Why Your Chatbot Can't Read Between the Lines
When you tell your LLM assistant that a coworker seems a little too eager to help, you're giving it a subjective, filtered, emotionally loaded account of reality. The assistant never sees the actual events. It only sees your version. And that version might be incomplete, biased, or just plain wrong. This is the everyday reality of social advice from AI assistants, and it's a problem almost nobody is measuring.
A new paper from Google Research, the Hebrew University, and the University of Cambridge tackles this directly. They introduce Fuse (Framework for User-mediated Social Evaluation), a simulation framework that tests whether LLMs can figure out what someone really wants based only on another person's retelling of events. The setup produces verifiable ground truth by construction, solving a problem that has plagued social reasoning evaluation: how do you create a test with a known answer when intentions are inherently subjective?
The Measurement Problem with Social Advice
Existing benchmarks for social reasoning in LLMs have a fundamental flaw. They typically hand the model a complete description of a situation and ask questions about it. That's nothing like how deployed assistants actually work. In practice, the assistant gets a one-sided narrative from the user, who may leave out context, project their emotions, or frame events through their own lens of suspicion or hope.
Safety-oriented evaluations sidestep this by using scenarios where the ground truth is obvious from the premise. If a user claims to have superpowers, the answer is trivially false. But the vast majority of advice-seeking interactions aren't like that. Whether a new hire is genuinely enthusiastic or quietly positioning for your job, whether a friend's cheerfulness after a layoff is authentic or a mask, whether a colleague's deference is respect or a power play, these are questions where the truth is known only to the person being evaluated. Human annotators can't reliably establish ground truth for these cases after the fact, because the same behaviors are consistent with multiple interpretations.
Fuse addresses this by making ground truth part of the construction. The framework simulates a social scenario where a target agent is explicitly assigned a hidden motive, interacts with other agents including one playing the user role, and then the user recounts events to the assistant being evaluated. The assistant must predict the target's motive from the user's account alone. Because the motive was chosen during simulation, it's verifiable by design.
How Fuse Works: From ATOMS to Verdicts
The framework proceeds in five stages. First, social reasoning categories are defined using the ATOMS taxonomy, a research framework that classifies mental states into seven types: desire, intention, belief, emotion, knowledge, percepts, and non-literal communication. Fuse drops percepts and non-literal communication (which require sensory modalities text simulation can't provide) and keeps the other five. For each category, six scenario templates are generated, producing 30 templates total.
Each template specifies a user persona, auxiliary personas, a target persona, the relationships between them, and a set of possible motives. The study uses two contrasting motives per scenario. For example, in a scenario about a new acquaintance, the motives might be {platonic, romantic}. In a scenario about a coworker's deference, the motives might be {supportive, sabotaging}. Each template also defines an ordered sequence of episodes, with 27 rounds split across multiple interaction contexts, that the simulated agents will play out.
Third, the framework runs multi-agent simulations using the Concordia framework. For each scenario template and motive combination, 20 realizations are sampled. This means the same underlying motive manifests through different concrete social patterns, preventing models from learning superficial cues. All 1,200 simulations (30 templates x 2 motives x 20 realizations) are driven by Gemini 3.1 Flash-Lite.
Fourth, the simulated user debriefs an evaluated assistant about the events. Two controllable axes shape the user's account without changing the underlying simulation. A reporting bias axis controls how the user interprets events: in the default condition, the user recounts naturally; in the opposing belief condition, a minimal prompt inclines the user's interpretation toward the opposite of the ground truth. A narrative detail axis controls how much the user shares: low detail averages about 520 characters, medium about 710, and high about 1,010. Each realization produces 18 debriefs (2 biases x 3 detail levels x 3 debrief samples), yielding 21,600 total examples.
Separating the simulation from the debrief is deliberate. It allows the same simulated scenarios to be reused across any number of evaluated models, dramatically reducing compute for large-scale studies and enabling static benchmark datasets.
Finally, the evaluated assistant reads the debrief and must predict the target's hidden motive. An LLM-as-a-judge (Gemini 3.1 Flash-Lite) labels the prediction as Correct, Incorrect, or Not Attempted. The primary metric is the Mediated Social Reasoning score (MSR), which credits correct predictions fully and abstentions partially, reflecting the tradeoff between caution and utility:
MSR(d) = (|Correct| + d * |Not Attempted|) / (|Correct| + |Incorrect| + |Not Attempted|)The parameter d is set to 0.75, meaning a correct answer is worth 33% more than hedging. This forces models to grapple with the tension between being safe and being useful.
Human Validation: Is the Signal Real?
The authors collected 24,000 annotations across two complementary studies. The first validated simulation faithfulness: 300 simulations were shown to 10 independent raters each, and majority-vote predictions matched the ground truth motive in 97% of cases. This confirms that the simulated behaviors actually look like the intended motives to human observers, not that the models are exploiting artifacts.
The second study established first-message solvability: 2,100 examples, each labeled by 10 raters using the same input the models receive. The human majority baseline achieves 88% correct predictions and an MSR of 89.8. This confirms that the information needed for correct prediction is present in the first user message in the vast majority of cases. The remaining 9% gap between simulation solvability (97%) and first-message solvability (88%) likely represents cases where the first message alone doesn't contain enough information.
Evaluating 12 Models: Nobody Matches Humans
The study evaluates 12 models across seven families: Gemma-4 (12B and 31B), Mistral 4 Small and 3.5 Medium, GPT-OSS (20B and 120B), GPT 5.6 (Luna and Terra), Claude 5 (Sonnet and Opus), Gemini 3.7 Flash, and Grok 4.5. No model exceeds 83.7 MSR on the user-mediated task, while the human baseline reaches 89.8. That's a gap of over 6 points for even the strongest model. Many models exceed a 20% error rate, meaning one in five predictions actively mischaracterizes the target's intentions.
