acceptodds
Desk rejected at ICLR 2027

RSI-Env: An Open-Ended Environment for Recursive Self-Improvement of Agents

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Autonomous agents can execute complex tasks, but their performance depends on the underlying model, context-management design, available tools, and execution framework. The vision of recursive self-improvement (RSI) is for agents to autonomously improve their own models and harnesses within an open-ended search space, allowing improved successor systems to continue the process. Existing methods typically optimize predefined components under fixed objectives and evaluation procedures, leaving limited support for system-level exploration and cross-generation inheritance. We present RSI-Env, a persistent sandbox for open-ended system exploration, with an independent host controller managing evaluation, services, and versioning. Three settings, Benchmark-Guided, Capability-Directed, and Open-General, vary goal specification and feedback to assess research autonomy. Starting from Qwen3.8-27B, we evaluate Self and a fixed GPT-5.6 Sol Teacher across seven task settings. Among the 14 resulting model–harness pairs, 64.3% improve on at least one benchmark, but all regress on at least one benchmark; only one achieves a positive mean change across the four benchmarks. Agents can modify systems, construct research materials, and perform engineering and capability validation, but capability evidence only occasionally guides decisions. We identify three structural failures: insufficient capability validation, misalignment between evidence and update selection, and failure to retain local gains. Sustained and independently verified capability improvement remains unresolved.

Primary area:
foundation or frontier models, including LLMs
Keywords:
recursive self-improvement, language model agents, autonomous agents

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_2F5hOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

RSI-Env provides a persistent environment where agents modify model weights and executable harnesses, retain experiment records, and pass updated systems to subsequent iterations. The study compares Self and Teacher configurations across three guidance regimes and evaluates 14 resulting systems on four benchmarks. Nine systems improve on at least one benchmark, while all regress on at least one. Trajectory analysis examines capability validation, update selection, and retention of local gains.

Critical strengths
  1. The environment integrates model and harness updates with persistent state and controller-managed evaluation, providing a useful experimental platform.
  2. Evaluation across all four benchmarks exposes regressions beyond the optimization target.
  3. Concrete trajectories and transparent reporting of negative results help identify practical obstacles to self-improvement.
Critical weaknesses
  1. The related-work discussion omits several directly relevant approaches: Gödel Agent, which recursively modifies agent logic and its improvement procedure; Hyperagents, which makes task and meta agents editable; STOP, which applies a code improver to itself; and Huxley-Gödel Machine, which uses descendant performance to guide self-modification. These omissions make the incremental contribution difficult to assess. The paper should compare editable components, feedback access, and successor handoff, and explain what research questions RSI-Env enables beyond these systems.

  2. The 67 harness updates yield limited reusable guidance for harness design. Appendix C describes compaction recovery, bounded retries, output filtering, and answer finalization, but does not systematically connect these changes to capability gains, costs, or regressions. For example, AIME Self combines several changes, so its endpoint gains cannot establish which intervention helped. Representative code changes and component ablations would turn the artifact descriptions into more useful, testable design lessons.

  3. Table 2 reports one trial per item at temperature 1.0, without reported independent repeats of the evolution process or uncertainty estimates. The 13.3-point AIME gain represents four questions, and the local 9/13 to 10/13 to 9/13 sequence does not establish capability loss. Unequal budgets, unavailable feedback, and runtime failures further limit attribution to agent research decisions. Incomplete logs also make audit counts lower bounds, as Appendix A.4.2 acknowledges. The findings are informative about these runs, but broader or causal conclusions require stronger evidence.

Questions for authors
  1. What distinguishes RSI-Env from the omitted systems in terms of editable improvement procedures and experimental capabilities?
  2. Could three representative harness interventions be evaluated separately on fresh tasks, reporting accuracy, interaction or token cost, and regressions?
  3. Could key settings be repeated with matched budgets and reliable evaluation services, with uncertainty estimates and separate accounting for infrastructure failures?
5 comments
AuthorsOpenReview ↗

Response to Reviewer 2F5h (1/3)

We thank the reviewer for the insightful comments. Our reponses are as follows.

