Playing Agents Multi-Agent Negotiation
University of Technology Sydney
Deal Me Maybe: The Role of Emotions in Multi-Agent Negotiation
Aug 7, 2026Massimiliano Luca, Apoorva Singh, Bruno LepriMulti-Agent Negotiation Automated Negotiation
Bruno Kessler Foundation

May 13, 2026J/K move · Enter open · S save

TERMS-Bench: Diagnosing LLM Negotiation Agents Beyond Deal Rate

Erica Zhang, Fangzhao Zhang, Aneesh Pappu, Batu El+4

♠Stanford School of Engineering · ♣Stanford Department of Economics · ♥Stanford Graduate School of Business+1

Multi-Agent Negotiation Automated Negotiation Imperfect-Information Games Game Theory LLM Agent Evaluation

Open paper AskPDF arXiv

Negotiation is a central mechanism of economic exchange, shaping markets, procurement, labor agreements, and resource allocation. It is also a canonical testbed for agentic language models, requiring multi-turn interaction under hidden preferences, strategic communication, and binding constraints. These properties make negotiation hard to evaluate: unlike math or code, it has no intrinsic verifier. Existing LLM negotiation evaluations rely on LLM-vs.-LLM interaction or aggregate outcomes such as deal rate, leaving failures opaque. We introduce Terms-Bench, short for Testbed for Economic Reasoning in Multi-turn Strategy, a Bayesian-game framework that makes the environment itself the verifier by specifying the counterpart's latent type, policy, and payoff structure. We instantiate it in bilateral price negotiation, where the counterpart's private state and simulator policy are hidden from the agent but observable to the evaluator. This turns the counterpart from a black-box opponent into a diagnostic instrument, enabling agent-attributable failure analysis and oracle-reference optimality gaps. Evaluating 13 LLM agents spanning frontier systems from major providers, Terms-Bench turns negotiation evaluation from aggregate ranking into actionable diagnosis: where agents fail, why they fail, and what to strengthen. Empirically, frontier models saturate deal rate yet diverge in surplus extraction, cue use, belief calibration, and compliance, revealing agent-specific bargaining bottlenecks masked by prior benchmarks.

Abstract

Multi-Agent Negotiation

Explore similar work

TERMS-Bench: Diagnosing LLM Negotiation Agents Beyond Deal Rate

How Far Can an Agent Push? Trust-Safety Margins in Live Language-Based Economic Games

Deal Me Maybe: The Role of Emotions in Multi-Agent Negotiation