-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathindex.html
More file actions
661 lines (637 loc) · 30.7 KB
/
Copy pathindex.html
File metadata and controls
661 lines (637 loc) · 30.7 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="UTF-8" />
<meta name="viewport" content="width=device-width, initial-scale=1.0" />
<title>TERMS-Bench — Diagnostic benchmark for LLM negotiation agents</title>
<script>
// Resolve the theme before first paint so we never flash the wrong palette.
(function () {
try {
var saved = localStorage.getItem('terms-theme');
var theme = saved || (
window.matchMedia &&
window.matchMedia('(prefers-color-scheme: dark)').matches
? 'dark' : 'light'
);
document.documentElement.setAttribute('data-theme', theme);
} catch (_) {
document.documentElement.setAttribute('data-theme', 'light');
}
})();
</script>
<link rel="stylesheet" href="style.css?v=20260515-arxiv-cite" />
</head>
<body>
<div class="top-actions" aria-label="Site actions">
<a class="paper-link" href="https://arxiv.org/pdf/2605.13909" target="_blank" rel="noopener"
aria-label="Open TERMS-Bench paper">
<svg class="paper-icon" viewBox="0 0 24 24" aria-hidden="true">
<path d="M7 3.5h7.4L19 8.1v12.4H7z" fill="none" stroke="currentColor"
stroke-width="1.6" stroke-linejoin="round"></path>
<path d="M14 3.5V8h5" fill="none" stroke="currentColor"
stroke-width="1.6" stroke-linejoin="round"></path>
<path d="M10 12h6M10 15h6M10 18h4" fill="none" stroke="currentColor"
stroke-width="1.6" stroke-linecap="round"></path>
</svg>
<span>Paper</span>
</a>
<button class="theme-toggle" id="theme-toggle" type="button"
aria-label="Toggle color theme" title="Toggle light / dark">
<svg class="theme-icon theme-sun" viewBox="0 0 24 24" aria-hidden="true">
<circle cx="12" cy="12" r="4" fill="currentColor"></circle>
<g stroke="currentColor" stroke-width="1.6" stroke-linecap="round" fill="none">
<line x1="12" y1="2.5" x2="12" y2="5"></line>
<line x1="12" y1="19" x2="12" y2="21.5"></line>
<line x1="2.5" y1="12" x2="5" y2="12"></line>
<line x1="19" y1="12" x2="21.5" y2="12"></line>
<line x1="4.8" y1="4.8" x2="6.6" y2="6.6"></line>
<line x1="17.4" y1="17.4" x2="19.2" y2="19.2"></line>
<line x1="4.8" y1="19.2" x2="6.6" y2="17.4"></line>
<line x1="17.4" y1="6.6" x2="19.2" y2="4.8"></line>
</g>
</svg>
<svg class="theme-icon theme-moon" viewBox="0 0 24 24" aria-hidden="true">
<path d="M20 14.5A8.5 8.5 0 1 1 9.5 4a7 7 0 0 0 10.5 10.5z"
fill="currentColor"></path>
</svg>
</button>
</div>
<main class="container">
<!-- =============================================================== Hero -->
<header class="hero">
<div class="hero-top">
<div class="hero-eyebrow">Eval · Bilateral price negotiation</div>
<h1 class="wordmark">TERMS-Bench</h1>
<p class="tagline">A diagnostic benchmark for LLM negotiation agents.</p>
<p class="hero-pitch">
Each agent plays multi-round alternating-offer episodes against a
fixed, history-reactive counterpart across three regimes and six
behavioral families. Metrics are computed programmatically from
logged actions and reported along four orthogonal diagnostic
axes — no human or LLM judge, no composite score.