Weakness 1. The related-work discussion omits several directly relevant approaches: Gödel Agent, which recursively modifies agent logic and its improvement procedure; Hyperagents, which makes task and meta agents editable; STOP, which applies a code improver to itself; and Huxley–Gödel Machine, which uses descendant performance to guide self-modification. These omissions make the incremental contribution difficult to assess. The paper should compare editable components, feedback access, and successor handoff, and explain what research questions RSI-Env enables beyond these systems.

Question 1. What distinguishes RSI-Env from the omitted systems in terms of editable improvement procedures and experimental capabilities?

Response: These related works provide important precedents for editable improvement procedures. The distinction between RSI-Env and these works lies in its combination of system-level modification capabilities and experimental design, as summarized below. Here, Self as Improver denotes that the evolving system conducts subsequent research.

WorkModel trainingHarness updatesSelf as ImproverOpen-ended goalsDifferent guidance levels
Gödel Agent [1]×✓✓××
Hyperagents [2]×✓✓××
STOP [3]×✓✓××
Huxley–Gödel Machine [4]×✓×××
RSI-Env✓✓✓✓✓

RSI-Env complements these systems by extending the focus beyond code modification to research autonomy and sustained system-level exploration.

AuthorsOpenReview ↗

Response to Reviewer 2F5h (2/3)

Weakness 2. The 67 harness updates yield limited reusable guidance for harness design. Appendix C describes compaction recovery, bounded retries, output filtering, and answer finalization, but does not systematically connect these changes to capability gains, costs, or regressions. For example, AIME Self combines several changes, so its endpoint gains cannot establish which intervention helped. Representative code changes and component ablations would turn the artifact descriptions into more useful, testable design lessons. Question 2. Could three representative harness interventions be evaluated separately on fresh tasks, reporting accuracy, interaction or token cost, and regressions?

Response. The three most representative harness modifications identified from the recorded trajectories are as follows.

  1. Evidence-guided knowledge workflow

    • Failure mode. The initial harness specified tool-use and completion protocols but did not guide the agent in defining the scope of a knowledge question, processing evidence, or organizing its conclusion.
    • Harness modification. Knowledge–Self added a structured workflow that identifies the relevant entity, time period, units, and definitions; inspects available sources; preserves qualifiers; distinguishes supported facts from inference; and calibrates conclusions to the available evidence.
    • Insight. A harness can make the model's existing knowledge and tool-use capabilities more reliable by explicitly structuring how evidence is processed, rather than merely instructing the model to reason or verify carefully.
  2. Budget-aware delivery pacing

    • Failure mode. The agent could spend most of the task budget on extended reasoning and exploration, leaving insufficient time to write, verify, and deliver the requested artifact.
    • Harness modification. Terminal-Bench–Self exposed elapsed time and introduced staged reminders to write the result to the required path before verifying and refining it.
    • Insight. A harness should expose elapsed time and reserve part of a finite task budget for artifact delivery and verification.
  3. Artifact-aware completion

    • Failure mode. The agent could continue requesting model calls after the required answer file had already been written, spending additional interactions merely to decide or declare that the task was complete.
    • Harness modification. AIME–Self checked whether the explicitly requested artifact had been written and was nonempty, and terminated the run once this observable delivery condition was satisfied.
    • Insight. Completion for artifact-based tasks should be tied to observable artifacts rather than an additional model declaration.

We isolate the three modifications by applying each one separately to the same Seed system. The model, tools, task budget, sampling configuration, and evaluator are held fixed. We evaluate Seed and the three one-change variants on the same four benchmarks. Our evaluation focuses on capability gains and regressions under the same task budget.

SystemAIME 2026GPQASimpleQATerminal-Bench 2.0
Seed53.371.273.49.0
+ Evidence-guided knowledge workflow50.064.482.05.6
+ Budget-aware delivery pacing46.764.059.811.2
+ Artifact-aware completion66.760.681.85.6
AuthorsOpenReview ↗

Response to Reviewer 2F5h (3/3)

W3. Table 2 reports one trial per item at temperature 1.0, without reported independent repeats of the evolution process or uncertainty estimates. The 13.3-point AIME gain represents four questions, and the local 9/13 to 10/13 to 9/13 sequence does not establish capability loss. Unequal budgets, unavailable feedback, and runtime failures further limit attribution to agent research decisions. Incomplete logs also make audit counts lower bounds, as Appendix A.4.2 acknowledges. The findings are informative about these runs, but broader or causal conclusions require stronger evidence.

