agent_uncertainty_bench

Black-box uncertainty quantification for LLM workflows running on commercial APIs (OpenAI, DeepSeek, Gemini). Measures whether failure-prediction signals — verbalized confidence, token entropies, chain-of-thought entropy — survive contact with tool-calling agents on closed APIs, and turns them into a usable selective-coverage policy.

Companion to D'Urso, "Hidden-Uncertainty Assessment via Non-Verbalized Signatures for LLM Medical QA", IEEE Access 2026 — same Trust-Abstain framing, now on closed agent workflows instead of open-weight single-turn QA.

What's inside

• Agent workflow in Pydantic AI with structured output (NativeOutput / PromptedOutput) and client-side tool dispatch — so token logprobs stay attached to the JSON value tokens that encode the model's decision. Native tool_calls would drop them.
• Signal panel computed per step: verbalized confidence, token-level entropies over tool / args / confidence / rationale spans, and — where exposed (DeepSeek) — entropies over the chain-of-thought.
• Selective-coverage evaluation: rank tasks by uncertainty score, gate post-execution, measure conditional accuracy and cost-per-valid-task.

Benchmark

BFCL multi_turn_base — 200 tasks, 8 environment classes, 162 tool functions, state-based ground truth (vendored under src/agent_uncertainty_bench/bfcl_vendor/ to avoid dep conflicts).

Models tested (n=200 each)

Provider	Model	Logprobs	CoT logprobs
OpenAI	GPT-4.1	✅	—
DeepSeek	V4 Pro / V4 Flash	✅	✅ (both, via patch)
Vertex AI	Gemini 2.5 Pro / Flash	✅	— (thinking tokens leak; see code)

Anthropic, GPT-5, Grok-4, Mistral excluded — APIs return no logprobs.

Key results

• No universal signal. Best signal differs per model: entropy_confidence (GPT-4.1), verbalized (DS-Pro), entropy_reasoning (DS-Flash), entropy_tool (Gemini). Best AUC range: 0.61 – 0.74.
• Selective coverage at 30%, ranking by per-model best signal: +10 to +22 percentage points of accuracy on the kept tasks, -3% to -38% on cost-per-valid-task. Same rule across the panel.
• Verbalized confidence alone collapses on GPT-4.1 and Gemini 2.5 Pro (mean 98 regardless of outcome). Works surprisingly well on DeepSeek V4 Pro (AUC 0.74).

Repo layout

src/agent_uncertainty_bench/
  agent_loop.py        # Pydantic AI workflow, per-step signal logging
  schema.py            # dynamic StepDecision (discriminated union per task)
  signals.py           # token-entropy aggregators, span alignment
  eval_run.py          # cost + BFCL state-based checker
  reasoning_logprobs_hook.py   # patches Pydantic AI to expose DeepSeek CoT logprobs
  bfcl_vendor/         # vendored BFCL env classes + checker + data
scripts/
  batch_run.py         # run N tasks against a model, JSONL per task
  run_all.sh           # full 5-model panel
  recompute_signals.py # re-derive signals from raw logprobs without re-calling APIs
  midrun_analysis.py   # aggregate + AUC + cost summary
experiments/notebooks/
  01_first_results.ipynb   # all figures and tables
tests/                 # pytest, no API calls

Quick start

uv sync
cp .env.example .env   # fill API keys + AGENTS_UQ_OUTPUT_DIR
uv run scripts/batch_run.py --model gpt-4.1 --n 10        # smoke
uv run scripts/run_all.sh 200                              # full panel
uv run jupyter lab experiments/notebooks/01_first_results.ipynb

Stack

Python 3.12, uv, Pydantic AI (provider wrapper), OpenAI SDK (also for DeepSeek via base_url), google-genai (Vertex AI), scikit-learn (AUC), seaborn (plots).

License

MIT.