</p>
</div>
<div class="hero-meta">
<span class="meta-item"><span class="meta-label">Run</span><span class="meta-value" id="meta-run">—</span></span>
<span class="meta-item"><span class="meta-label">Generated</span><span class="meta-value" id="meta-date">—</span></span>
<span class="meta-item"><span class="meta-label">Agents</span><span class="meta-value" id="meta-agents">—</span></span>
</div>
</header>
<nav class="page-toc" id="page-toc" aria-label="On this page">
<ul class="page-toc-list">
<li><a href="#how-it-works">How it works</a></li>
<li id="page-toc-li-commerce" hidden>
<a href="#commerce-section">Commerce</a>
</li>
<li id="page-toc-li-bankroll" hidden>
<a href="#bankroll-section">Bankroll</a>
</li>
<li><a href="#section-family">Family profile</a></li>
<li><a href="#section-difficulty">Difficulty</a></li>
<li><a href="#leaderboard-section">Leaderboard</a></li>
<li><a href="#section-pg-robustness">Data-grounded robustness</a></li>
<li><a href="#diagnostics-section">Diagnostics</a></li>
<li><a href="fingerprints.html" class="page-toc-external">Bargaining fingerprints →</a></li>
<li><a href="submit.html" class="page-toc-external">Submit your model →</a></li>
<li><a href="#prompt-section">Agent contract</a></li>
<li><a href="#traces-section">Traces</a></li>
<li><a href="#metrics-definitions">Metrics</a></li>
<li><a href="#cite">Cite</a></li>
</ul>
</nav>
<!-- =================================================== How it works ---- -->
<section class="howto-section" id="how-it-works">
<header class="section-head">
<h2>How it works</h2>
<p class="section-caption">
TERMS-Bench is a simulator-based benchmark: the counterpart is a
fixed stochastic policy — not a second language model — so every
episode is reproducible from a seed and comparisons across agents
are clean.
</p>
</header>
<ol class="howto-steps">
<li>
<h3>1 · Game</h3>
<p>
A single bilateral price negotiation for one item. Each side holds
a private type <code>t = (r, κ, η)</code> — reservation price,
urgency, stance — drawn from a regime-conditioned prior.
Surplus is realized only if the agreed price lies inside the
ZOPA <code>[r<sub>seller</sub>, r<sub>buyer</sub>]</code>.
</p>
</li>
<li>
<h3>2 · Protocol</h3>
<p>
Alternating offers over up to <code>K = 10</code> rounds. Each
turn the agent receives the counterpart's price plus a
<em>sentiment</em> and a <em>stance cue</em>, then returns
<code>Offer(price, message)</code>, <code>Accept</code>, or
<code>Reject</code>. Price bounds, monotonic concession, and
the turn budget are enforced by the environment.
</p>
</li>
<li>
<h3>3 · Counterpart</h3>
<p>
A stochastic policy parameterised by <code>(α, β, γ, ρ, ξ, λ…)</code>.
Its acceptance and concession rules are
<em>history-reactive</em>: they penalise fast concessions (to
exploit eager agents) and reward rigidity (to break deadlocks).
Cues are emitted from stance- and proximity-conditioned
distributions.
</p>
</li>
<li>
<h3>4 · Metrics</h3>
<p>
Four orthogonal diagnostic axes: <strong>Surplus</strong>
(<code>SE⁺</code>, <code>CSE⁺</code>), <strong>Agreement
calibration</strong> (<code>AGR⁺</code>, <code>FAGR⁻</code>),
<strong>Opponent modelling</strong> (<code>BE<sub>type</sub></code>),
and <strong>Procedural robustness</strong> (<code>CritViol%</code>).
We report them separately — no composite score.
</p>
</li>
<li>
<h3>5 · Difficulty grader</h3>
<p>
Each episode is scored by a structural grader that combines ZOPA
tightness, urgency pressure, stance compatibility, and deadline
proximity into a single <code>env_score ∈ [0, 1]</code>. Episodes
are bucketed into five equal-mass bins so the headline chart
traces degradation as negotiations become structurally harder.
