$τ$-Emotion-Bench: Exposing the Interaction Fragility of LLM Agents through Real-World User Emotions

AI summary

Overview

Research area: Large language model (LLM) agents that call tools over long, multi-turn interactions with users — specifically the evaluation and training of such agents when the simulated user is emotionally variable rather than cooperative.

Technical level: Intermediate. The paper assumes familiarity with tool-calling agents, user simulators, trajectory fine-tuning, and metrics such as Pass^k, but the core idea is stated plainly.

Scope: The paper introduces τ-Emotion-Bench, a 580-task ten-domain benchmark that holds the underlying task fixed while varying the user's emotional condition, and Tracer, a data-generation pipeline that uses the same emotion-conditioned harness to collect and filter agent-training trajectories.

What This Paper Is About

Existing tool-calling benchmarks and agent-training pipelines largely assume users who are cooperative and behaviorally stable, and they judge agent quality almost entirely by whether the task eventually succeeds. That means a trajectory can score as a success even if the agent repeatedly asks unnecessary questions, requests information it could have gotten from tools, or leans on the user instead of the environment. This paper asks how much agent reliability changes when the same task is paired with users expressing controlled emotions, and whether emotion-conditioned interaction can double as a training signal that rewards efficient, tool-driven problem solving rather than task completion alone.

Key Contributions

1. τ-Emotion-Bench. A ten-domain benchmark with 580 executable tasks and an emotion-conditioned agent–user–environment harness that varies user behavior — five emotion families with 22 fine-grained emotions plus behavioral dimensions — while preserving the underlying task objective.
2. An interaction-aware data-generation pipeline. Tracer builds grounded, policy-valid tool workflows, reverse-refines them into natural user instructions, and collects trajectories in which a dissatisfied user can terminate the interaction, so successful-but-inefficient trajectories are filtered out.
3. Generalizable agent training. Tracer supervision improves Qwen3 models at 1.7B, 4B, and 8B scales, transfers to six unseen domains with a small training set (1,498 trajectories), and transfers to τ-Bench and τ-Trait, which use different tasks and user simulators.
4. Behavioral analysis of failure modes. The paper examines policy violations, the balance between tool calls and follow-up questions, empathy, and simulator hallucination rates to explain why emotion-conditioned interaction exposes fragility.