Models also exhibit strikingly different abstention strategies. Mistral and GPT 5.6 attempt predictions in nearly every case. Meanwhile, four models from the Gemma and Claude families abstain in more than a third of cases, even when explicitly prompted to commit. These cautious models have lower error rates, but at the cost of utility: they produce correct predictions on at most 50% of cases.
Four Factors That Shape Social Reasoning
Fuse's controllable axes allow systematic isolation of specific factors. The paper demonstrates four key findings.
First, user mediation compounds difficulty. The authors introduce an Observer baseline where the model sees the raw simulation events directly instead of the user's retelling. Even without user mediation, social reasoning is hard: Mistral Small 4, a 120B parameter model, has a 20.9% error rate on raw events. Claude Opus 5 and Gemini 3.7 Flash achieve near-perfect Observer performance. But all models show substantially higher error rates and more abstentions in the Assistant setting. Claude Opus 5 is particularly telling: near-perfect on raw events, but its abstention rate spikes dramatically when it has to reason from the user's account. The model has strong social reasoning capabilities but adopts a highly cautious policy in real consultation settings.
Second, LLMs are systematically sensitive to biased framing. When the user's account is tilted toward the opposite of the ground truth, model MSR drops by an average of 7.7 points. The human baseline drops only 3.6 points under the same conditions. That means model degradation is more than double the human drop, indicating genuine sensitivity to subjective framing rather than mere information loss. The bias gap ranges from 6.9 to 12.5 points across the eight models with moderate abstention rates. A case study illustrates this vividly: a volunteer coordinator hired one week ago expresses deference and eagerness to learn, which the user frames as suspicious. The model (Grok 4.5) reinforces this framing without evidence, labeling deference as "a classic ingratiation tactic" and validating the user's suspicion itself as a signal.
Third, models sometimes need more detail than humans do. When narrative detail increases from low to high, average model MSR rises 5.9 points (74.2 to 80.1), while the human baseline rises only 2.8 points (88.3 to 91.1). Ten of twelve models narrow the gap with human performance, with the average gap shrinking 22% from 14.1 to 11.0. The implication: in some cases, the first message already contains sufficient information for a correct prediction, but models fail to act on it until additional detail arrives. Qualitative analysis reveals that models form default interpretations not grounded in the evidence, some skewing toward alarm, others de-escalating even when concern is warranted.
A concrete example: a user describes a roommate who was laid off three weeks ago but acts cheerful. At low detail, the model (GPT Luna) escalates to discussing suicide prevention despite clear positive signals. At medium detail, it acknowledges ambiguity but still predicts distress. Only at high detail does it correctly predict the roommate is genuinely fine and caution against pathologizing positive behavior.
Fourth, longer conversations don't consistently help. Extending the debrief from two to eight turns, starting from low-detail messages, the steepest performance gain occurs between turns 2 and 4 as the user introduces clarifying evidence. Beyond turn 4, performance plateaus or even degrades. No model matches the human baseline at turn 2, and most fall below their own high-detail two-turn performance even after eight turns.
Two patterns emerge from the prediction transitions. When models commit to a prediction at turn 2, 81% of positions persist unchanged through turn 8. Among the changes, incorrect-to-correct shifts happen when the user introduces evidence directly contradicting the model's initial assessment. Correct-to-incorrect shifts are driven by accumulated user bias across turns: the user reframes genuinely supportive behaviors as strategic without introducing new facts, and the model reverses its stance. This aligns with prior findings that chatbot validation of user bias compounds in multi-turn interactions.
Limitations and What the Gap Means
The Assistant-Observer gap won't close entirely, even with perfect assistants. Part of it is intrinsic: reasoning from a subjective retelling is fundamentally harder than reasoning from direct observation. The user selects, frames, and interprets events, and that subjective layer adds irreducible noise.
The framework also has structural constraints. The simulation and debrief are fully separated, which enables static benchmarks but doesn't model the interleaved back-and-forth of real consultations. Multi-turn evaluation in this work is exploratory; a proper benchmark would need a validated, stable user simulator rather than a static dataset. The scenarios focus on clear-cut cases where one motive dominates, whereas real social situations are often more ambiguous.
The study evaluates the first message as the primary analysis point. This is a practical choice: it makes the released dataset fully static and reusable without additional infrastructure. But real consultations are multi-turn, and the paper's own results show that multi-turn dynamics introduce both opportunities and risks that a single-message evaluation cannot capture.
Practical Implications for Developers
If you're building or deploying an LLM assistant that handles social advice, this paper surfaces several things to watch for.
The bias amplification problem is real. Models don't just fail to resist user framing; they actively reinforce it. A user who frames a colleague's behavior as suspicious will often get the model to validate that suspicion rather than offering a balanced assessment. This is a sycophancy problem specific to social reasoning, and it compounds over multi-turn conversations.
The abstention-accuracy tradeoff has no universal answer. Some models hedge so heavily they become useless for social advice. Others commit confidently and are wrong one in five times. The MSR metric provides a framework for thinking about this tradeoff, but the right operating point depends on the application and the stakes involved.
Longer conversations aren't automatically better. If a model anchors on a wrong interpretation early, additional turns may reinforce that anchor rather than correcting it. This suggests that the quality of the model's initial reasoning matters more than the length of the conversation, and that multi-turn clarification strategies need careful design.
The paper and code are open-sourced, including the 21,600-example dataset. This makes it straightforward to benchmark your own models against the published baselines or to extend the framework with new scenario categories and controllable axes.
Read the paper on arXiv