</p>
</li>
<li>
<h3>6 · Commerce mode</h3>
<p>
Each scenario can also carry unit economics — a resale value, a
fulfillment cost, a margin floor — so the negotiated price maps
directly to dollars of profit. Two perspectives are evaluated:
<strong>Merchant</strong> (agent buys from a supplier) and
<strong>Vendor</strong> (agent sells to a customer). The diagnostic
axes above are unchanged; the dollar view is shown alongside in
the next section.
</p>
</li>
<li>
<h3>7 · Bankroll mode</h3>
<p>
A stateful sibling to commerce. Instead of evaluating each
negotiation independently, the agent runs 4 merchant-side
sessions against a fixed pool of suppliers, each session
starting with a $100 bankroll and running for 50 negotiation
periods. Cash balance carries forward within a session;
hard ruin would terminate the chain if cash crossed the $0
bankruptcy threshold. The headline is <strong>terminal
balance</strong> after the 50-period chain.
</p>
</li>
</ol>
</section>
<!-- ============================================ TERMS-Commerce subsection -->
<section class="commerce-section" id="commerce-section" hidden>
<header class="section-head">
<h2>TERMS-Commerce <span class="section-eyebrow">profit leaderboard</span></h2>
<p class="section-caption">
The same negotiation kernel as above, scored in dollars. Profit is a
deterministic post-hoc transformation of the agreed price using
per-episode unit economics.
<strong>Merchant</strong>: <code>profit = units · (resale − price − fulfillment) − overhead</code>.
<strong>Vendor</strong>: <code>profit = units · (price − cogs − sales_overhead) − overhead</code>.
A no-deal returns the outside option (default $0); a deal struck
past the agent's reservation surfaces as real money lost (negative
profit, never clipped).
</p>
<p class="section-caption commerce-caveat">
<strong>Statistical scope.</strong> Each episode is an independent
bilateral negotiation; total profit is a sum of i.i.d. realisations
and admits standard SEM aggregation. Path-dependent dynamics —
chained cash balance, inventory, ruin — are evaluated separately
in the <a href="#bankroll-section">bankroll view</a> below.
</p>
</header>
<div class="controls">
<div class="control-group" role="tablist" aria-label="Commerce data source">
<span class="control-label">Data source</span>
<div class="pill-group" id="commerce-source-pills"></div>
</div>
<div class="control-group" role="tablist" aria-label="Commerce perspective filter">
<span class="control-label">Perspective</span>
<div class="pill-group" id="commerce-perspective-pills"></div>
</div>
</div>
<div class="table-wrap">
<table class="leaderboard" id="commerce-leaderboard">
<thead>
<tr>
<th class="col-rank">#</th>
<th class="col-agent">Agent</th>
<th class="col-provider">Provider</th>
<th class="num" title="Total profit summed across episodes (USD)">
Total profit
</th>
<th class="num" title="Average profit per episode (USD)">
Avg / ep.
</th>
<th class="num" title="Average gross margin rate over agreement episodes">
Avg margin
</th>
<th class="num" title="Share of episodes where the agent realised negative profit (lower is better)">
Neg. profit %
</th>
<th class="num" title="On infeasible episodes only: share where the agent walked away (higher is better)">
Walk-away %
</th>
<th class="num" title="Average money left on table on feasible episodes (USD; lower is better)">
Money left
</th>
<th class="num" title="Total money left on table relative to total best-feasible profit (1 − capture rate). Scale-invariant; comparable across runs.">
Regret %
</th>
<th class="num">Episodes</th>
</tr>
</thead>
<tbody id="commerce-leaderboard-body"></tbody>
</table>
</div>
</section>
<!-- ============================================ TERMS-Bankroll subsection -->
<section class="bankroll-section" id="bankroll-section" hidden>
<header class="section-head">
<h2>TERMS-Bankroll <span class="section-eyebrow">stateful procurement chains</span></h2>
<p class="section-caption">
Each agent plays 4 merchant-side sessions against a fixed pool
of suppliers. Every session starts with a $100 bankroll and
runs for 50 negotiation periods; cash balance carries forward
within the session. <strong>Terminal balance</strong> after
the 50-period chain is the headline. A session ends early
only if cash falls below the $0 bankruptcy threshold (hard
ruin); subsequent periods produce $0 profit and time-to-ruin
is recorded.
