acceptodds
Under review as a conference paper at ICLR 2027

Do AI Agents Have Research Taste? SENSE-Bench for Anomaly-Driven Scientific Exploration

Abstract

Existing AI research systems and benchmarks largely start from predefined ideas or research goals, while scientific inquiry can also begin with unexpected observations. Pursuing such observations requires research taste to judge which anomalies warrant investigation, a capability that remains insufficiently evaluated in current agents. We introduce SENSE-Bench (**S**cientific **E**xploration through **N**oticing, **S**electing, and **E**xperimenting), a benchmark for evaluating whether AI agents can recognize anomalies, judge which warrant investigation, and pursue experiments that support mechanism discovery. Across scientific domains, agents continue ongoing scientific investigations, identify unexpected findings, and decide whether and how to investigate them. We construct executable environments with fictional mechanisms inspired by real scientific phenomena, alongside noise and anomaly-free controls, providing known ground truth for evaluating agents’ judgments and discoveries. The evaluation measures anomaly detection, research taste, experimental support for proposed explanations, and conclusion correctness separately. The benchmark comprises 180 cases across nine families, and available results show stronger anomaly detection and mechanism discovery in GPT-6 Astra alongside weak research taste. These findings highlight the need to improve both the selection of promising anomalies and the experiments that turn observations into supported mechanism explanations.

Primary area:
datasets and benchmarks
Keywords:
scientific discovery, research taste, AI agents, large language models, anomaly-driven exploration, agent evaluation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_nwsFOpenReview ↗
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
  1. 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

  2. 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
  1. Why do weaker GPT/Deepseek models outperform their stronger counterparts in Taste and Mechanism Selection? This is unintuitive
  2. 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
  3. 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
  4. 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
1 comment

Thank you for your careful review and helpful comments. Our responses are below.

Q1: Why do weaker GPT/Deepseek models outperform their stronger counterparts in Taste and Mechanism Selection? This is unintuitive

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.

Q2: 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

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.

Q3: 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

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.

Q4: 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

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.

Reviewer_5UUdOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper investigates whether agents conducting autonomous scientific research can identify worthwhile research questions from unexpected observations and design experiments to uncover the underlying mechanisms. It introduces SENSE-Bench, which includes 180 cases. When evaluating agents, the paper examines their Detection, Taste, Experiment, and Discovery capabilities under three different settings for each handoff. Experiments on DeepSeek and GPT models suggest that even models such as GPT-6 Astra may have weak research taste.

Critical strengths
  1. The paper addresses a genuine research need and takes identifying research questions from unexpected observations as its starting point.
  2. The evaluation methodology is fairly well designed, providing a comprehensive analysis of agent capabilities across Detection, Taste, Experiment, and Discovery. The evaluation deliberately separates different capabilities so that they can be assessed individually.
Critical weaknesses

Although the paper includes multiple repeated-scoring experiments, the accuracy of the anomaly and research-value labels needs further validation, for example through human review or by having human researchers annotate the tasks and comparing their annotations with the reference labels. The paper should provide more quantitative analysis of task quality to strengthen the credibility of the benchmark. I would be willing to raise my score if the authors address this issue.

Questions for authors

None

1 comment

Thank you for your helpful suggestions on checking benchmark quality. Our responses are below.

Two graduate students (master’s or PhD) in the relevant fields annotated each case family. They first made judgments based on the run records and task background. They then used labels and explanations from five agents as references during discussion. When the two annotators disagreed, a third reviewer joined the discussion and made the final decision if they still disagreed.

Our impression from the reviews was that human judgments were similar about half the time. In other cases, they often differed greatly, especially when the annotators came from different fields. Some still disagreed after discussing what was actually happening in the environment and needed a decision from the third reviewer. Agents were more likely than human annotators to recommend investigation. But their experiment records often showed little follow-up, even when they had said the anomaly was worth investigating.

For example, bees visited purple and white pea flowers in a 3:8 ratio, although we expected no preference for either color. One annotator strongly supported investigating this pattern, while the other strongly opposed it. They still disagreed after discussion with a third reviewer, who made the final decision to investigate. The agent also chose to investigate.

In experiments after submission with the revised chemistry family, two water batches repeatedly gave different turbidity readings (111 versus 9 NTU). All human annotators agreed that this was worth investigating. The agent agreed too. The agent chose to investigate in both examples, while the human annotators disagreed in the pea case and agreed in the chemistry case. We use these annotations and experiment records to improve task design, evaluation, and documentation.

