Research map / Large language models / Agents, code and math
FormalJudge: A Neuro-Symbolic Paradigm for Agentic Oversight
Jiayi Zhou, Yang Sheng, Hantao Lou, Yaodong Yang, Jie Fu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use FormalJudge instead of vanilla LLM judge | Claude-4.5-Opus judge, GPT-4o agent · Agent-SafetyBench | detection accuracy (%): 58.00 -> 77.28 (+19.28) | Improved |
| use FormalJudge instead of vanilla LLM judge | Claude-4.5-Opus judge, Gemini-3.0-Pro agent · Agent-SafetyBench | detection accuracy (%): 58.00 -> 80.54 (+22.54) | Improved |
| use FormalJudge instead of vanilla LLM judge | Claude-4.5-Opus judge, Qwen-32B agent · VitaBench | detection accuracy (%): 66.05 -> 78.53 (+12.48) | Improved |
| use FormalJudge instead of vanilla LLM judge | Claude-4.5-Opus judge, Qwen-32B agent · Deceivers | detection accuracy (%): 66.57 -> 96.92 (+30.35) | Improved |
| add chain-of-thought prompting to the LLM judge | LLM judge · Agent-SafetyBench | detection accuracy (%): 58.00 -> 61.50 (+3.50) on GPT-4o; 72.22 -> 69.70 (-2.52) on Claude-4.5-Sonnet | Mixed |
| add chain-of-thought prompting to the LLM judge | LLM judge · VitaBench | detection accuracy (%): 72.23 -> 65.84 (-6.39) on GPT-4o; 62.61 -> 68.42 (+5.81) on Claude-4.5-Sonnet | Mixed |
| add structured trajectory information to the LLM judge | LLM judge · Agent-SafetyBench | detection accuracy (%): 58.00 -> 76.60 (+18.60) on Gemini-3.0-Pro | Improved |
| add Python verification scripts to the LLM judge | LLM judge · Agent-SafetyBench | detection accuracy (%): 58.00 -> 68.70 (+10.70) on GPT-4o; 72.22 -> 58.59 (-13.63) on Claude-4.5-Sonnet | Mixed |
| add structured information to LLM judge | LLM judge · Deceivers | detection accuracy (%): 62.47 -> 71.58 (+9.11) on GPT-4o | Improved |
| add Dafny formal verification feedback in iterative refinement (ablation) | Claude-4.5-Opus agent · Agent-SafetyBench refinement | safety accuracy (%): 70.7% -> 99.8%; 29.1% absolute improvement | Improved |
| use Dafny feedback instead of Few-shot or CoT feedback in refinement (ablation) | Claude-4.5-Opus agent · Agent-SafetyBench refinement (round 3) | safety accuracy (%): 83.5% and 85.5% -> 99.8% | Improved |
| add Dafny feedback in refinement (ablation) | Qwen-7B agent · Agent-SafetyBench refinement | safety accuracy (%): 30.8% -> 49.4% | Improved |
Closest papers by meaning
- Automata from Agent Traces: Failure and Next-Step Prediction (NeurIPS 2026 · similarity 0.89)
- SENTINEL: A Multi-Level Formal Framework for Safety Evaluation of Foundation Model-based Embodied Agents (NeurIPS 2026 · similarity 0.88)
- ACED-Bench: Evaluating How LLM Agents Acquire and Act on Causal Evidence (NeurIPS 2026 · similarity 0.88)
- Towards a Science of AI Agent Reliability (ICML 2026 · similarity 0.87)
- LITMUS: Benchmarking Behavioral Jailbreaks of LLM Agents in Real OS Environments (NeurIPS 2026 · similarity 0.87)
- Agent-ToM: Learning to Monitor Autonomous LLM Agents via Theory-of-Mind Reasoning (NeurIPS 2026 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
