acceptodds
Under review as a conference paper at ICLR 2027

Values Shape What Self-Improving Coding Agents Build

Abstract

Self-improving agents are typically optimized for the results they produce, with less attention to how they produce them. Their instructions give designers limited ways to shape how they work, especially when the task is open-ended. This paper studies whether a short value statement in an agent's persistent instruction file changes not only how well the agent performs but also what it builds. Four higher-order value types in Schwartz's theory of basic human values were evaluated: openness to change, conservation, self-enhancement, and self-transcendence. Across diverse coding tasks, the values were held fixed within each run over multiple iterations of self-improvement. Three findings emerged: (1) Values matter for performance across a diverse set of tasks. The openness values increase performance on held-out data, whereas the opposite conservation values result in scores below those of an agent given no values. (2) Values change what agents build, producing solution algorithms that differ in size, complexity and approach. (3) Human oversight is still needed for designing values. Allowing the agent to revise its own values narrows the differences between sets of values, and values distilled from the AI literature provide no detectable gain over having no values. Properly designed values can thus influence what a self-improving coding agent builds and, where the task permits, how well it performs.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
self-improving agents, recursive self improvement, agentic loops

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

The paper test whether a statement on "value" (function) in the instruction file of a self-improving coding agent changes what it builds and how will it performs. Specifically, the value statements come from Schwart’s high-order value types, openness, conservation, enhancement and transcendence. Evaluated on six coding tasks, this paper finds that 1) openness is better than none while conservation and transcendence is worse; 2) value statements do change what the agent builds, e.g., openness makes larger and more complex solvers; 3) when the agent can revise its own values, the difference between different value statements shrinks.

Critical strengths
  1. I like the direction that looking at coding agent’s actual code beyond simple metrics from some benchmarks. Also the idea of starting from an existing theory of human values (Schwartz) is also interesting.
  2. The experiment settings are rigorious and comprehensive---6 tasks, 2 controls and full reports.
Critical weaknesses
  1. My main concern is that the paper calls a short prompt a "value", but the results do not fully support this framing. The openness and conservation statements are mainly codind/search strategies, e.g., prefer the bold transformation a rewrite, a new representation, a different algorithm vs.take the smallest, most-verified step, so the gap between them is largely expected and relatively intuitive. Moreover, section 4.8 shows that rewording openness while "retaining their original dispositions" significantly lowers its score, and the off-task placebo also has a significant effect on Game 2048. Does this mean that the results reflect prompt wording, not human values which are supposed to be more fundemantal, stable and long-term?
  2. Related to point 1, the link to Schwartz’s theory is mainly in the names. The statements were drafted with an LLM, and the paper does not check that agent behavior matches the intended value. How to define these values mathematically and formally and embed to LLM agent’s value function like classic RL does?
  3. The none control is itself an instruction (You do not have any inherent values). It is not the same as leaving the values section out.
Questions for authors

See weaknesses.

1 comment
AuthorsOpenReview ↗

Response to Reviewer MPqa

Hi, thank you for the review. We appreciate the feedback.

Regarding paper weaknesses:

  1. We define “values” to mean guiding principles designed for AI coding agents, drawing on Schwartz's theory. We do not claim that the agents possess human values. The search guidance is intentional and allows us to evaluate what algorithms the agents develop and how well they perform on held-out data across six tasks.

    Our experiments do show that wording matters moderately. In Section 4.8, the rewrites were designed to express openness as rules, outcomes, or character traits. The wording do not change the gap in performance between different ends of the value spectrum such as openness vs conservation. However, we admit that we cannot fully separate the influence of the guiding principles from the wording for each value. We will clarity the details of the prompt wording comparison in the Introduction and Section 4.8.

  2. Schwartz's theory is used not only as names but to guide the content of the instructions: openness favors trying new approaches, conservation favors preserving proven ones, enhancement emphasizes achievement, and transcendence emphasizes others' welfare (Appendix B). An LLM drafted the statements from Schwartz's accounts, and we edited them (Section 3.1). We will make this mapping clearer in Section 3.1.

    We examine behavior in Appendix F. Its LLM-based analysis finds more exploration and method adoption under openness than conservation on ALE-Bench and Circle Packing. This is consistent with the intended openness–conservation distinction, but is not an independent validation of the value categories. We will state this limitation explicitly in Appendix F.

    In Algorithm 1, V is the instruction text supplied at each iteration. We do not define the four values mathematically or embed them in an RL value function. The model weights and benchmark objective stay fixed. We will make this distinction explicit in Section 3.1 and Algorithm 1. An RL formulation remains future work.

  3. You are correct that “no values” is an explicit instruction: “You do not have any inherent values” (Appendix B). In preliminary experiments, we observed negligible performance difference between this instruction and omitting the values section. We included it to keep the instruction file structure consistent across all values and controls. We will clarify the control wording and rationale in Section 3.1.