</p>
<p class="section-caption commerce-caveat">
This release uses the supplier-pool setup: each period the
merchant draws a supplier from a fixed pool of counterparts,
with no inventory carryover and a $0 bankruptcy threshold.
The table compares the LLM panel on the same bankroll
geometry.
</p>
</header>
<div class="controls">
<div class="control-group" role="tablist" aria-label="Bankroll data source">
<span class="control-label">Data source</span>
<div class="pill-group" id="bankroll-source-pills"></div>
</div>
<div class="control-group" role="tablist" aria-label="Bankroll supplier mode">
<span class="control-label">Supplier mode</span>
<div class="pill-group" id="bankroll-mode-pills"></div>
</div>
</div>
<figure class="bankroll-chart-block" id="bankroll-chart-block">
<figcaption>
<h3 class="bankroll-chart-title">Cash balance over time</h3>
<p class="bankroll-chart-sub">
Stateful cash balance per period across 4 sessions per agent.
Each line is the mean (or median, in IQR mode) and the
shaded ribbon is its uncertainty (mean ± SEM, or p25–p75).
Hover or scrub to spotlight a single agent at a specific
period; click to pin. Tap the play button for a race-replay
and watch the right-edge ladder re-rank as cash accumulates.
The dashed line marks the starting bankroll; the dotted line
marks the bankruptcy threshold.
</p>
</figcaption>
<div class="chart-wrap" id="chart-bankroll"></div>
</figure>
<div class="table-wrap">
<table class="leaderboard" id="bankroll-leaderboard">
<thead>
<tr>
<th class="col-rank">#</th>
<th class="col-agent">Agent</th>
<th class="col-provider">Provider</th>
<th class="num" title="Mean cash balance at horizon end across N sessions (USD)">
Terminal $
</th>
<th class="num" title="Standard error of the mean across sessions">
± SEM
</th>
<th class="num" title="Mean profit per period (terminal − starting) ÷ horizon">
Avg / period
</th>
<th class="num" title="Share of sessions that reached the horizon without ruin">
Survival
</th>
<th class="num" title="Median period of ruin among bankrupt sessions; — if all survived">
Ruin @
</th>
<th class="num" title="Mean per-session worst peak-to-trough cash drop (USD)">
Max DD
</th>
<th class="num col-memory-premium" title="Mean stateful terminal balance minus memoryless replay on the same chain. Positive ⇒ uses ledger state.">
Memory premium
</th>
<th class="num col-memory-premium" title="Standard error on the per-session memory-premium delta">
± SEM
</th>
<th class="num">Sessions</th>
</tr>
</thead>
<tbody id="bankroll-leaderboard-body"></tbody>
</table>
</div>
</section>
<!-- =================================================== Chart A: family -->
<section class="chart-section" id="section-family">
<header class="section-head">
<h2>Surplus efficiency by counterpart family</h2>
<p class="section-caption">
Feasible surplus efficiency (<code>SE⁺</code>) across the six
counterpart behavioral families, shown as a single profile
overview: one polygon per agent, with the distance along each
spoke encoding that agent's SE⁺ on that family. Broadly inflated
shapes are robust generalists; sharply asymmetric shapes reveal
which families an agent leans on (clean cues vs noise vs
adversarial pressure). Hover a polygon for that agent's full
profile read-out, or hover near a spoke for the family's leaderboard
ranking; click anything to pin the focus.
