Non-Collaborative User Simulators for Tool Agents

TL;DR

Existing user simulators always cooperate. Ours asks for the impossible, drifts off topic, loses patience, and types half-sentences, yet still delivers every goal item. Under these users, state-of-the-art tool agents lose up to half of their success rate, and fine-tuning on collaborative data alone does not fix it.

1. Which behaviors? Unavailable service, tangential talk, impatience, and incomplete utterances.
2. How do we keep the goal intact? A dialogue state tracker and an ending verifier on a τ-bench-style backbone; 97% of dialogues convey the full goal on the first try.
3. How much do agents suffer? Every model drops on MultiWOZ and τ-bench; tangential users cost 29.1% relative success on average.
4. Does fine-tuning help? Only when non-collaborative dialogues are in the training mix (average success 78.8 → 86.9).
5. Is the simulator realistic? Harder on agents than a prompt-only simulator at equal goal alignment, and preferred by annotators about 70% of the time.

Four non-collaborative behaviors

Grounded in marketing research on difficult customers and in real user–LLM chat logs. Each card: the behavior, an example, and how the simulator produces it (Figure 3 of the paper, one panel per behavior; click a panel to enlarge).

A simulator that misbehaves but still finishes the job

Every behavior is layered on a collaborative simulator built to deliver the whole user goal, so a failed task is the agent's fault, not a user who forgot to mention something.

• Backbone. The τ-bench prompt simulator (GPT-4.1-mini): user goal, instruction, dialogue history; emits <END> when done.
• Dialogue state tracker. Shards the goal into information pieces and delivers any that are still unsent before the dialogue ends.
• Ending verifier. Blocks <END> while the agent still has to act or is waiting for confirmation.
• Goal alignment check. GPT-4o-mini verifies that every goal item was conveyed; failures are regenerated. Only 3.0% of 1,230 dialogues fail on the first attempt, and the judge agrees with a human at MCC 0.77.

Results

Every agent loses under every behavior, each in its own way: unavailable services and incomplete utterances break tool use, tangents and impatience break dialogue management.

Table 1 of the paper. Relative SR is SR divided by the model's collaborative SR. Values below 60 are marked.

Unavailable service tool use

Tangential dialogue management

Impatience dialogue management

Incomplete utterances tool use

Two behaviors at once

Beyond tool use: ColBench and MINT

The simulator ports with small adjustments. ColBench (backend programming, no tools) repeats the tool-agent pattern; MINT (user–agent collaboration on HotpotQA) does not, so the findings hold for domains built around fulfilling a user goal.

Table 2 of the paper. Success rates averaged over 4 trials.

Fine-tuning on collaborative data leaves agents fragile

SFT on successful collaborative MultiWOZ dialogues takes 3B and 7B models from under 4% to over 90% success with collaborative users. The gains do not transfer: unavailable services (61 to 77%) and incomplete utterances (72 to 82%) lag far behind, with more duplicated API calls and parameter hallucinations than the zero-shot baselines.

Adding non-collaborative dialogues in equal proportion lifts Qwen2.5-3b on every behavior and raises the average from 78.8 to 86.9. Up-weighting incomplete-utterance data closes that gap further at a small cost elsewhere.

Table 3 of the paper. Success rate on MultiWOZ. Checkpoints and training sets are released (see Code).

Is the simulator itself any good?

Baseline: a prompt-based user simulator (PBUS) that adds the behavior descriptions to the τ-bench prompt, with no extra modules. Tested with two behaviors at once. PBUS barely moves the agent; ours does, at equal or better initial goal alignment (IGA), so the difficulty comes from the behaviors, not from missing information.

Table 4 of the paper, agent GPT-4.1-mini. COL collaborative, IMP impatience, UNA unavailable service, TAN tangential, INC incomplete utterance.

Human evaluation

Nine annotators compared dialogues pairwise. Ours wins about 70% overall and every single-behavior mode:

Figure 5 of the paper, single-behavior mode.

Takeaways

• Test agents against users who do not cooperate. A collaborative simulator hides the failures that appear the moment a user asks for the impossible, changes the subject, gets angry, or types “I want to res”.
• Each behavior needs its own fix. Better grounding in API docs for unavailable services and terse users; dialogue strategies that acknowledge without stalling for tangents and anger.
• Put non-collaborative dialogues in the training set. A uniform mix of the four behaviors recovers most of the gap that collaborative-only SFT leaves.

Run the simulator on your own agent

Both environments ship with their own conda setup. Behaviors are switched on by dropping YAML files into non_coll_list/; models, trials, and API keys are set in the run script.

# MultiWOZ environment
git clone https://github.com/holi-lab/NCUser
cd NCUser/multiwoz_env
conda create -n multiwoz_env python=3.11 && conda activate multiwoz_env
pip install -r requirements.txt

# choose the user behaviors to simulate (any subset; each YAML is one mode):
#   normal (collaborative) · unavailable_service · tangential
#   emotional_acts (impatience) · fragment_dumping (incomplete utterance)
cp non_coll_normal.yaml non_coll_tangential.yaml non_coll_list/

# set N (trials), the models array, and API keys in run_dialogue_simulation.sh, then
./run_dialogue_simulation.sh
./run_error_analysis.sh

# τ-bench environment: same workflow
cd ../tau_env && conda create -n tau_env python=3.11 && conda activate tau_env
pip install -e .
./run_dialogue_simulation.sh

Fine-tuned agents and training sets: Hugging Face. All dialogue logs behind the tables: Google Drive. Configuration, vLLM serving, and training: README.

@inproceedings{shim2026noncollaborative,
    title     = {Non-Collaborative User Simulators for Tool Agents},
    author    = {Shim, Jeonghoon and Song, Woojung and Jin, Cheyon and Kook, Seungwon and Jo, Yohan},
    booktitle = {The Fourteenth International Conference on Learning Representations},
    year      = {2026},
    url       = {https://openreview.net/forum?id=UAUimofy3W}
}