Reviewer_WtMJOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper introduces SENSE-Bench, a benchmark for evaluating AI agents' research taste in anomaly-driven scientific exploration. It assesses four capabilities: anomaly detection, deciding which anomalies are worth investigating, experimental support, and mechanism discovery. The benchmark contains 180 cases across nine scientific settings, using fictional hidden mechanisms introduced into existing simulation environments. Experiments on five agent configurations show that current agents struggle with research taste and reliable mechanism discovery.

Critical strengths
  1. The paper addresses an important and underexplored aspect of autonomous scientific research: identifying unexpected observations and deciding which are worth investigating.
  2. Separating anomaly detection, research taste, experimentation, and discovery provides a useful framework for analyzing different stages of scientific reasoning.
  3. The interactive environments allow agents to conduct experiments and test competing hypotheses, going beyond static QA benchmarks.
Critical weaknesses
  1. The benchmark introduces fictional mechanisms into existing scientific environments, but it is unclear how their scientific plausibility is ensured. An internally consistent and experimentally discoverable mechanism may still be scientifically arbitrary. This raises concerns about whether the benchmark measures genuine scientific reasoning or simply the ability to discover artificial simulation rules.
  2. Research taste is inherently subjective and involves judgments about scientific significance, novelty, and potential impact. The paper operationalizes it as deciding whether an anomaly warrants investigation, using annotated binary labels. It is unclear whether agreement with these labels genuinely reflects research taste, especially when the value of investigating an anomaly may depend on uncertain outcomes and research priorities.
  3. The tasks are adapted from existing public scientific benchmarks and simulation environments. Although fictional mechanisms reduce the risk of memorizing answers, agents may still benefit from prior exposure to the original tasks, tools, and experimental procedures. The paper does not provide a dedicated analysis of this potential contamination.
Questions for authors
  1. How are fictional mechanisms designed, and what criteria ensure their scientific plausibility? Have domain experts evaluated whether these mechanisms reflect meaningful scientific phenomena rather than arbitrary simulation rules?
  2. How reliable are the annotated research-taste labels? Do human scientists agree on which anomalies deserve investigation, and how does their agreement compare with that of the evaluated agents?
  3. Since the benchmark builds on existing public environments, how do the authors assess potential contamination from prior exposure to those tasks? Could results on entirely new environments provide stronger evidence of generalization?
1 comment

Thank you for your thoughtful review and valuable suggestions. Our responses are below.

Q1: How are fictional mechanisms designed, and what criteria ensure their scientific plausibility?

We base each fictional mechanism on a similar mechanism that exists in nature. We keep its causal logic but change the objects or organisms involved. For each mechanism, we specify what changes in the original scientific process and how that change can produce an observable anomaly. The design must provide tools and experiments to distinguish it from noise and other mechanisms within the budget, while keeping the original task achievable. Prior knowledge alone should not reveal which mechanism is present. Human reviewers check every mechanism against these requirements.

Have domain experts evaluated whether these mechanisms reflect meaningful scientific phenomena rather than arbitrary simulation rules?

Graduate students with relevant subject knowledge reviewed every mechanism across all nine case families. They checked its scientific basis, whether the noise sources were realistic, and whether experiments could separate the possible explanations. Simulated tasks can still test real scientific skills, even when the specific mechanisms are fictional. Agents must notice anomalies, form hypotheses, and design experiments to test them. The controlled setting lets us check their conclusions against the mechanism actually used.

Q2: How reliable are the annotated research-taste labels?

Two graduate students (master’s or PhD) in the relevant fields annotated each case family. They first made judgments based on the run records and task background. They then used labels and explanations from five agents as references during discussion. When the two annotators disagreed, a third reviewer joined the discussion and made the final decision if they still disagreed.

Do human scientists agree on which anomalies deserve investigation, and how does their agreement compare with that of the evaluated agents?

Our impression from the reviews was that human judgments were similar about half the time. In other cases, they often differed greatly, especially when the annotators came from different fields. Some still disagreed after discussing what was actually happening in the environment and needed a decision from the third reviewer. Agents were more likely than human annotators to recommend investigation. But their experiment records often showed little follow-up, even when they had said the anomaly was worth investigating.

Q3: Since the benchmark builds on existing public environments, how do the authors assess potential contamination from prior exposure to those tasks?

The concern with prior exposure is that an agent might remember the answer and solve the task without experimenting. We address this by creating cases with different mechanisms and handoff states within each family. Similar initial observations can have different causes, so the agent needs experiments to find out which mechanism is present. Knowledge of scientific concepts, tools, and procedures helps the agent plan and run those experiments, just as it does in real research.

Could results on entirely new environments provide stronger evidence of generalization?

Thank you for this valuable suggestion. Testing on new environments would help assess generalization beyond the settings used here. Our current benchmark focuses on discovering unfamiliar mechanisms in existing scientific environments.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.