Reviewer_ACmWOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper tests whether Schwartz value statements embedded in a self-improving coding agent's instructions affect both its task performance and its algorithmic choices, across six tasks. It reports that opposing values (openness vs. conservation) produce measurably different code and performance, and includes ablations over self-modification, literature-derived values, ethical framing, and model substitution.

Critical strengths
  1. The question is interesting and under-explored. Whether high-level value statements shape an agent's algorithmic approach, not just its performance, is a genuinely novel angle for the agent-alignment literature.
  2. The design is ambitious and systematic, constituting a substantial experimental effort.
Critical weaknesses
  1. The paper rests its novelty on the assertion that whether a value statement changes what an agent builds "has not been tested" (Intro & Related Work), but Serapio-García et al. (2025) and Jiang et al. (2023), which the paper itself cites, already use Schwartz value profiles to characterize LLM behavior, so the increment is to move that profiling from questionnaire responses to a new output modality (code and code metrics), a change of modality rather than of proposition. Framing the manipulation as "Schwartz values" rather than persona prompting reads as relabeling, since the paper's own cited Bardi & Schwartz (2003) notes the value–behavior link is strong for some values and weak for others in humans.

In practice the value statements reduce to explore-versus-exploit instruction variants in the style of role-play prompting. The value-alignment and value-learning literature is absent, Constitutional AI (Bai et al., 2022) is dismissed in one sentence with no empirical comparison, and the direction conflict with Zheng et al. (2024), where personas do not improve accuracy, is cited but never reconciled against this paper's "values improve performance" result.

  1. Table C2 claims openness exceeds AlphaEvolve and ShinkaEvolve, but the actual values are 2.635983085 for openness versus 2.635982839 for ShinkaEvolve, a difference of 2.46e-7 in the seventh decimal place.
  2. The Entity Matching placebo row reports LOC = 2,182, roughly twice openness's 1,128, while its cyclomatic complexity (266) and cognitive complexity (264) are comparable to much smaller solvers. LOC and complexity correlate strongly in every other row, so either LOC wrongly includes embedded data, contradicting the claim that "LOC excludes embedded data" (Line 1102), or the complexity metrics are wrong.
  3. There is selective reporting. The abstract states that openness increases performance, but Table 1 shows a negative δ (−0.10, non-significant) on Game 2048, and the counterexample is omitted.
Questions for authors

See Weakness.

2 comments
AuthorsOpenReview ↗

Response to Reviewer ACmW

Hi, thank you for the review. We appreciate the feedback.

Regarding paper weaknesses:

  1. Our contribution is a controlled test of how persistent value instructions affect an agent's algorithms and held-out performance over repeated coding sessions on six tasks. Prior work already examines prompted Big Five traits in generated posts and vignette responses (Serapio-García et al., 2025; Jiang et al., 2023). We will narrow the broad “has not been tested” claim and expand our discussion of Schwartz-based profiling and alignment (Rozen et al., 2025; Yao et al., 2024).

    We use “values” to mean instruction priorities based on Schwartz's theory: openness and conservation favor exploration versus incremental improvement, while enhancement and transcendence emphasize achievement versus others' welfare. We test these prompts as a whole, including their strategy guidance, and do not separate the effects of value framing from that guidance. We will clarify this scope in the Introduction.

    Constitutional AI (Bai et al., 2022) and BaseAlign (Yao et al., 2024) shape models through training, whereas our study changes instructions while keeping the model fixed. We will expand this discussion and make clear that we did not compare these methods head to head. Zheng et al. (2024) find no average factual-accuracy benefit from personas. Our results concern task-dependent changes in iterative solver design.

  2. For 26-circle packing, the best openness and no values runs both reach 2.635983085, only 2.46 × 10⁻⁷ above ShinkaEvolve. We will update the paper to mention that this is marginal numerical difference, not a meaningful or value-specific advantage. Table C2 also mentions that AlphaEvolve and ShinkaEvolve are not directly comparable with our results because the models and search protocols differ.

  3. The Entity Matching discrepancy is due to a LOC-counting error: a multiline lookup table in one placebo solver was counted as code despite the stated data exclusion. We will correct the affected LOC results and any analyses that use them.

  4. We will revise the abstract to mention that while our findings are generally correct, there are some task-specific exceptions. For example, openness shows no significant gain over no values on Game 2048 (Cliff's δ = −0.10). Nonetheless, the pooled openness–conservation contrast remains statistically significant and is one of our main findings.

Thank you for your reply, I have already improved my score.

Reviewer_VRr1OpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper tests whether short value statements in a coding agent's persistent instructions affect how it revises a solver over repeated sessions. It compares four higher-order values from Schwartz's theory, a length-matched placebo, and a no-values control, using Claude Opus 5 on six tasks: ALE-Bench, Circle Packing, Entity Matching, 2048, Symbolic Regression, and SZ-Tetris. Each condition receives ten independent 15-iteration runs, with sessions capped at 30 minutes; the agent sees development data and a separate scorer evaluates held-out data after each commit. Openness generally produces higher held-out scores and more complex solvers, whereas conservation tends to produce lower scores and smaller solvers. The effects vary by task: enhancement leads on 2048, while differences are weak on Symbolic Regression. Follow-up experiments find that letting agents revise their values narrows the differences, literature-derived values yield no detectable gains, and results vary by model. It also reports task baselines, code metrics, and solver case studies.

