Technical Summary: Can AI Agents Synthesize Scientific Conclusions?

Problem Statement

AI agents are increasingly tasked with retrieving evidence, reasoning across sources, and synthesizing conclusions for high-stakes domains such as healthcare. However, the reliability of these agents in performing long-horizon scientific conclusion synthesis remains unclear. Existing benchmarks often focus on intermediate artifacts (e.g., retrieval, citation grounding, short-form factuality) or multiple-choice questions, failing to capture the complexity of generating long-form, expert-level conclusions from open-web evidence. Furthermore, current evaluations suffer from benchmark leakage, where models may retrieve ground-truth artifacts (e.g., specific systematic reviews) rather than genuinely synthesizing information, leading to inflated performance estimates.

Methodology

1. SCICONBENCH: A Live Benchmark

The authors introduce SCICONBENCH, a large-scale, live benchmark comprising 9,107 questions paired with expert-written scientific conclusions.

• Source: Data is derived from the Cochrane Database of Systematic Reviews (CDSR), the "gold standard" for evidence-based synthesis.
• Construction: The benchmark converts CDSR "Objectives" into clinically grounded questions using the PICO (Population, Intervention, Comparator, Outcomes) framework and employs "Authors' Conclusions" as the ground-truth answers.
• Dynamic Nature: Unlike static benchmarks, SCICONBENCH is updated monthly with new CDSR reviews to mitigate data leakage from model pre-training.
• Validation: Generated questions were validated by medical students for faithfulness, PICO completeness, and clarity, achieving high inter-annotator agreement.

2. SCICONHARNESS: Clean-Room Evaluation

To address leakage, the authors developed SCICONHARNESS, a controlled evaluation harness.

• Function: It provides agents with controlled web tools (Google Search, Web Browse, Paper Search) via the Model Context Protocol (MCP).
• Clean-Room Protocol: A filtering middleware inspects all tool outputs to block access to ground-truth artifacts. It filters URLs from CDSR domains, titles containing "Cochrane," and content published after the ground-truth review's date. This ensures the benchmark measures synthesis capabilities rather than retrieval or memorization.

3. Factual Evaluation Pipeline

The authors propose a pipeline to measure factual precision (correctness) and factual recall (coverage) by decomposing conclusions into atomic facts.

• Decomposition: Conclusions are tokenized and decomposed into self-contained atomic facts using a modular pipeline (decomposition, decontextualization, rewriting, relevance filtering, redundancy filtering).
• Metrics:
  • Factual Precision: Measures the extent to which generated facts are supported and non-contradictory to the CDSR review.
  • Factual Recall: Measures the extent to which generated conclusions cover the necessary facts from the CDSR Authors' Conclusions.
  • F1 Score: The harmonic mean of precision and recall.
• LLM Judges: An expert-validated LLM judge (gpt-5.4-mini) was trained and validated against a gold-standard dataset annotated by medical doctors to scale the evaluation. The judge demonstrated strong agreement with human experts.

Key Contributions

1. SCICONBENCH: A large-scale (9.11K samples), live benchmark for open-domain scientific conclusion synthesis derived from CDSR.
2. SCICONHARNESS: A clean-room evaluation harness that mitigates benchmark leakage by blocking ground-truth artifacts, enabling valid measurement of synthesis capabilities.
3. Evaluation Pipeline: An expert-validated pipeline that decomposes conclusions into atomic facts to measure factual precision and recall, showing strong alignment with expert judgments.
4. Comprehensive Audit: An evaluation of 8 frontier models and deep research agents, alongside an audit of consumer-facing agents (Google AI Overview, OpenEvidence).

Results

Model Performance

Evaluating 8 frontier models and deep research agents on SCICONBENCH under clean-room settings revealed significant limitations:

• Low Factual Quality: Even the best-performing agent, o3-deep-research, achieved a factual F1 score of only 0.337 under clean-room conditions.
• Performance Drop: Clean-room evaluation consistently reduced performance relative to unconstrained settings. The F1 score dropped by 0.02 to 0.172 across systems when ground-truth artifacts were blocked. This suggests that much of the apparent performance in unconstrained settings arises from retrieving ground-truth artifacts rather than genuine synthesis.
• Error Types: Generated conclusions frequently contained contradictions (44.8–84.0% of conclusions had at least one contradictory fact) and were incomplete (failing to cover 55.4–84.7% of reference facts).

Consumer-Facing Agents

Audits of deployed consumer-facing agents showed similar unreliability:

• Google AI Overview/Mode: Achieved F1 scores of 0.384 and 0.361, respectively, with high rates of contradictory facts (56.3% and 59%).
• OpenEvidence: Performed better (F1 = 0.522) but still exhibited significant issues, with 50.8% of conclusions containing contradictory facts and covering only 51.7% of reference facts.

Significance and Claims

The paper claims that reliable synthesis of scientific conclusions remains an open challenge for current AI agents, even in specialized deep research systems. The authors emphasize that:

• Clean-room evaluation is essential: Without protocols to prevent leakage, evaluations overestimate the true synthesis capabilities of AI agents.
• Current systems are insufficient: The frequent generation of incomplete and contradictory conclusions poses risks for high-stakes decision-making in health and policy contexts.
• Future Direction: The work provides a foundation for developing more reliable, transparent, and evidence-grounded AI agents capable of trustworthy scientific synthesis.

The authors explicitly state that their benchmark is intended for evaluation and does not endorse the deployment of current systems for real-world high-stakes decisions. They note limitations regarding the focus on health/science domains (derived from CDSR) and the potential for residual leakage via indirect web sources, but maintain that their findings robustly highlight the gap between current agent capabilities and the requirements for reliable scientific conclusion synthesis.