Can AI Agents Synthesize Scientific Conclusions in Health?

arXiv:2606.11337 · cs.AI, cs.CL, cs.CY · Submitted 2026-06-09 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: We're looking at the paper: "Can AI Agents Synthesize Scientific Conclusions in Health?".

Jane: Here's the gist of the paper. SCICONBENCH is a large-scale, live benchmark designed to evaluate open-domain scientific conclusion synthesis, drawing on 9,110 questions paired with expert-written conclusions sourced from the Cochrane Database of Systematic Reviews (CDSR).

Tom: The task assessed involves the long-horizon process of retrieving relevant sources, appraising evidence quality, and integrating heterogeneous evidence to construct an expert-level conclusion.

Jane: To ensure valid measurements and mitigate data leakage—a critical concern in evaluating AI synthesis capabilities—the research introduces SCICONHARNESS, a clean-room evaluation harness that provides agents with controlled web interaction tools.

Tom: This harness allows for controlled web search and browsing, ensuring the evaluation measures genuine synthesis rather than access to ground-truth artifacts.

Title and authors: Tom: So, we're looking at this paper today called "Can AI Agents Synthesize Scientific Conclusions in Health?". It sets up this whole big challenge.

Jane: That title makes you wonder if an AI can actually do the hard work of pulling together complex medical evidence into a solid conclusion.

Lu: The core idea here is testing if these agents can handle that long-horizon task—finding the right sources, judging the quality of that evidence, and then putting it all together into something expert-level.

Meng: I’m curious about how they define "expert-level" because that’s where most of these synthesis tasks fail in production environments.

Lalam: The paper is using this massive benchmark called SCICONBENCH, which has nine thousand one hundred ten questions paired with conclusions written by actual experts from the Cochrane Database of Systematic Reviews.

Tom: Right, so it’s not just asking a simple question; it’s asking an agent to retrieve sources, appraise evidence quality, and integrate different types of information for a health-related conclusion.

Jane: And to make sure the evaluation is fair and isn't just the AI finding a shortcut, they introduced something called SCICONHARNESS.

Lu: SCICONHARNESS is this clean-room evaluation harness that gives agents controlled web interaction tools so we can measure genuine synthesis without letting them just grab the answers from somewhere.

Tom: That’s important because data leakage—where an AI just memorizes the answer instead of reasoning—is a huge problem, and this tool tries to block that access to ground-truth artifacts.

Meng: From a deployment standpoint, I worry that even with clean tools, agents might still rely on patterns they learned from those initial artifact accesses.

Lalam: The paper breaks down the generated conclusions into atomic facts and then uses an expert-validated pipeline to measure factual precision and recall, which gives them a score called an F1 score.

Jane: So, they are quantifying how accurate the conclusion is in terms of what it got right versus what it missed.

Tom: And when they tested eight frontier models and deep research agents under these clean conditions, the results showed that scientific conclusion synthesis is still a significant challenge.

Lu: The best system they tested, o3-deep-research, only achieved a factual F1 score of zero point three three seven.

Meng: zero point three three seven is pretty low if you’re talking about high-stakes health decisions where accuracy matters so much.

Lalam: And what’s even more telling is that using this clean-room evaluation consistently reduced performance by zero point zero two to one seventy two relative to settings where agents could access the ground truth artifacts.

Jane: That suggests that when we see high performance, it might just be because the agent is good at retrieving and rephrasing known answers, rather than actually synthesizing something new.

Tom: I also saw some issues when they audited consumer-facing agents like Google AI Overview and OpenEvidence, where they found them generating incomplete or sometimes contradictory scientific conclusions even when the right answer was available.

Lu: That points to a real problem with how these systems handle the complexity of integrating heterogeneous evidence from different sources.

Meng: If we’re deploying these, we need to know what breaks at scale—and contradiction rates are definitely one of those things that break trust in healthcare applications.

Lalam: The paper's main contribution is introducing SCICONBENCH as this large-scale benchmark, developing SCICONHARNESS for clean-room evaluation, and creating this factual evaluation pipeline.

Jane: So, the overall conclusion they draw from this work is that reliable synthesis of scientific conclusions remains an open challenge, and clean-room evaluation is absolutely essential for seeing the true capability of these open-domain AI agents.

Tom: It confirms what we suspected: it’s not a solved problem yet when we ask an AI to be a genuine scientific synthesizer.

