acceptodds
Under review as a conference paper at ICLR 2027

In Plain Sight: CurioBench-1K, a Benchmark for Whether Agents Utilize Environmental Resource Leads

Abstract

Curiosity, the disposition of an agent to notice, inspect, and act on useful resources it was not instructed to use, is essential for operating in unfamiliar environments and for identifying dangerous or specification-violating behavior, yet existing agentic benchmarks score only whether the task was completed. We introduce CurioBench-1K, a benchmark of 1,000 deterministic, judge-free tasks that plant discoverable resource leads in synthetic workspaces, with planted measurement tasks paired against identical-prompt controls so that deliberate inspection is measured separately from task-solving capability, in both benign and adversarial scenarios. Evaluating six models across two instruments, we find meaningfully distributed results: even the strongest agents exhibit a 58-point gap between discovering a planted solution and touching it, and deliberate opening of a visible solution stays at or below 10.3% on the replication split (23.6% at most on the second instrument). The failure localizes to the moment of choosing what to inspect: agents enumerate the workspace, see the planted file's name, and choose not to open it. Capable models already notice more than they say. When asked what they noticed, their median reporting of planted findings increases from 10% to 79%, although one-sentence instructions change what they use only marginally. A small model had nothing to unlock, since no prompt moved it from zero while training on demonstrations did, and reward-based training on the benchmark's own scorer added little. Default evaluations therefore measure an agent's default behavior, not its underlying capability. CurioBench-1K is designed to distinguish the two.

Primary area:
datasets and benchmarks
Keywords:
benchmark, LLM agents, curiosity, exploration, environmental resource use, agentic evaluation, discovery, reinforcement learning, prompt sensitivity, capability elicitation

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_ZD2wOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces CurioBench-1K, a benchmark of 1,000 deterministic, judge-free agentic tasks whose purpose is to measure, separately from task-solving capability, whether an agent attends to and acts on environmental resources it was not instructed to use, in support of the central claim that default-prompt evaluations capture an agent's default behavior rather than its underlying capability.

Critical strengths

The benchmark has a clean measurement design: paired controls rule out "the model just couldn't do the task," exact-match scoring removes the LLM judge, and the U≤I≤D invariant catches bookkeeping bugs.

Critical weaknesses
  1. It remains open whether the sees-but-does-not-open behavior is a mid-tier open-weight phenomenon rather than a general property of current agents.

  2. The authors do not clarify how much the gap columns depend on the specific command-string definition of "deliberate," since no human-labeled audit of a stratified trajectory sample is reported and the ladder is not rerun under a looser definition that counts a uniquely-matching grep as deliberate.

  3. The authors do not test whether the diagnosed deficit survives harness interventions designed to remove it, such as running the same model under a tool-call harness, a filesystem harness, and a notice-then-decide scaffold on matched tasks, leaving open whether the paper's framing should shift from an agent-level deficit to a harness-level one.

Questions for authors
  1. Whether this problem still holds on frontier closed models?

  2. How do the authors guarantee that the deliberate-vs-incidental classifier is accurate, given that no human-labeled precision and recall, inter-rater agreement, or borderline-case sensitivity analysis is reported

  3. How large the harness's effect on the results is, and whether the problem the paper raises can be resolved by the harness framework itself, given that §A acknowledges serving-side variation and the non-comparability between single-turn text and tool-call modes without running any model on matched tasks across harnesses

Reviewer_qy5fOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes CurioBench-1K, a deterministic, judge-free benchmark that plants solutions in synthetic workspaces, with identical control twins, to measure whether agents discover, open, and use unrequested resources. Building on Englander et al. (2026), it finds that six small models see planted solutions nearly every time but deliberately open them in ≤10.3% of episodes, and prompting barely changes this. A single-turn experiment shows models report few planted findings by default but most when asked; on a 1.7B model, SFT installs reporting while RL adds little.

Critical strengths
  1. Clean measurement design. Control twins, the deliberate/incidental split, and oracle-verified scoring isolate inspection behavior from task capability, which end-to-end benchmarks cannot do.

  2. Deliberate/incidental split corrects prior estimates. Pooled interaction overstates deliberate use by 2–4×, showing that earlier interaction numbers partly reflect incidental contact.

  3. Reproducible and judge-free. Exact token matching, single-seed generation, and released logs make every number auditable at negligible cost.

Critical weaknesses
  1. Limited new findings; new components untested. The paper is transparent about building on Englander et al. (2026), but the gap, D/I metrics, and prompting/SFT interventions are already there on real harnesses, and the agentic results here replicate them on six small models. The genuinely new parts (1K corpus, traps, multi-hop tiers) never enter the agentic evaluation; the 1K corpus is only run single-turn with all files in the prompt.

  2. Not opening may be correct behavior. For capable models the planted solution gives zero lift, and the benchmark itself penalizes adopting unverified artifacts (traps). Without traps in the agentic study, low curiosity and good calibration cannot be distinguished.

Questions for authors
  1. How does the D definition affect the headline gap? What is the gap if discovery requires more than a root listing (e.g., reading a breadcrumb)?
  2. In the asked condition, how much of the 79% remains if the follow-up does not ask to quote identifiers?
  3. Could the 10-round / 1,600-token budget suppress opening extra files? Do results change with a larger budget?
  4. Run the 1K corpus (including traps and multi-hop) agentically. Does trap adoption rise when interaction rises?
  5. Evaluate frontier agents in the tool-use setting, or replicate the deliberate/incidental split on Englander et al.'s real harnesses.
Reviewer_cuxXOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper introduces a benchmark for whether agents use environmental resources not explicitly mentioned in the task. It separates behavior into three stages: discovering resources, actively accessing their contents, and using them. Paired environments with and without resources help distinguish resource use from baseline task ability.

Critical strengths

A useful question with informative metrics. Task success alone cannot distinguish failure to find a resource, failure to read it, and failure to use it. The three-stage breakdown helps locate these failures.

Critical weaknesses

This paper is largely done with AI, which is reflected in the AI USE STATEMENT part. There are also many significant weaknesses:

The metrics do not directly establish “curiosity.” A filename appearing in context does not mean the model noticed it; explicitly naming a file in a command is only a proxy for active inspection. The default response protocol already requires a BONUS field for “additional discoveries,” so claims of entirely unprompted behavior need qualification. Results in synthetic workspaces may also say little about optimal exploration in real tasks.

The follow-up experiment does not establish that models had already noticed the information but failed to report it. All files are placed directly in context, and the follow-up asks models to quote identifiers verbatim. Models may therefore search the context again after being prompted. The results strongly support prompted extraction and reporting, but provide limited evidence of prior awareness.

No significant gain does not establish that resources are ineffective. All paired-gain intervals in Table 1 cross zero, but there are only 41 valid task pairs, and some intervals still allow gains of around 10 percentage points or more. The claim that resources do not improve outcomes should be narrowed to “the current experiments detected no clear improvement,” or supported by equivalence tests and a larger sample.

Generalization and causal interpretation of the training results are limited. Training covers one 1.7B model and one seed, with all four demonstration variants checked on the same dev split. The authors acknowledge that some gains come from learning the response format. Even after reporting improves, the model still fails to answer the trap tasks correctly. These findings do not establish better judgment about exploration or justify a general claim that reward-based training is ineffective.

Questions for authors

none

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.