Main Findings

• Frontier agents remain unreliable under emotion-conditioned users. With Qwen3-235B-A22B-Instruct-2507 as the user simulator, the best model, DeepSeek-V4.1-Flash, reaches 55.9% Avg@3, followed by GLM-5.3 at 51.9% and Kimi-K2.7-Code at 49.3%. Average Pass^3 tops out at only 32.2%.
• Consistency across repeated trials drops sharply. Under GLM-5.2 as the simulator, GLM-5.3 reaches 78.6% Avg@3 but only 66.5% average Pass^3, leaving a substantial gap between average success and consistent completion.
• Performance is uneven across domains. Under the Qwen3 simulator, DeepSeek-V4.1-Flash scores 44.3% Pass^3 on Retail but 17.1% on Telecom, while GLM-5.3 reaches 62.0% on Basketball but only 2.9% on Telecom. Under GLM-5.2, GLM-5.3 reaches 94.3% Pass^3 on Retail but only 40.0% on Telecom.
• Tracer tuning improves small models substantially. Under the Qwen3 simulator, Tracer-High improves Avg@3 by 13.8–19.7 over the corresponding bases, and Pass^3 rises from 1.2% to 10.4% at 1.7B, 3.9% to 17.2% at 4B, and 8.4% to 19.8% at 8B. Tracer-High-8B reaches 38.1% Avg@3, above MiniMax-M3 at 37.0% and Qwen3-235B-A22B-Thinking-2507 at 36.0%.
• Gains hold under a second simulator. Under GLM-5.2, Tracer-High raises Avg@3 from 11.1% to 36.3% at 1.7B, 31.2% to 51.4% at 4B, and 37.6% to 54.7% at 8B — an improvement of 17.1–25.2 percentage points — with Pass^3 rising from 4.0% to 19.4%, 16.2% to 34.2%, and 22.6% to 37.5% respectively. Tracer-High-8B reaches 54.7% Avg@3, exceeding Qwen3-235B-A22B-Instruct-2507 at 51.7% and matching Muse-Glimmer-30B.
• Out-of-domain transfer with less data. Tracer-Low, trained on 1,498 trajectories from four domains, achieves the highest average Pass^3 among compared methods at 1.7B, 4B, and 8B on six unseen domains. At 8B it reaches 20.7% average Pass^1 and 8.3% Pass^3, outperforming EnvScaler, Arctic-AWM, Simia-Tau-SFT, and Simia-Tau-SFT-RL on average Pass^3 — with prior methods using 9,000 SFT trajectories (EnvScaler) and 91,204 synthetic trajectories (Simia-Tau-SFT).
• Cross-benchmark transfer. Tracer-tuned models improve over their bases on τ-Bench (Airline and Retail) and τ-Trait (Telehealth and Telecom), including on the stricter Pass^2 and Pass^3 metrics, despite different tasks and user simulators.
• Policy violations persist. Violation rates are higher for base Qwen3 and weaker models and lower for Tracer-High and stronger models, but some conditions still reach approximately 67% in Telehealth and 91% in Telecom even among stronger models.
• Success tracks a balance between tools and follow-up. Higher Avg@3 is associated with a moderate ratio of tool execution to conversational follow-up across all five emotion families, and Tracer-tuned models reach higher performance without shifting toward tool-heavy interaction.
• Empathy correlates with more follow-up. Empathetic responses are accompanied by a higher fraction of follow-up questions in every model group; empathy is associated with lower success under Enjoyment for all groups and under Fear for most groups, with the effect under Disgust varying across models.
• The user simulator is not destabilized by emotion conditioning. Of 6,961 failed trajectories examined, 449 user turns were flagged as potential hallucinations and only 124 were confirmed as user errors — 1.1% of failed trials and at most 1.4% for any emotion family.
• Training data composition differs sharply. Tool calls exceed follow-up actions in 76.2% of Tracer-Low and 78.9% of Tracer-High trajectories, versus 55.6% for AReaL and 22.0% for Simia; 51.9% of Simia trajectories contain more follow-up than tool calls, versus 12.8% for Tracer-Low and 10.6% for Tracer-High.

Methodology in Plain English

The authors keep the environment fixed and vary only the user. Each task is a partially observable Markov decision process where the agent can call tools and send messages, and the user simulator responds according to the task instruction plus an emotion condition. The simulator can end the episode with either a normal stop or an "unsatisfied" signal; if dissatisfaction is expressed before the task completes, the trajectory receives a reward of -1. This is the key mechanism: an agent that leans too heavily on the user can have its own trajectory cut off and denied success.

To build the tasks, the authors reuse customer-service environments, database schemas, and tool APIs from prior work (τ-Bench, τ-Trait, ToolOrchestra) without modifying the environment logic. For each of the ten domains they construct a directed tool dependency graph, where edges indicate which tools can validly follow which others under domain policy, verify the graph manually, then sample paths through it to form workflows. A grounding agent fills in concrete entities and arguments using the executable environment, producing a reference trajectory; a preference agent turns that into a natural user request with feasible preferences; and a separate feeling generator describes the user's likely emotional state from user context only, without access to the environment database.

Each free-form feeling is mapped into the emotion space — five families (Anger, Fear, Disgust, Sadness, Enjoyment) containing 22 fine-grained emotions — using natural-language anchors and a multi-head classifier over fine-grained emotion and behavioral attributes, then combined with politeness, urgency, and trust dimensions to yield 594 possible conditions. The simulator receives the natural-language instruction rather than discrete labels, so the same condition can be expressed differently as the conversation evolves.

Before inclusion, tasks are verified for a concrete objective, sufficient information, an executable reference trajectory, and policy consistency, using deterministic replay with model-based adjudication only when replay is inconclusive. All 580 evaluation tasks were reviewed manually, which identified 101 tasks requiring correction. As a solvability check, a human tool-calling agent completed selected tasks that the evaluated models failed across all trials.