Lu: The authors are pointing toward the need for better representation learning or more sophisticated reasoning pathways to handle that kind of complex integration correctly.

Meng: From an engineering view, the fact that even their best system only hit zero point three three seven suggests we need much stronger constraints on what we ask these models to do in this domain.

Lalam: We’re looking forward to seeing how the next generation of models tackles this because it seems like a critical hurdle for real-world health applications.

The paper's summary: Tom: So we're looking at the core message of this paper now, what they actually claim about these AI agents trying to make scientific conclusions in health research?

Jane: They’re saying that when you ask an AI to synthesize a big conclusion from medical evidence, it’s still really hard. It’s not just about finding a few facts; it’s about the whole process of gathering and judging that information.

Tom: Exactly. They set up this massive test with nine thousand one hundred ten questions based on actual expert conclusions from Cochrane reviews. It shows that even when we give them high-quality source material, getting the final synthesis right is a major hurdle for these systems.

Jane: And to test this fairly, they built this clean-room setup called SCICONHARNESS. This means they controlled the AI's web access so it couldn't just cheat by finding the answer directly in a textbook or an old paper.

Tom: Right, so the main point is that true synthesis requires more than just retrieval; it demands this long process of appraising evidence quality and weaving different pieces together coherently.

Jane: The results show that even the best model tested, o3-deep-research, only got a factual score of zero point three three seven. That’s pretty low for something dealing with health research where accuracy matters so much.

Tom: And they’re being honest about how this clean testing affects performance—it actually made things worse by about zero point zero two to one seventy two compared to letting the agents cheat and access the answer when they were supposed to be synthesizing it.

Jane: That suggests a lot of what we see in the real world, where an AI might look good initially because it found a shortcut, but then fail when you test it under proper conditions.

Tom: Plus, they found that even consumer-facing tools often spit out conclusions that are incomplete or actually contradictory to the ground truth answer, even when the right information is available to them.

Jane: So what this means for us is that we can't just look at a high score and assume the AI has genuinely understood the science. We need better ways to measure if it’s reasoning or just remembering patterns.

Tom: This paper really hammers home that reliable synthesis of scientific conclusions is still an open challenge, and using these clean-room evaluations isn't optional anymore; it’s essential for seeing what these agents can actually do in a real setting.

Jane: If we want to build health applications on top of this AI, we need to move past just looking at accuracy scores and focus on how the model actually builds its argument from scratch.

Tom: And that leads us right into how they did this measurement—they broke the final conclusion down into tiny atomic facts and then used a specific pipeline to check for precision and recall.

Jane: That whole idea of breaking it down into atomic facts is what’s really cool; it gives us a concrete way to see exactly where the synthesis process goes wrong.

Tom: We’ll get into how that fact-checking pipeline works next, but before we do that, we gotta talk about why this benchmark matters so much for real-world deployment.

The paper's improvements: Tom: So we’re looking at what the authors themselves suggest they need to fix about their original benchmark and their testing setup, what they call improvements to SCICONBENCH and SCICONHARNESS?

Jane: They're basically saying that if we want this whole evaluation process to be useful, we need to make it much tighter. They want to stop the leakage of information before it even happens.

Tom: Right. First up is making those clean-room harnesses better by filtering the tool outputs—meaning they’re going to look at things like only letting search results come from specific Cochrane domains or only looking at publications from certain years.

Jane: That way, we ensure the AI isn't just memorizing a single perfect answer; we force it to actually synthesize something relevant and current.

Tom: Then they’re developing this whole new pipeline for fact-checking that breaks the conclusion down into atomic facts—and this pipeline has six sequential steps, starting with preprocessing and ending with redundancy filtering.

Jane: That sounds like a lot of work, but if it works, it means we can reliably measure precision and recall using expert judgments instead of just guessing.

Tom: And they’re also tweaking the way questions are generated to make sure the prompts are "Google-Proof," meaning they have to avoid adding any extra information and focus on preserving that core research question structure, like the PICO structure in medical studies.

Jane: That makes sense because if you feed an AI a messy prompt, you get a messy answer, so cleaning up the input helps clean up the output.

Tom: And they’re looking at consumer-facing agents too, specifically auditing them to look for those contradiction rates we talked about earlier—where they generate conflicting conclusions even when the right answer is sitting right there.

