- Rating
- 1
- Confidence
- 2
- Informativeness
- 1
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
This paper argues that scientific breakthrough can come from ideas or anomalies, but the majority of AI4Science benchmarks focus on the former. To make progress towards the latter, the authors introduce SENSE-Bench (Scientific Exploration through Noticing, Selecting, and Experimenting), which constructs synthetic environments with anomalies that models must reason over, in order to plan new experiments. The authors compare the anomaly detection (can models detect anomalies), research taste (can models detect fruitful anomalies to pursue), and task completion (did models come up with the right explanation) abilities of GPT in Codex and DeekSeek in the DeepSeek harness. The authors show the task is challenging, study how sensitive the results are, and present a case study with a failure mode of GPT Astra
Critical strengths
- The motivation of this paper is strong: breakthroughs in research can stem from studying anomalies, not just a hypothesis. The authors make progress towards this problem by constructing a new dataset to measure various abilities needed to complete such tasks (taste, discoverability, exploration, ...)
Critical weaknesses
-
The details of the benchmark construction are described at a very high level, which makes it hard to determine the soundness of the protocol. There is only one example of a task at the end of the paper (a case study in Section 4.4), but it only describes the initial task description without outlining the environment, tools, anomalies, etc., which are essential to understanding the paper. Further, the entire benchmark construction is AI-generated, and I do not see any rigorous human evaluation protocol. The paper vaguely states there is "human review at key stages" (L269), "Humans review all case families and anomaly annotations disputed during AI review" (L296), and "Errors and disagreements receive arbitration and human review" (L325). The authors need to provide more example of their tasks, clearly outline the protocol, and add scientific quality control checks to quantify the usefulness of the dataset
-
The experiments are extremely sparse, only taking up around one page. They include a single evaluation run comparing DeepSeek and GPT, a reliability check on a subset (which is not really a separate research question, and would be considered standard to report uncertainty estimates), and a qualitative example of a failure with Astra. The authors fail to interpret their results and takeaways, especially as they find that weaker models can do better than Astra on their benchmark, separate runs do not largely change the results, and one failure mode. To reach a conference submission, the authors would have to expand the sections greatly (e.g., discuss why the trends matter, turn the case study into more of a taxonomy of failure modes)
Questions for authors
- Why do weaker GPT/Deepseek models outperform their stronger counterparts in Taste and Mechanism Selection? This is unintuitive
- All of the Cohen's kappas values are extremely small, often near 0 which suggests near random agreement with the benchmark. Why is this? I feel the authors need to have a more rigorous human annotation/quality control protocol to disentangle the obvious case that the benchmark is impossible due to errors
- Do the authors have a definition of "taste" anywhere in the paper? I am unconvinced that the current evaluation setup is measuring taste, both because it is never defined and the evaluation protocol is described at too much of a high level
- I find this paper to follow a very classic style of AutoResearch papers: they generate a dataset, do a single evaluation run, assess reliability, and show a single case study. Can the authors clarify the extend of their AI use? The disclosure statement is vague and states it was used to generate the entire benchmark and translate the draft from chinese
Thank you for your careful review and helpful comments. Our responses are below.
F1 partly reflects a model's willingness to investigate, not just its ability to choose anomalies worth investigating. We therefore also report Cohen's κ, which measures agreement with the reference labels beyond chance. DeepSeek Flash chose to investigate in 85.3% of its valid Taste responses, compared with 59.7% for Pro. The models had different numbers of valid responses. Flash had higher F1 (0.7755 versus 0.6915), but Pro had higher κ (0.1234 versus 0.0422). So the ranking depends on which metric we use. For GPT-6 Astra, the pattern suggests that it is less willing to investigate when the initial evidence is limited. Its higher scores in the later stages suggest that it makes better use of new observations and tests explanations more effectively. Astra had the highest Experiment (0.2125) and Discovery (0.2000) scores.
Low κ values mean little agreement with the reference labels beyond chance. For Taste, an agent that recommends investigating every anomaly gets κ = 0 when both reference labels are present, even if its F1 is high. The agents' tendency to recommend investigation helps explain this gap between F1 and κ. For each family, two graduate students (master’s or PhD) in the relevant fields first made judgments based on the run records and task background. They then used labels and explanations from five agents as references during discussion. A third reviewer joined the discussion and made the final decision when disagreements remained. Graduate students also reviewed every mechanism across the nine families. They checked its scientific basis and the experiments needed to separate the possible explanations.
The Introduction defines research taste as “the judgment of which anomalies deserve further investigation.” We focus on deciding whether the evidence for an anomaly makes it worth further experiments. For each confirmed anomaly, the agent reads the handoff records and chooses whether to investigate. We compare its choice with the final label from human review and report F1 and Cohen's κ.
We designed the benchmark framework and wrote the paper's outline and initial draft. Human reviewers checked all scenarios, mechanisms, and labels across the nine families. We built a workflow in which AI agents screened source environments, proposed mechanisms, wrote code, adapted environments, and produced annotations. Human reviewers checked mechanisms before implementation. Two human annotators first judged the labels from the run records and task background. They used the agents' labels and explanations as references in later discussions. A third reviewer made the final decision when disagreements remained. We repeatedly revised the framework, mechanisms, tools, task descriptions, and scoring instructions after reviewing experimental records. AI assisted with translation and language editing.