Research map / Multi-agent systems and games / LLM agents
VERITAS: A Multi-agent Co-scientist for Verifiable Image-Derived Hypothesis Testing
Lucas Stoffl, Benedikt Wiestler, Johannes Paetzold
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 |
|---|---|---|---|
| add multi-agent discussion over single-model pipeline (SMe to Veritas) | SMe pipeline, local · 64-hypothesis benchmark, majority vote | verdict accuracy: 71.2% vs. 61.0% | Improved |
| add multi-agent discussion over single-model pipeline (SMe to Veritas) | SMe pipeline, frontier · 64-hypothesis benchmark, majority vote | verdict accuracy: 81.4% vs. 61.0% | Improved |
| add code execution to direct reasoning (SMa to SMb) | SMa direct reasoning · 64-hypothesis benchmark | verdict accuracy: 56.1 -> 63.8 | Improved |
| add iterative self-correction (SMc to SMd) | SMc code via API · 64-hypothesis benchmark | evidence-label accuracy: +7.3 local, +11.8 frontier | Improved |
| add iterative self-correction (SMc to SMd) | SMc code via API · 64-hypothesis benchmark | verdict accuracy: +2.8 local, +3.8 frontier | Improved |
| add structured phase guidance (SMd to SMe) | SMd agentic, frontier · 64-hypothesis benchmark | verdict accuracy: 76.6 -> 65.5 | Hurt |
| halve context window to 8k from 16k (ablation) | Veritas local · UCSF-PDGM | evidence-label accuracy: -7.7pp (36.7 -> 29.0) | Hurt |
| extend context window to 32k from 16k (ablation) | Veritas local · UCSF-PDGM | evidence-label accuracy: -5.0pp (36.7 -> 31.7) | Hurt |
| set temperature to zero instead of 0.2 (ablation) | Veritas local · UCSF-PDGM | evidence-label accuracy: -9.0pp (36.7 -> 27.7) | Hurt |
| add critic during execution phases (ablation) | Veritas local · 64-hypothesis benchmark combined | verdict accuracy: +1.8pp combined verdict (71.4 vs 73.2 without critic) | No clear effect |
| extend critic to planning and interpretation phases (ablation) | Veritas local · 64-hypothesis benchmark combined | evidence-label accuracy: -10.2pp evidence, -4.9pp verdict | Hurt |
| use GT expert masks instead of SAT-Nano masks (ablation) | Veritas local · UCSF-PDGM | evidence-label accuracy: -23.7pp (36.7 -> 13.0) | Hurt |
Closest papers by meaning
- MedAgent-Pro: Towards Evidence-based Multi-modal Medical Diagnosis via Reasoning Agentic Workflow (ICLR 2026 · similarity 0.87)
- DDx-TRACE: A Benchmark for Medical Diagnostic Trajectories in VLMs (NeurIPS 2026 · similarity 0.87)
- Can AI Agents Synthesize Scientific Conclusions? (NeurIPS 2026 · similarity 0.87)
- MicroVQA: A Multimodal Reasoning Benchmark for Microscopy-Based Scientific Research (CVPR 2025 · similarity 0.86)
- Medical thinking with multiple images (ICLR 2026 · similarity 0.86)
- ACED-Bench: Evaluating How LLM Agents Acquire and Act on Causal Evidence (NeurIPS 2026 · similarity 0.86)
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.