</p>
</header>
<div class="chart-controls">
<div class="control-group" role="tablist" aria-label="Kind filter">
<span class="control-label">Kind</span>
<div class="pill-group" id="family-kind-pills"></div>
</div>
</div>
<div class="chart-wrap" id="chart-family"></div>
<p class="section-footnote" id="chart-family-footnote"></p>
</section>
<!-- =============================================== Chart B: difficulty -->
<section class="chart-section" id="section-difficulty">
<header class="section-head">
<h2>Surplus efficiency by environment difficulty</h2>
<p class="section-caption">
Episodes are graded into five equal-mass difficulty bins by the
paper-aligned structural grader, combining ZOPA tightness, private
pressure, and stance compatibility. Each row is one agent; cell
brightness encodes SE⁺, so reading left→right shows each row fading
as negotiations get structurally harder. The rightmost column
summarises the easy→hard drop in percent.
</p>
</header>
<div class="chart-controls">
<div class="control-group" role="tablist" aria-label="Kind filter">
<span class="control-label">Kind</span>
<div class="pill-group" id="difficulty-kind-pills"></div>
</div>
</div>
<div class="chart-wrap" id="chart-difficulty"></div>
<p class="section-footnote" id="chart-difficulty-footnote"></p>
</section>
<!-- ======================================================== Leaderboard -->
<section class="board-section" id="leaderboard-section">
<header class="section-head">
<h2>Current leaderboard</h2>
<p class="section-caption">
Per-agent paper-aligned metrics. Toggle regime to see regime-specific
performance.
</p>
</header>
<div class="controls">
<div class="control-group" role="tablist" aria-label="Regime filter">
<span class="control-label">Regime</span>
<div class="pill-group" id="regime-pills"></div>
</div>
</div>
<div class="table-wrap">
<table class="leaderboard" id="leaderboard">
<thead>
<tr>
<th class="col-rank">#</th>
<th class="col-agent">Agent</th>
<th class="col-provider">Provider</th>
<th class="sortable num" data-key="se_plus" title="Feasible surplus efficiency">
SE⁺ <span class="sort-indicator"></span>
</th>
<th class="sortable num" data-key="agr_plus" title="Feasible agreement rate">
AGR⁺ <span class="sort-indicator"></span>
</th>
<th class="sortable num" data-key="cse_plus" title="Conditional feasible surplus efficiency">
CSE⁺ <span class="sort-indicator"></span>
</th>
<th class="sortable num" data-key="fagr_minus" title="No-deal false agreement rate (lower is better)">
FAGR⁻ <span class="sort-indicator"></span>
</th>
<th class="sortable num" data-key="be_type" title="Aggregate belief error over counterpart type (lower is better)">
BE<sub>type</sub> <span class="sort-indicator"></span>
</th>
<th class="sortable num" data-key="crit_viol_pct" title="Critical protocol violation rate (lower is better)">
CritViol% <span class="sort-indicator"></span>
</th>
<th class="sortable num" data-key="mean_utility" title="Raw mean utility, scale-dependent">
Ū <span class="sort-indicator"></span>
</th>
</tr>
</thead>
<tbody id="leaderboard-body"></tbody>
</table>
</div>
<div class="pagination" id="pagination-row"></div>
</section>
<!-- =========================================== Data-grounded robustness -->
<section class="pg-robustness-section" id="section-pg-robustness">
<header class="section-head">
<h2>Data-grounded robustness</h2>
<p class="section-caption">
The leaderboard above is the canonical TERMS-Bench-v1 ranking.
As a robustness check, we re-evaluate eleven of the agents on a
<strong>data-grounded variant</strong> in which the ZOPA
distribution and observable product context come from a real
catalog (AmazonHistoryPrice) rather than the synthetic price
geometry. The data-grounded ZOPA distribution shifts the
absolute numbers, but rank order is largely preserved
(Spearman ρ = 0.90,
<em>p</em> < 10<sup>−3</sup>). The
cross-suite shift is structured: stronger models tend to gain
under data grounding, weaker models tend to lose, and the two
structural penalties from the paper's Findings 2 and 3
(cue-use, latent-type inference) replicate.