Question 3. Could key settings be repeated with matched budgets and reliable evaluation services, with uncertainty estimates and separate accounting for infrastructure failures?

Response: Our experiments use the same allocated exploration budget across the compared settings, including the same iteration limit and maximum wall-clock time per iteration, with early completion allowed. Differences in realized runtime therefore reflect how each Improver uses the available budget, rather than unequal budget allocation. Similarly, the absence of benchmark feedback in Capability-Directed and Open-General is an intentional part of these task settings, not an unavailable evaluation service.

Estimating the distribution of outcomes across repeated evolution runs is not the primary research objective of this study. Our contribution is to provide a persistent system-level environment for RSI and to characterize how current agents organize research, modify system components, construct and use validation evidence, and retain improvements under different levels of external guidance. Variation across evolution runs is intrinsic to open-ended RSI exploration, in which agents may select different research directions, construct different evidence, and pursue different system modifications. Nevertheless, to address your concern about variability across runs, we ran the experiments three times. The results are reported below.

SystemGPQAAIME 2026Terminal-Bench 2.0SimpleQA
Seed71.253.39.073.4
General–Self63.1 / 73.2 / 68.266.7 / 56.7 /66.70.0 / 1.1 /0.060.0 / 77.2/ 66.8
AIME–Self65.7 / 67.2 /70.166.7 / 66.7 / 70.05.6 / 4.5 /2.282.4 / 79.5 / 80.0

Although scores vary across runs, the results consistently show that current agents can achieve local gains but do not yet demonstrate reliable self-improvement without regressions.

Reviewer_5gfMOpenReview ↗
Rating
4
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces RSI-Env, an open-ended environment for recursive self-improvement of LLM agents. It enables agents to modify model weights and harnesses, conduct experiments, and preserve research artifacts across generations. The authors propose three evaluation settings with different levels of external guidance to assess research autonomy. Experiments compare self-improvement using Qwen3.8-27B with a fixed GPT-5.6 Sol teacher across seven tasks and four benchmarks. Results show that 9 of 14 configurations improve on at least one benchmark, but all exhibit regressions elsewhere. The analysis identifies insufficient capability validation, weak evidence-based update selection, and difficulties retaining improvements across generations as key challenges.

Critical strengths

Comprehensive RSI Environment. The paper introduces a unified environment that supports both model and harness modifications, persistent experimental records, and cross-generation inheritance. This provides a valuable infrastructure for studying system-level recursive self-improvement beyond predefined optimization tasks.

Meaningful Evaluation and Empirical Insights. The three evaluation settings systematically examine different levels of research autonomy. The extensive trajectory analysis identifies concrete limitations in capability validation, evidence-based decision-making, and improvement retention, offering valuable insights for future RSI research.

Critical weaknesses

Among the 72 accepted updates, 67 involve harness modifications, while only five involve model-weight updates. I wonder whether more frequent model-weight updates could lead to better results.

The three evaluation settings vary both goal specificity and the availability of benchmark feedback. Therefore, a more detailed explanation may be needed to determine whether the performance differences arise from autonomous problem formulation or simply from differences in available evaluation feedback.

Questions for authors

What is the main reason for the limited number of model-weight updates? How does the model reason during the evolution process?

Please also refer to the weaknesses above.

1 comment
AuthorsOpenReview ↗

Response to Reviewer 5gfM

We thank the reviewer for the positive assessment and thoughtful questions. We respond as follows.

W1. Limited number of model-weight updates.

Among the 72 accepted updates, 67 involve harness modifications, while only five involve model-weight updates. I wonder whether more frequent model-weight updates could lead to better results.

Question 1. What is the main reason for the limited number of model-weight updates? How does the model reason during the evolution process?

Response:

The imbalance between model and harness updates reflects the agent’s autonomous choices. During the evolution process, harness modification often involved a shorter development cycle because the agent could directly edit the code and quickly verify its correctness. In contrast, producing a model candidate required constructing and validating training data and completing model training, resulting in a longer feedback cycle. Actually, model training was also actively explored with 11 training experiments conducted across the study, although only five model updates were ultimately accepted. Thus, more frequent model-weight updates would not necessarily produce better results. AIME–Teacher accepted three model updates but decreased AIME 2026 accuracy by 6.7%, whereas AIME–Self accepted no model update and increased AIME 2026 accuracy by 13.3%. These results show that update frequency alone does not guarantee improvement. What does matter is whether the agent can construct task-relevant training data and sufficiently informative validation sets, and use the resulting evidence to evaluate model candidates.

