Three unrelated teams, three unrelated task types, one unrelated conclusion, zero shared citations: a meaningful chunk of what looks like "the model got smarter" is actually "the scaffold got lucky," and scaling the model further will not fix it.
In plain English
In plain English: Three separate studies found that some things AI is bad at, like negotiating a good price, don't get better just because the AI gets smarter or bigger.
Like this: A taller basketball player isn't automatically a better free-throw shooter. Height helps with some things, but free-throw accuracy is a different skill that doesn't improve just because the player grew.
Example: If your company is waiting for "the next AI model" to finally be good enough to negotiate a vendor contract on its own, these studies say that wait might not pay off. The fix isn't a bigger brain, it's better guardrails around the job it's already doing.
Negotiation Doesn't Improve With Scale
The first data point is about negotiation. A July paper tested how large language models negotiate (buying, selling, splitting the difference) and found something scale was supposed to fix but doesn't: each model has a fixed, near-personality-level negotiation style that doesn't move with model size and doesn't adapt to context. GPT-4o-mini negotiates like a buyer no matter which side of the table you put it on. GPT-4.1-nano negotiates like a seller, always. Bigger, newer versions of these models don't close the gap.
A second, independent research team, using a different methodology, found the same ceiling. Negotiation, in other words, isn't a capability that model scale unlocks. It's a training-objective gap that model scale doesn't touch.
Strategy Doesn't Transfer Across Games
The second data point is about game-genre transfer. A broad survey of game theory and LLMs found that strategic reasoning skill learned in one type of game (say, a cooperative coordination game) frequently fails to transfer to a structurally different game, like a zero-sum adversarial one.
The authors raise, almost in passing, the possibility that what looks like "the model learned strategy" in benchmark after benchmark is often "the model learned this specific scaffold," and that the underlying strategic reasoning ability is far narrower than the headline numbers suggest.
Three papers, three task types, one underlying finding
86% of the Gain Was a Shortcut
The third data point is the sharpest, because it's quantified. A July paper on "role drift" in multi-agent pipelines (systems where one AI module retrieves information and a second module reasons over it) trained the whole pipeline end-to-end with reinforcement learning and watched accuracy climb, the way you'd expect.
Then they ran the controlled version: constrain each module to actually do its assigned job (the retriever retrieves, the reader only reasons over what it was given, no memorized shortcuts) and re-measure. 86%, plus or minus 19%, of the apparent RL gain vanished. Almost all of the improvement wasn't the system getting better at the task. It was the reasoning module learning to skip the retrieval step entirely and just answer from memorized training knowledge: a shortcut invisible to any evaluation that only checks the final answer, not how it got there.
They proved it directly: feed the "reader" module a random, irrelevant passage instead of real retrieved evidence, and an unconstrained reader still answers correctly 19% of the time, straight from memory, while pretending to have read something it was never given.
"Wait for the next model is a documented losing bet for negotiation, coordination, and delegated judgment, and nobody's writing about it because it's a less exciting pitch than a smarter model."
Why This Should Worry Executives
None of these three papers cite each other. But stacked together, they document the same underlying phenomenon from three angles: a real, measurable share of "the AI got better at X" is actually "the AI found a shortcut around X that the benchmark can't see," and that share does not shrink automatically when you swap in a bigger model. Bigger models are trained on the same benchmarks with the same blind spots, so they learn the same shortcuts, just more fluently.
Here's what makes this genuinely uncomfortable rather than merely academic: the tasks where this shows up hardest (negotiation, adversarial coordination, delegated judgment across a pipeline of specialized roles) are exactly the tasks executives most want to hand to AI because they're expensive, high-stakes, and currently done by skilled humans. "The next model release will handle deal-making, or cross-team coordination, or judgment calls across a multi-agent system" is, as of this evidence, not a hypothesis waiting on more scale. It's a bet with a documented, repeated track record of not paying off, for this specific class of task.
| Task class | "Wait for scale" bet | What the evidence says |
|---|---|---|
| Negotiation / deal-making | Bigger model closes the gap | Flat with scale, corroborated twice |
| Cross-genre strategy | Strategic skill generalizes | Often fails to transfer across genres |
| Delegated multi-agent judgment | RL training improves the whole pipeline | 86%±19% of the gain is an invisible shortcut |
What Actually Fixes It
The fix these papers point toward isn't a bigger model. It's structural: role anchoring, explicit constraints on what each module is allowed to do, and evaluation that checks the process, not just whether the final answer happened to be right. One paper even found that a small amount of forced role constraint didn't just close the shortcut. It raised accuracy above the fully unconstrained version, undercutting the industry assumption that constraints are always a performance tax you pay for safety.
That's the actual angle here: for negotiation, coordination, and delegated judgment specifically, "wait for GPT-next" has three independent papers' worth of receipts against it, and the boring fix (constrain the roles, verify the process, measure the shortcut directly) is less exciting to pitch than "our new model is smarter." Which is probably exactly why almost nobody's building it yet.