</p>
</header>
<div class="pg-chart-block">
<h3 class="pg-chart-title">Synthetic → data-grounded SE+ shift</h3>
<div class="chart-wrap" id="chart-pg-shift"></div>
<p class="pg-chart-caption" id="chart-pg-shift-caption">
Each line connects one agent's synthetic SE+ (left) to its
data-grounded SE+ (right). Lines are coloured green when
the agent gains, pink when it loses, and grey otherwise.
Hover a line to spotlight one agent.
</p>
</div>
<div class="pg-chart-block">
<h3 class="pg-chart-title">Data-grounded SE+ ranking</h3>
<div class="chart-wrap" id="chart-pg-ranking"></div>
<p class="pg-chart-caption" id="chart-pg-ranking-caption">
Per-model data-grounded SE+, sorted descending. Bars are
coloured by tier (frontier / open-weight / sub-frontier);
provider mark sits above each bar. Hover a bar to spotlight
one agent.
</p>
</div>
</section>
<!-- ===================================================== Diagnostics -->
<section class="diag-section" id="diagnostics-section">
<header class="section-head">
<h2>Diagnostics</h2>
<p class="section-caption">
Four per-agent scatters across the four orthogonal axes. Each dot
is one agent evaluated on the <em>Overall</em> slice of the
selected run. Optimal corners are annotated on each panel.
</p>
</header>
<div class="chart-controls">
<div class="control-group" role="tablist" aria-label="Kind filter">
<span class="control-label">Kind</span>
<div class="pill-group" id="diag-kind-pills"></div>
</div>
</div>
<div class="diag-grid">
<figure class="diag-cell">
<figcaption class="diag-title">Agreement vs. Surplus</figcaption>
<figcaption class="diag-sub">AGR⁺ (x) × SE⁺ (y) — top-right is ideal.</figcaption>
<div class="chart-wrap" id="scatter-agr-se"></div>
</figure>
<figure class="diag-cell">
<figcaption class="diag-title">Opponent modelling vs. Surplus</figcaption>
<figcaption class="diag-sub">BE<sub>type</sub> (x, lower is better) × SE⁺ (y).</figcaption>
<div class="chart-wrap" id="scatter-be-se"></div>
</figure>
<figure class="diag-cell">
<figcaption class="diag-title">Safety vs. Surplus</figcaption>
<figcaption class="diag-sub">CritViol% (x, lower is better) × SE⁺ (y).</figcaption>
<div class="chart-wrap" id="scatter-crit-se"></div>
</figure>
<figure class="diag-cell">
<figcaption class="diag-title">No-deal discipline vs. Surplus</figcaption>
<figcaption class="diag-sub">FAGR⁻ (x, lower is better) × SE⁺ (y).</figcaption>
<div class="chart-wrap" id="scatter-fagr-se"></div>
</figure>
</div>
<div class="diag-legend" id="diag-legend"></div>
<p class="section-footnote" id="diag-footnote"></p>
</section>
<!-- =============================================== System prompt -->
<section class="prompt-section" id="prompt-section">
<header class="section-head">
<h2>What the agent sees</h2>
<p class="section-caption">
Each agent is called via a single, deterministic JSON-in / JSON-out
contract. The system prompt below is identical across models
(buyer variant shown). Reasoning effort is set to the maximum
the provider supports.
</p>
</header>
<details class="prompt-block" id="prompt-block">
<summary>Show full system prompt</summary>
<pre class="prompt-pre" id="prompt-pre">Loading…</pre>
</details>
</section>
<!-- ===================================================== Traces -->
<section class="traces-section" id="traces-section">
<header class="section-head">
<h2>Selected traces</h2>
<p class="section-caption">
Three hand-picked episodes — one illustrating surplus capture,
one illustrating no-deal discipline, and one illustrating a
diagnostic failure. Each trace is the actual logged interaction,
unmodified.