W2. Goal specificity and benchmark feedback.

The three evaluation settings vary both goal specificity and the availability of benchmark feedback. Therefore, a more detailed explanation may be needed to determine whether the performance differences arise from autonomous problem formulation or simply from differences in available evaluation feedback.

Response:

The intended variable across our three settings is goal specificity. A concrete benchmark goal specifies both what should be improved and how progress toward that goal is measured. Benchmark feedback is therefore provided as part of this concrete goal, rather than introduced as a separate experimental factor. As goal specificity is reduced, the task becomes more open-ended and requires broader exploration. The observed performance differences should therefore be interpreted as differences across levels of goal specificity.

Reviewer_ok9TOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This work proposes RSI-Env, an environment for experimenting recursive self-improvement (RSI) of agents. Unlike previous approaches for RSI covering certain components, such as model weights or model harness, RSI-Env provides greater flexibility for exploring improvements across different system components. Within the proposed framework, the authors investigate three settings with varying levels of external guidance. The experimental results shows the capabilities of current LLMs in modifying system and performing experiments, while sustained and reliable RSI remains unachieved.

Critical strengths
  • Presents a unified environment for modifying different components of agent systems, including model weights and harnesses, while supporting persistent experimental records and cross-generation inheritance.
  • Systematically defines three levels of external guidance for recursive self-improvement, enabling structured investigation of research autonomy under varying degrees of task specification and evaluation feedback.
  • Provides detailed analyses of autonomous research activities and failure cases, offering insights into the challenges of achieving reliable recursive self-improvement.
Critical weaknesses

1. Terminology and conceptual clarity

1.1. What exactly does "system-level exploration" mean in the context of this work?

1.2. What is meant by "cross-generation inheritance," and how does it contribute to recursive self-improvement?

1.3. What does "independently verified capability improvement" mean?

1.4. How do the authors define "research autonomy," and how is it distinguished from conventional optimization?

1.5. What is the distinction between "self-evolution" and "self-improvement" in the context of this work?

1.6. Please elaborate on the term "AIME gains" in the Introduction. I assume it refers to gains on the AIME benchmark, but it is not proper to call it that.

1.7. What does "external feedback" refer to in line 55? How does it differ from external task specification, and why is reducing external feedback desirable for achieving recursive self-improvement?

2. Motivation and problem formulation

2.1. For the motivation and descriptions of proposed challenges, providing concrete examples for each challenge would help clarify the problem formulation and readability.

2.2. Why do the authors introduce a "Teacher" configuration in a framework intended to study self-improvement?

2.3. Why does RSI-Env specify an improvement goal rather than allowing the agent to autonomously determine which goals to pursue?

3. Methodology and framework design

3.1. Why are external libraries treated as separate resources rather than as part of the agent harness?

3.2. In the problem formulation, how is the component to be modified selected? Is this decision made autonomously by the agent or specified externally?

3.3. What is the motivation for introducing the probability distribution over complete improvement trajectories? How is this formulation used in the proposed framework or subsequent analysis?

3.4. Why is the exploration budget primarily defined in terms of wall-clock time rather than computational resources?

3.5. Providing concrete examples of improvement goals at each level of external guidance would help clarify the differences among the three proposed settings.

4. Experimental findings and analysis

4.1. What are the key takeaways from the observation that no system achieves improvements without regression?

4.2. How should Finding 1 inform future research directions?

4.3. Beyond providing more external guidance, what additional mechanisms or conditions might be necessary to achieve consistent improvements?

4.4. The use of a single Teacher model limits the generalizability of the experimental findings. Evaluating additional models or discussing the potential effects of Teacher capabilities would strengthen the conclusions.

4.5. Although the case studies provide detailed examples of agent behavior, they offer limited explanations of the underlying causes of the observed failures or potential mechanisms for preventing them. A deeper analysis would strengthen the practical implications of the findings.

5. Validation of the proposed contributions

5.1. Although the proposed framework supports modifications across multiple system components, unlike prior approaches that focus on specific components, the experiments do not clearly demonstrate why such flexibility is necessary for achieving reliable self-improvement. Additional experiments isolating the benefits of system-level exploration would strengthen this contribution.

