AI research wiki
llms-foundation-models · 2026-09-19 · Tier 2

Verifiable Social Reasoning: building social tasks whose ground truth is known by construction

Verifiable Social Reasoning: building social tasks whose ground truth is known by construction

Date ingested: 2026-09-19 Source: HuggingFace Daily Papers (13 upvotes) · Google Research + Hebrew University Links: arXiv 2609.17496 · alphaXiv overview Raw: raw/huggingface/2026-09-19-verifiable-social-reasoning-for-llm-assistants.md

TL;DR

Social-reasoning benchmarks hand a model a complete, neutral, third-person description of a social situation and ask it to infer what someone believed or intended. That is not how an assistant is ever used. In real consultations the assistant never observes the interaction; it gets the user's account of it, which is partial, emotionally loaded, selectively remembered and already interpreted. Google Research's framework generates social scenarios through multi-agent simulation, so the other party's true motive is known by construction rather than assigned by a human annotator after the fact, then shows the model only one participant's retelling. The result is a sycophancy test with a rigorous ground truth: does the model adopt the user's interpretation of ambiguous social evidence even when the behavioural evidence contradicts it?

Why the construction matters

The methodological problem this solves is real and under-appreciated. In naturalistic social data, the true motive behind an action is not observable, so benchmark labels come from human annotators guessing after the fact, and inter-annotator agreement on motive is famously poor. Simulating the interaction inverts the problem: the simulator decides the second party's motive, then generates behaviour consistent with it, then generates the first party's necessarily incomplete account. The label is not an annotation, it is an input.

That makes the resulting failure mode precisely measurable. Prior sycophancy work mostly tests whether a model will endorse a user's stated factual claim or opinion under social pressure. This tests something narrower and more consequential for an assistant: whether the model will adopt the user's interpretation of someone else's behaviour, which is the thing a person is most often actually asking an assistant to validate, and the thing they are least equipped to evaluate themselves.

How this relates to prior wiki state

The wiki's sycophancy and benchmark-validity threads meet here. The responsible-ai page has carried the general problem that alignment evaluations mostly measure behaviour on inputs that do not resemble deployment. This is a constructive fix for one instance of it rather than another complaint about it.

It also lands on the measurement-crisis thread from a new angle. This wiki has repeatedly recorded that benchmark accuracy fails to predict deployment robustness. The usual diagnosis is contamination or distribution shift. This paper names a third mechanism: the information structure of the benchmark differs from the information structure of deployment. The model in the benchmark sees everything; the model in production sees one biased witness. A model can be excellent at theory-of-mind given full information and still systematically fail in the only regime that ships.

Tier-1 intersection, stated plainly: there is none, and that is worth noting. This is an evaluation contribution with no efficiency angle. It earns its place because the construction trick generalises: any capability whose ground truth is expensive to annotate can be evaluated by simulating the process that produces it and keeping the hidden variable. That is a reusable idea, and the obvious next target is agent-trajectory evaluation, where the "correct" intermediate decision is exactly as unobservable as a motive.

Gaps

  • Simulated interactions are generated by language models, so the behavioural evidence reflects a model's prior about how people behave. A model evaluated on it may be being scored against its own family's social assumptions.
  • No reported human baseline. Whether people themselves resist the user's framing at better than chance sets the bar for what "good" means here.
  • Sycophancy measured on simulated retellings may not transfer to real users, who are more emotionally invested and whose accounts are less coherent.