Jane: So it seems like a lot of work is going into making sure that when we do test these models, we aren't just measuring how well they parrot what they already know, but if they have a genuine ability to reason through complex evidence.

Tom: The overall implication is that the authors are pushing for a more rigorous standard for testing AI in sensitive fields like healthcare; it’s not enough to just run it against an unconstrained system.

Jane: It really underlines that reliable synthesis isn't just a matter of making the models bigger; it's about building better ways to evaluate their reasoning process.

Tom: We’ll see how this new fact-checking pipeline performs in the next segment, because that’s where we get to see if they actually solved the problem of measuring quality reliably.

Conclusion: Tom: So we’re wrapping up this deep dive into "Can AI Agents Synthesize Scientific Conclusions in Health?" by summarizing what this whole benchmark study actually tells us about where we are right now?

Jane: Basically, they’re concluding that reliable synthesis of scientific conclusions is still a major open challenge, and to test AI properly in health domains, clean-room evaluation isn't just nice to have anymore; it’s necessary.

Tom: They showed that even the top models only scored zero point three three seven on factual accuracy when they were tested strictly. That number tells us that there’s a big gap between what these systems *think* they know and what they can actually produce reliably.

Jane: And the results strongly suggest that high performance we see often might just be the AI being good at finding and rephrasing existing answers, not actually creating new scientific insight.

Tom: The implication for anyone building something on top of this is that you can't just look at raw accuracy scores; you have to demand evidence of genuine reasoning through those complex synthesis steps.

Jane: It pushes us toward needing better representation learning or more sophisticated reasoning pathways, because the current methods aren't quite getting there when it comes to weaving together heterogeneous medical evidence.

Tom: We’re going from seeing models that can retrieve information to demanding models that can actually construct a sound, verifiable scientific argument.

Jane: If we want AI agents to be useful tools in medicine, they need to demonstrate this deep integration skill under controlled conditions.

Tom: The authors are setting the stage for what needs to happen next—better benchmarks and better evaluation methods if we’re going to make progress on this specific problem.

Jane: Before we wrap up, let’s hear from Lu on what you see as the next big step in tackling these synthesis challenges.

Lu: I think the focus needs to shift toward creating representation learning that inherently understands causal relationships within scientific literature so the AI doesn't just learn correlations but understands why one piece of evidence supports another.

Meng: From a deployment standpoint, if we can't run this kind of rigorous testing efficiently on real traffic, then we can’t trust it in production; the cost and risk of deploying something that fails these benchmarks is just too high for us right now.

Lalam: I'm confident because I have seen the data from the paper; it shows a clear path forward by defining those atomic facts precisely and checking them against expert consensus, which makes the measurement concrete.

Tom: So, to recap, this paper on "Can AI Agents Synthesize Scientific Conclusions in Health?" confirms that synthesis remains tough but provides a necessary framework for testing how well we measure it.

Jane: It’s a call for better evaluation tools to ensure these agents are doing real science, not just sophisticated pattern matching.

Tom: We’ll keep an eye on how the next set of models tackle this because we need to see if they can close that gap between retrieval and true synthesis.

Hayoung Jung, Pedro Viana Diniz, José Reinaldo Corréa Roveda, Abner Fernandes da Silva, Haeun Jung, Enoch Tsai ●

Princeton University · Stony Brook University

cs.AI, cs.CL, cs.CY

Submitted: 2026-06-09

Updated: 2026-11-10

Comments: To appear at NeurIPS 2026 (Evaluations & Datasets Track). Please cite the NeurIPS version when the proceedings are available

Code: https://github.com/hayoungjungg/SciConBench

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 96/100

The gist: SCICONBENCH is a large-scale, live benchmark designed to evaluate open-domain scientific conclusion synthesis…

Key concepts

SCICONBENCH
A massive benchmark containing 9,110 questions paired with expert-written conclusions from the Cochrane Database of Systematic Reviews. It tests how well AI can read multiple sources, judge evidence quality, and create a final scientific summary.
SCICONHARNESS
A controlled evaluation harness designed to prevent data leakage during AI testing. It uses restricted web interaction tools to ensure agents are synthesizing knowledge rather than simply retrieving pre-written answers or ground-truth artifacts.
Factual Evaluation Pipeline
A system that breaks down an AI's conclusion into individual facts and measures them using LLM judges. This pipeline calculates factual precision (how correct the facts are) and recall (how much of the required information is included).
Clean-Room Evaluation
Testing AI models in a controlled environment where they cannot access external, perfect answers. This method reveals the true synthesis ability of an agent by forcing it to generate conclusions from scratch, rather than memorizing existing data.