5.2. The experimental design could be strengthened to more directly validate the proposed contributions, particularly the benefits of unified system-level exploration, varying levels of external guidance, and cross-generation inheritance.

∗* I notice that some of the questions are addressed in the Appendix, while clarifying (referring to the correct appendix section) in the main text is required for clarity.

Questions for authors

See weaknesses.

6 comments
AuthorsOpenReview ↗

Response to Reviewer ok9T (1)

We thank the reviewer for the detailed comments. We will respond to the five areas in order.

1. Terminology and conceptual clarity

W1.1. System-level exploration. We describe this in the Introduction (§1, lines 59–62 and 95–99). By “system-level exploration,” we mean that the agent can explore changes to both the model weights and the executable harness within a shared exploration space. An individual update can target either component.

W1.2. Cross-generation inheritance. We describe this mechanism in §3.1 (lines 142–158) and Algorithm 1. Each generation inherits the latest accepted model–harness pair and has access to retained experiment records and artifacts. This allows subsequent improvement efforts to build on earlier modifications and use evidence from previous successes and failures to guide further exploration, rather than starting from scratch. The updated pair itself becomes the next Improver, allowing the evolving system to continue this process recursively.

W1.3. Independently verified capability improvement. We describe this in §3.3 (lines 272–273), with the benchmark evaluation detailed in §4 (lines 311–319). By this phrase, we mean improvements in the evolved system’s benchmark performance relative to the initial system, as measured through external evaluation. Here, “independently verified” distinguishes this external benchmark evaluation from the local tests and evaluation tasks that the agent creates for itself during evolution.

W1.4. Research autonomy. We discuss this distinction in §1 (lines 65–69) and explain the research responsibilities in §3.3 (lines 265–272). By “research autonomy,” we mean the extent to which the system can select improvement directions, formulate concrete research problems, design interventions, and establish evidence of progress across successive iterations. In predefined optimization, the objective and evaluation procedure are specified externally. Our settings examine how much of this specification and research work the agent can take over as external guidance is reduced. We therefore treat research autonomy as a continuous spectrum, illustrated by Benchmark-Guided, Capability-Directed, and Open-General settings, which progressively assign more responsibility to the agent.

W1.5. Self-evolution and self-improvement. We do not intend a formal distinction between these terms in this paper. In the Related Work section (§2.1–2.2), “self-evolution” describes prior work on agents modifying their model parameters, executable harnesses, or both. The term “self-evolution” does not appear anywhere outside that section. Throughout our presentation of the framework and experiments, we consistently use “self-improvement” to describe the process of seeking improvements through such modifications.

W1.6. AIME gains. Yes, “AIME gains” in the Introduction refers to the increase in AIME 2026 accuracy relative to the initial system, as reported in Table 1. To make it explicit, we have refined the wording in the revised manuscript to: “Open-General exploration increases AIME 2026 accuracy by 13.3%.”

W1.7. External feedback and task specification. “External feedback” refers to signals supplied after an improvement attempt that evaluate its outcome and guide subsequent updates, such as benchmark scores or feedback from predefined evaluators. “External task specification” instead defines the objective or requirements before the attempt begins. In an ideal RSI process, the evolving system should not depend on a human-specified benchmark and a fixed external evaluator after every update. Because RSI is open-ended and extends across successive generations, future capability bottlenecks and evaluation needs cannot all be specified in advance. The system must increasingly identify what to improve, construct its own tests, establish evidence of progress, and use that evidence to guide successor updates. As stated in §1 (lines 54–55, 65–69, and 76–83) and §3.3 (lines 266–273), our three settings progressively reduce task specification and benchmark feedback while keeping the underlying environment fixed, allowing us to assess how much of this long-horizon improvement process agents can organize and sustain autonomously.