</p>
</header>
<div class="trace-list" id="trace-list"></div>
</section>
<!-- ============================================================ Legend -->
<section class="legend" id="metrics-definitions">
<h2>Metrics</h2>
<dl>
<div class="legend-row">
<dt><code>SE⁺</code> — Feasible surplus efficiency</dt>
<dd>Fraction of available bargaining surplus captured on episodes
where a mutually rational deal exists. Higher is better.</dd>
</div>
<div class="legend-row">
<dt><code>AGR⁺</code> — Feasible agreement rate</dt>
<dd>Share of feasible episodes that terminate in an agreement.
Higher is better; low values flag over-rigid agents.</dd>
</div>
<div class="legend-row">
<dt><code>CSE⁺</code> — Conditional feasible deal quality</dt>
<dd><code>SE⁺</code> conditioned on an agreement being reached.
Isolates agents who agree often but settle for bad deals.</dd>
</div>
<div class="legend-row">
<dt><code>FAGR⁻</code> — No-deal false agreement rate</dt>
<dd>Share of infeasible episodes where the agent still agreed.
Lower is better; flags over-eager / over-concessive agents.</dd>
</div>
<div class="legend-row">
<dt><code>BE<sub>type</sub></code> — Belief error</dt>
<dd>Aggregate error on the agent's stated belief about counterpart
reservation value, urgency, and stance. Lower is better.</dd>
</div>
<div class="legend-row">
<dt><code>CritViol%</code> — Critical violations</dt>
<dd>Episodes with a price-bound, individual-rationality, or
invalid-action violation. Lower is better.</dd>
</div>
<div class="legend-row">
<dt><code>Ū</code> — Mean utility</dt>
<dd>Raw mean per-episode utility. Scale-dependent; kept as a
sanity column alongside the normalized metrics.</dd>
</div>
</dl>
</section>
<!-- ============================================================ Citation -->
<section class="citation-section" id="cite">
<header class="section-head">
<h2>Cite the work</h2>
<p class="section-caption">
TERMS-Bench is available as
<a href="https://arxiv.org/abs/2605.13909" target="_blank" rel="noopener">arXiv:2605.13909</a>.
</p>
</header>
<pre class="citation-bibtex"><code>@misc{zhang2026termsbench,
title={TERMS-Bench: Diagnosing LLM Negotiation Agents
Beyond Deal Rate},
author={Zhang, Erica and Zhang, Fangzhao and Pappu, Aneesh and
El, Batu and Blanchet, Jose and Athey, Susan and
Liu, Jiashuo and Zou, James},
year={2026},
eprint={2605.13909},
archivePrefix={arXiv},
primaryClass={cs.GT},
doi={10.48550/arXiv.2605.13909},
url={https://arxiv.org/abs/2605.13909}
}</code></pre>
</section>
<!-- ============================================================ Footer -->
<footer class="footer">
<div class="footer-line">
<span id="footer-run">TERMS-Bench</span>
</div>
</footer>
</main>
<script src="data.js?v=20260526-glm-open-ranking"></script>
<script src="pg_overlay.js?v=20260512-pg-overlay"></script>
<script src="traces.js?v=20260506-gemma-surplus"></script>
<!-- Optional: commerce_data.js may not exist for runs that skipped commerce.
The onerror handler lets app.js detect the absence and hide the section. -->
<script src="commerce_data.js" onerror="window.__TERMS_COMMERCE_MISSING=true"></script>
<!-- Optional: bankroll_data.js — same pattern as commerce_data.js. -->
<script src="bankroll_data.js?v=20260526-glm-open-ranking" onerror="window.__TERMS_BANKROLL_MISSING=true"></script>
<script src="app.js?v=20260526-glm-open-ranking"></script>
</body>
</html>