Terminology

Summary

SCICONBENCH is a large-scale, live benchmark designed to evaluate open-domain scientific conclusion synthesis, drawing on 9,110 questions paired with expert-written conclusions sourced from the Cochrane Database of Systematic Reviews (CDSR). p. 0

Where the paper says it

We introduce SCICONBENCH, a large-scale live benchmark of 9. 11K questions and expert-written conclusions from systematic reviews to evaluate open-domain scientific conclusion synthesis.

The task assessed involves the long-horizon process of retrieving relevant sources, appraising evidence quality, and integrating heterogeneous evidence to construct an expert-level conclusion. p. 2
Where the paper says it

We audit widely-deployed consumer-facing agents, including Google AI Overview and OpenEvidence, and find they synthesize incomplete and sometimes contradictory scientific conclusions, raising concerns for high-stakes decision-making relying on them in real-world health contexts. 2 The SCICONBENCH Dataset SCICONBENCH leverages the Cochrane Database of Systematic Reviews to evaluate models on the long-horizon task of scientific conclusion synthesis: retrieving relevant sources, appraising evidence quality, and integrating heterogeneous evidence to construct a long-form, expert-level conclusion.

To ensure valid measurements and mitigate data leakage—a critical concern in evaluating AI synthesis capabilities—the research introduces SCICONHARNESS, a clean-room evaluation harness that provides agents with controlled web interaction tools. p. 0

Where the paper says it

To mitigate data leakage, we further introduce SCICONHARNESS, a clean-room evaluation harness that equips agents with controlled web interaction to ensure valid measurement. Evaluating 8 frontier models and deep research agents, we find that factual quality remains low: under clean-room settings, the best agent achieves only a factual F1 of 0.

This harness allows for controlled web search and browsing, ensuring the evaluation measures genuine synthesis rather than access to ground-truth artifacts. p. 3
Where the paper says it

We address this challenge with SCICONHARNESS, an MCP-based harness that provides controlled access to web search and browsing tools. SCICONHARNESS enforces a “clean-room setting” preventing models’ access to ground-truth systematic reviews, while retrieving, integrating, and reasoning over scientific evidence from the open web [48].

The core of the factual evaluation pipeline decomposes generated conclusions into atomic facts. p. 1

Where the paper says it

(3) SCICONHARNESS enforces clean-room evaluation by blocking ground-truth artifacts. (4) Generated conclusions are evaluated against ground-truth references using an expert-validated pipeline that decomposes both into atomic facts and computes factual precision, recall, and F1.

These facts are then measured by LLM-based judges to quantify factual precision (correctness) and recall (coverage), resulting in an overall F1 score. p. 1
Where the paper says it

To further mitigate leakage, we introduce SCICONHARNESS, a clean-room evaluation harness with controlled web search and browsing tools. Finally, we develop a factual evaluation pipeline that decomposes generated conclusions into atomic facts2 and uses LLM-based judges to measure factual precision (correctness), factual recall (coverage), and F1 (overall quality), showing strong agreement with expert judgments.

When evaluating eight frontier models and deep research agents on SCICONBENCH under clean-room settings, the results indicate that scientific conclusion synthesis remains a significant challenge. p. 1
Where the paper says it

Finally, we develop a factual evaluation pipeline that decomposes generated conclusions into atomic facts2 and uses LLM-based judges to measure factual precision (correctness), factual recall (coverage), and F1 (overall quality), showing strong agreement with expert judgments. Evaluating 8 frontier models and deep research agents on SCICONBENCH, we find scientific conclusion synthesis remains an open challenge under clean-room evaluation: the best system, o3-deep-research, achieves only F1= 0.

The best system tested, o3-deep-research, achieved only a factual F1 of 0. p. 1
Where the paper says it

Finally, we develop a factual evaluation pipeline that decomposes generated conclusions into atomic facts2 and uses LLM-based judges to measure factual precision (correctness), factual recall (coverage), and F1 (overall quality), showing strong agreement with expert judgments. Evaluating 8 frontier models and deep research agents on SCICONBENCH, we find scientific conclusion synthesis remains an open challenge under clean-room evaluation: the best system, o3-deep-research, achieves only F1= 0.

