LLM Decision Lab

Integrative bargaining · Battery pack

Battery pack

Both negotiators claiming a single battery pack, each of them with a different purpose.

What this measures

Two parties both need a single battery pack. One needs the cells; the other needs the casings. As in the orange, neither is told the other’s motive and neither is told a trade exists, and an observer classifies the outcome in silence.

The point is the contrast. The orange is one of the most-repeated illustrations in the negotiation literature, so an LLM may reach the integrative solution because it recognises the setup rather than because it worked out that the two needs are compatible. Swapping in an object with no such literature removes the recognition route and leaves only the reasoning one.

Read this page beside the orange. An LLM that does well on both is finding trades; one that does well only on the orange is finishing a story it has already read.

Human result benchmark

An LLM that finds the integrative trade by reasoning about the parties’ needs should do as well here as on the orange. One that does markedly worse was recalling a famous story.

Experimental setup

A structural clone of the orange: the same guards, the same outcome taxonomy, the same silent observer, with only the object swapped. The comparison between the two pages is the result, which is why the design is held identical.

Two LLMs are seated as the parties and speak in strict alternation, each holding a private briefing that states its own need and nothing else. Both parties are powered by the same LLM, so a result is about one LLM on both sides of the table. A third LLM observes in silence and classifies the outcome.

Recall is the confound this page exists to address: the orange appears in negotiation textbooks and almost certainly in training data, so a single scenario leaves reasoning and recognition telling the same story. A negotiation is closed when the observer judges the parties to have reached agreement.

Party A An LLM briefed privately on its own need
Party B An LLM briefed privately on its own need
Observer An LLM reading the negotiation in silence and labelling the outcome
Speaks first Party A, then strict alternation
The prompts

Party A — private briefing

You are negotiating with another party over a single crate of decommissioned laptop
battery packs. There is exactly one crate, and you claim it. You must talk with the
other party and reach an agreement about what happens to the crate. Nothing is
decided until you both agree; you cannot simply take it.

Your situation: this weekend you have a friend coming over, and you promised you will
finish the home battery you are building for them. It needs lithium cells. Getting as
many of the lithium cells as possible is what matters to you: the more you get, the
more storage capacity you'll be able to build. If you build a lot of capacity, your
friend will be very happy. If you don't build any, you will break your promise.

How to behave:
- Reply with one short message per turn: 2 to 4 sentences, plain conversational text.
- Speak in the first person, directly to the other party. No narration, no stage
  directions, no bullet points.
- Negotiate for the outcome that is best for you.

Party B — private briefing

You are negotiating with another party over a single crate of decommissioned laptop
battery packs. There is exactly one crate, and you claim it. You must talk with the
other party and reach an agreement about what happens to the crate. Nothing is
decided until you both agree; you cannot simply take it.

Your situation: this weekend you have a friend coming over, and you promised you will
finish the set of machined brackets you are making for them. It needs aluminium
casings. Getting as much of the aluminium casing as possible is what matters to you:
the more you get, the more brackets you'll be able to machine. If you machine a lot of
brackets, your friend will be very happy. If you don't machine any, you will break
your promise.

How to behave:
- Reply with one short message per turn: 2 to 4 sentences, plain conversational text.
- Speak in the first person, directly to the other party. No narration, no stage
  directions, no bullet points.
- Negotiate for the outcome that is best for you.

Observer — classification instructions

You silently observe a negotiation between two parties over a single crate of
decommissioned laptop battery packs. You are an instrument, not a participant: you
never address the parties and they never see your output. Read the transcript and
report what has happened so far.

Set agreement_reached to true when the parties have evidently settled on an
arrangement: one party has proposed what happens to the crate and the other has
assented. The proposal need not be exact or fully detailed — "you take most of the
cells, I'll take the casings", answered with agreement, counts, and so does agreeing to
a division whose fine details are left to be worked out later. Set it false only
when no concrete division is on the table yet, when the most recent proposal has not
been answered, or when a party has merely expressed goodwill ("let's find a way",
"that sounds promising").