Why This Matters

Impact on research. The paper argues that terminal task success is an incomplete measure of agent quality. By making the interaction itself affect the reward — through a user who can become dissatisfied and terminate — it supplies a training signal that distinguishes "eventually solved the task" from "solved the task efficiently." It also provides a benchmark where the task is held constant and only the user's emotional condition changes, which isolates a variable that prior interactive benchmarks did not systematically control.

Real-world applications (drawn from the ten domains covered by the benchmark):

• Airline and travel booking agents that must complete reservations while handling a frustrated or anxious traveler.
• Retail and restaurant customer-service agents handling orders, returns, and complaints from impatient users.
• Telecom and telehealth support agents where policy constraints (what may be done without confirmation) matter as much as task completion.
• Banking agents where unconfirmed actions are costly and user trust levels vary.

Industry relevance. The finding that a small model trained on 1,498 interaction-aware trajectories can outperform substantially more data-intensive training methods — and can match or exceed much larger models under one simulator — is directly relevant to teams deciding how to spend annotation and compute budgets. The 91% and 67% policy-violation figures in Telecom and Telehealth also indicate that deploying tool-calling agents in regulated workflows carries compliance risk that aggregate success rates would hide.

Future Directions

• Reinforcement learning. The authors state explicitly that Tracer-generated data is evaluated only through supervised fine-tuning; its effectiveness as RL supervision remains unexplored.
• Beyond tool calling. Extending τ-Emotion-Bench and Tracer to other agent settings, such as coding and web-based agents, is named as future work.
• Understanding the empathy trade-off. Empathy is consistently associated with more follow-up questions and, in several conditions, with lower success. How to acknowledge a user's emotional state without adding conversational overhead is left open.
• Reducing residual policy violations. Even stronger and Tracer-tuned models show high violation rates in Telecom and Telehealth, so what drives those domain-specific failures is unresolved.

Target Audience

Researchers and engineers working on tool-calling LLM agents, user simulation, and agent training data pipelines. It is most directly useful to those designing interactive benchmarks that need to control user behavior, and to practitioners deciding between small fine-tuned agents and large frontier models for user-facing deployments where interaction efficiency and policy adherence matter. Readers without background in agent harnesses or Pass^k evaluation will need to consult the cited prior work on τ-Bench and its descendants first.

Resources listed in the paper: the model at https://huggingface.co/tau-emotion-bench/models, the dataset at https://huggingface.co/tau-emotion-bench/datasets, code at https://github.com/sahsaeedi/tau-emotion-bench, and a project website at https://tau-emotion-bench.github.io. The paper is arXiv:2610.19560v1, licensed CC BY 4.0, with equal-contribution authorship and correspondence to ssaeidi1@asu.edu.

Authors’ abstract

In long-horizon tool-calling tasks, LLM agents can fail not only because of tool-use errors, but also because of ineffective interaction with users. Yet existing benchmarks largely assume cooperative and behaviorally stable users, leaving the effect of emotional variation on agent reliability underexplored. We introduce $τ$-Emotion-Bench, a ten-domain benchmark for evaluating tool-calling agents under controlled emotion-conditioned user interactions while preserving the underlying task objective. To scale both benchmark construction and agent training, we introduce Tracer, a scalable agentic data-generation pipeline that constructs grounded, policy-valid tasks and collects trajectories through an interactive agent--user--environment harness. By allowing dissatisfied users to terminate inefficient interactions, the harness provides an interaction-aware signal that filters trajectories with excessive user dependence and favors effective tool use. Experiments with two user simulators show that frontier agents remain unreliable, with substantial gaps between average task success and consistent task completion. Tracer-generated supervision consistently improves Qwen3 models across scales, generalizes to unseen domains, and transfers to $τ$-Bench and $τ$-Trait. Behavioral analyses further show that higher task success is associated with a better balance between tool execution and necessary user follow-up. Overall, $τ$-Emotion-Bench and Tracer provide a scalable framework for evaluating agent reliability under diverse user behaviors and improving tool-calling agents through interaction-aware supervision.