337. Furthermore, clean-room evaluation consistently reduces performance by 0. p. 0
Where the paper says it

Evaluating 8 frontier models and deep research agents, we find that factual quality remains low: under clean-room settings, the best agent achieves only a factual F1 of 0. Our cleanroom setting consistently reduces performance relative to unconstrained evaluation, suggesting that leakage inflates estimates of models’ true synthesis capabilities.

02–0.172 relative to unconstrained settings (where agents can access ground-truth artifacts), suggesting that apparent high performance often stems from artifact retrieval rather than true synthesis ability. p. 1
Where the paper says it

Across systems, clean-room evaluation reduces factual F1 by 0. 172 relative to unconstrained settings (agents can access ground-truth artifacts), indicating that much of the apparent performance arises from retrieving ground-truth artifacts rather than from genuine synthesis.

Audits of consumer-facing agents also reveal frequent issues, such as generating incomplete or contradictory conclusions even when the ground-truth answer is available. p. 0
Where the paper says it

Finally, we audit consumer-facing agents (e. , Google AI Overview, OpenEvidence) and find they frequently generate incomplete and sometimes contradictory conclusions, even when the ground-truth answer is available.

The paper's main contributions are: 1) the introduction of SCICONBENCH, a large-scale benchmark; 2) the development of SCICONHARNESS for clean-room evaluation; and 3) a factual evaluation pipeline measuring precision and recall. The overall conclusion is that reliable synthesis of scientific conclusions is an open challenge, and clean-room evaluation is essential for assessing the true capabilities of open-domain AI agents. p. 0

Where the paper says it

, Google AI Overview, OpenEvidence) and find they frequently generate incomplete and sometimes contradictory conclusions, even when the ground-truth answer is available. Overall, our results show that reliable synthesis of scientific conclusions remains an open challenge, and that clean-room evaluation is essential for assessing open-domain AI agents.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Implement clean-room evaluation harnesses like SCICONHARNESS to mitigate benchmark leakage by filtering tool outputs based on CDSR domains or publication dates, ensuring that measurements reflect genuine synthesis rather than memorization or shortcut retrieval.

  2. Develop a factual evaluation pipeline that decomposes generated conclusions into atomic facts and measures factual precision and recall using an expert-validated pipeline, showing strong agreement with expert judgments to assess quality reliably.

  3. Use a modular atomic fact generation pipeline involving six sequential steps—preprocessing, decomposition, decontextualization, incomplete fact rewriting, relevance filtering, and redundancy filtering—to ensure generated facts are self-contained and relevant to the question.

  4. System prompt engineering for question generation should enforce guidelines such as avoiding extraneous information and aiming to generate a question that is Google-Proof while preserving the PICO structure of the original research objective.

  5. Evaluate consumer-facing agents by auditing them on SCICONBENCH, specifically looking at contradiction rates, as results show they frequently generate incomplete and sometimes contradictory conclusions, even when the ground-truth answer is available.

Abstract

Scientific AI agents increasingly retrieve evidence, reason across sources, and synthesize conclusions used in consequential decisions. Yet, their ability to do so in high-stakes domains such as health remains unclear. We introduce SciConBench, a large-scale live benchmark of 9.21K questions and expert-written conclusions from systematic reviews to evaluate open-domain scientific conclusion synthesis in health and medical domains. The benchmark draws on an expert-validated automated evaluation pipeline that decomposes conclusions into atomic facts and measures correctness and comprehensiveness via factual precision and recall. To mitigate data leakage, we further introduce SciConHarness, a clean-room evaluation harness that equips agents with controlled web interaction to ensure valid measurement. Evaluating 8 frontier models and deep research agents, we find that factual quality remains low: under clean-room settings, the best agent achieves only a factual F1 of 0.337. Our clean-room setting consistently reduces performance relative to unconstrained evaluation, suggesting that leakage inflates estimates of models' true synthesis capabilities. Finally, we audit consumer-facing agents (e.g., Google AI Overview, OpenEvidence) and find they frequently generate incomplete and sometimes contradictory conclusions, even when the ground-truth answer is available. Overall, our results show that reliable synthesis of scientific conclusions in health remains an open challenge, and that clean-room evaluation is essential for assessing open-domain AI agents.

Sources

Related papers