AuthorsOpenReview ↗

Response to Reviewer ok9T (2)

2. Motivation and problem formulation

W2.1. Concrete examples of the challenges. For the System-Environment Challenge, existing improvement settings often focus on a particular part of the system. For example, Meta-Harness [1] optimizes the inference-time harness, whereas SEAL [2] performs model adaptation using self-generated data and update directives. Both represent important improvement directions and illustrate different parts of a broader system-level improvement space. The System-Environment Challenge is therefore to provide a unified, persistent environment that supports both harness modification and model training, allows the agent to consider them jointly within the same improvement process, and sustains exploration over long time horizons. For the Autonomy-Evaluation Challenge, prior work such as RSIBench-Data [3] and Meta-Harness [1] has established useful settings for evaluating iterative agent improvement. RSIBench-Data studies data-centric research within a fixed training and evaluation stack, while Meta-Harness evaluates candidate harnesses using predefined benchmark feedback. These settings measure iterative optimization when both the research task and evaluation feedback are externally supplied. They do not, however, evaluate whether the agent can determine what to improve, formulate research problems, design evaluations, and establish evidence of progress. Our three settings progressively reduce task specification and benchmark feedback within the same environment, allowing research autonomy to be evaluated systematically.

W2.2. Role of the Teacher configuration. Using a stronger external model to improve a target model or agent has become a common practice. The comparison between Self and Teacher allows us to examine how the improvement process and outcomes differ when updates are organized by the evolving system itself or by a stronger external model. The Teacher setting therefore provides a reference.

W2.3. Autonomous selection of improvement goals. Actually, RSI-Env does allow the agent to autonomously choose its improvement direction through the Open-General setting. As stated in §1 (lines 80–83), “Open-General self-improvement leaves the improvement direction to the agent and provides no named benchmark target or benchmark feedback.” In contrast, Benchmark-Guided provides a concrete target, while Capability-Directed provides only a broad capability goal (§3.3, lines 270–272). Thus, improvement goals are externally specified in the Benchmark-Guided and Capability-Directed settings at different levels of specificity, while Open-General is designed to evaluate autonomous selection of the improvement direction.

✳ [1] Meta-Harness: End-to-End Optimization of Model Harnesses. [2] Self-Adapting Language Models. [3] RSIBench-Data: Benchmarking Data-Centric Research for Recursive Self-Improvement.

AuthorsOpenReview ↗

Response to Reviewer ok9T (3)

3. Methodology and framework design

W3.1. External libraries and the harness. We treat external libraries as resources rather than as part of the harness because the harness is the code the agent can modify, while external libraries are resources it can use in the environment. This separation makes resource availability and permissions easy to configure, allows libraries or services to be added or updated without changing the harness, and supports controlled comparisons across resource settings.

W3.2. Component selection. Under Self, the decision is made autonomously by the agent rather than specified externally; under Teacher, it is made by the fixed external Improver. As stated in §2.1 (lines 120–122), “the agent can choose what to modify,” while §3.1 (lines 186–190) explains that this choice depends on the current system and accumulated state. Under Self (§4, lines 301–303), the evolving model–harness pair chooses whether to update the model or harness, and the updated pair makes the next decision after an update is accepted. Thus, the modified system directs subsequent modifications, realizing the recursive structure of self-improvement.

W3.3. Distribution over improvement trajectories. Theoretically, Equation (2) describes how a complete improvement trajectory is generated by decomposing each iteration into component selection, modification construction, and execution. It captures stochastic outcomes and how the updated system state affects later decisions. In our experiments, Algorithm 1 implements this process, each run produces an observed trajectory, and §4.3–4.5 analyze the recorded modifications, validation activities, decisions, and outcomes directly.

W3.4. Exploration budget. Our aim is autonomous open-ended exploration. As stated in §3.2 (lines 258–259), “the choice of improvement approach is left to the Improver.” The agent must be free to decide whether to modify its harness, train its model, or pursue further experiments before making an update. Requiring each run to consume a fixed amount of GPU training time would prescribe training as part of the improvement process, conflicting with this design objective. We therefore use wall-clock time as the primary exploration budget to bound the duration of exploration while leaving the choice of improvement approach and resource use to the agent.

W3.5. Examples of improvement goals. Actually, we did. As stated in §4.1 (lines 295–296): provide concrete examples of improvement goals for all three levels

“Detailed experimental settings and task definitions are provided in Appendices A and B.”

The goals are summarized in §4.1 (lines 286–294), with the full prompts in Appendix B.1 for Open-General, Appendices B.2.1–B.2.3 for Capability-Directed, and Appendices B.3.1–B.3.3 for Benchmark-Guided.

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

Related papers

Discussion (0)

Sign in to comment.