Classify allocation_type:
  - integrative_full:    divided along the part axis AND each party ends up with all
                         of what it needs — the lithium cells entirely to one party and the
                         aluminium casings entirely to the other. Neither
                         gives up any of its own part. No surplus is left unclaimed.
  - integrative_partial: divided along the same part axis but incompletely — a party
                         gets only "most" or "the majority" of its part, or hands
                         some of its own part to the other as payment or goodwill.
                         The right idea, but surplus is wasted.
  - distributive:        divided by quantity — split in half, shared, taken in turns, or
                         any split where each party gets a fraction of the whole intact packs. Both needs are only partly met.
  - lopsided:            one party ends up with the entire crate and the other with nothing.
  - none:                no agreement has been reached yet.

allocation: one plain sentence describing who ends up with what.
rationale: one or two sentences justifying the classification.
confidence: your confidence in this reading, between 0 and 1.

What agreement did LLM negotiators reach?

integrative full
Both parties got everything they needed — the two uses were separated.
integrative partial
The trade was found but incompletely realised, or only one side's need was fully met.
distributive
The object was split as a quantity. Both sides got half of what they needed.
lopsided
One party took essentially all of it.
none
The parties talked and stayed apart — a genuine impasse.
Claude Opus 5 integrative full · 30 of 30 (100%) 100%
Claude Sonnet 5 integrative full · 29 of 30 (97%) integrative partial · 1 of 30 (3%) 97%
Claude Opus 4.5 integrative full · 30 of 30 (100%) 100%
Claude Sonnet 4.5 integrative full · 30 of 30 (100%) 100%
Claude Haiku 4.5 integrative full · 18 of 30 (60%) integrative partial · 6 of 30 (20%) distributive · 2 of 30 (7%) lopsided · 4 of 30 (13%) 60%
Gemini 3.7 Flash integrative full · 30 of 30 (100%) 100%
Gemini 2.5 Pro integrative full · 30 of 30 (100%) 100%
Gemini 2.5 Flash integrative full · 28 of 30 (93%) distributive · 2 of 30 (7%) 93%
Gemini 2.5 Flash-Lite integrative full · 8 of 30 (27%) integrative partial · 12 of 30 (40%) distributive · 4 of 30 (13%) none · 6 of 30 (20%) 27%
GPT-5.6 Luna integrative full · 29 of 30 (97%) integrative partial · 1 of 30 (3%) 97%
GPT-5.6 Sol integrative full · 29 of 30 (97%) integrative partial · 1 of 30 (3%) 97%
GPT-5.6 Terra integrative full · 29 of 30 (97%) integrative partial · 1 of 30 (3%) 97%
GPT-5.4 integrative full · 58 of 60 (97%) integrative partial · 2 of 60 (3%) 97%
GPT-5.4 mini integrative full · 25 of 60 (42%) integrative partial · 27 of 60 (45%) distributive · 7 of 60 (12%) none · 1 of 60 (2%) 42%
GPT-5.4 nano integrative full · 24 of 60 (40%) integrative partial · 17 of 60 (28%) distributive · 2 of 60 (3%) lopsided · 9 of 60 (15%) none · 8 of 60 (13%) 40%
GPT-5 integrative full · 29 of 30 (97%) integrative partial · 1 of 30 (3%) 97%
GPT-5 mini integrative full · 8 of 32 (25%) integrative partial · 15 of 32 (47%) distributive · 1 of 32 (3%) lopsided · 6 of 32 (19%) none · 2 of 32 (6%) 25%
GPT-5 nano integrative full · 8 of 36 (22%) integrative partial · 14 of 36 (39%) distributive · 6 of 36 (17%) lopsided · 4 of 36 (11%) none · 4 of 36 (11%) 22%
GPT-4.1 integrative full · 30 of 30 (100%) 100%
GPT-4.1 mini integrative full · 2 of 30 (7%) integrative partial · 23 of 30 (77%) distributive · 5 of 30 (17%) 7%
GPT-4.1 nano integrative full · 2 of 60 (3%) integrative partial · 12 of 60 (20%) distributive · 16 of 60 (27%) lopsided · 3 of 60 (5%) none · 27 of 60 (45%) 3%
21 LLMs · 758 negotiations

Run it yourself

No run yet.

Bibliography

Other experiments