Critical strengths
  1. Sections 3 and Appendix A describe a reasonably controlled protocol: values, task, harness, model, and invariants are fixed, with fresh sessions and held-out scoring outside the agent workspace.
  2. The empirical breadth is useful for this question: six heterogeneous tasks, six instruction conditions, ten runs, and 15 iterations per run, with task-specific held-out data.
  3. Tables 1 and 2 report effect sizes alongside permutation p-values. The results show task-specific reversals, with enhancement and transcendence often indistinguishable from the control.
  4. Tables E2-E4 and Appendix F go beyond scores to compare solver size, code-space neighbors, and algorithm choices, including different ALE-Bench search families.
Critical weaknesses
  1. Tables 1, 2, and D2-D6 report many unadjusted p-values across task-by-value comparisons. Because several individual p < .05 results support value-specific claims, these tests are vulnerable to false positives, despite stronger pooled openness effects.
  2. The main experiment uses only Claude Opus 5. Appendix G.4 tests two cheaper Claude models only on ALE-Bench and shows large score drops, so the paper does not establish that the effects generalize across model families or current coding agents.
  3. Table A1 says temperature, top-p, and seed are “not set.” This obscures run reproducibility and whether the ten runs used distinct seeds. No code or run artifacts are identified.
  4. The values, solver complexity, and performance relationship is descriptive. Table E3's pooled correlations are modest (about 0.28 to 0.29 for code size and complexity), and Appendix F uses LLM judgments to classify exploration, method adoption, and maintainability. They do not establish that complexity mediates score gains.
Questions for authors
  1. Can you provide the exact API sampling defaults and run-level identifiers or seeds, and clarify how the ten runs were made independent?
  2. How do the task-specific significance claims change under a prespecified multiple-comparison correction, and which conclusions remain supported?
  3. Can you test the same value conditions on a non-Anthropic model or a second current coding-agent harness?
1 comment
AuthorsOpenReview ↗

Response to Reviewer VRr1

Hi, thank you for the review and we appreciate the feedback.

Regarding paper weaknesses:

  1. The original p-values were unadjusted (Appendix D). We have now added a conservative Bonferroni check separately within each table. Answer 2 reports the adjusted results and supported conclusions.

  2. On ALE-Bench (Appendix G.4), we have some preliminary results with different models. We show that openness has a higher mean than conservation on Sonnet, with little separation on Haiku. Evaluation beyond the Claude family or with another coding harness remains future work.

  3. We used Claude Opus 5 through Claude Code, with high reasoning effort in all main-experiment runs. Table A1's “not set” refers to the LLM parameters that were not user configurable due to limitations in Anthropic API. We will clarify this in the paper.

  4. Our claim is that value instructions affect the algorithms agents build, not that greater complexity causes the score gains. Table E3 already notes that its correlations do not establish this causal link, and Appendix F’s LLM-based assessments are descriptive. The algorithm analysis identifies concrete differences: free-cell/graph-cut search under openness versus column-based search under conservation on ALE-Bench (Appendix F). We will make this distinction clearer in the main text.

Regarding your questions:

  1. Our Claude Code and Anthropic API setup did not expose controls for temperature, top-p, top-k, or a sampling seed. For each task and value, we conduct ten separate runs, each from the task's starting solver in its own workspace. Each iteration starts a fresh conversation and inherits only that run's committed files.

  2. Using higher-precision permutation p-values, we applied Bonferroni correction to Table 2’s 21 comparisons, with separate within-table corrections for Tables 1 and D2–D6.

    Table 2: Bonferroni-adjusted p-values. Significant results (p < 0.05) are bold.

    Taskopenness vs. conservationenhancement vs. transcendenceplacebo vs. no values
    ALE-Bench0.0002271.0001.000
    Circle Packing0.0695621.0001.000
    Entity Matching0.7181261.0001.000
    Game 20481.0000.0031830.792007
    Symbolic Regression1.0001.0001.000
    SZ-Tetris0.5151230.9945551.000
    All tasks pooled0.00001050.4599841.000

    The significant contrasts are openness > conservation overall and on ALE-Bench, and enhancement > transcendence on Game 2048.

    Table 1: comparisons with no values

    TaskComparisonAdj final-score p-valueAdj AUC p-value
    All tasks pooledopenness > no values0.0434350.042070
    Game 2048no values > transcendence0.0151550.012882

    Tables D2–D5: final-score comparisons

    TaskComparisonTablesAdjusted p-value
    ALE-Benchopenness > conservationD2, D30.000325
    ALE-Benchopenness > enhancementD2, D40.005846
    ALE-Benchtranscendence > conservationD3, D50.002923
    Game 2048enhancement > transcendenceD4, D50.004547
    Game 2048no values > transcendenceD50.006495

    Overall, the pooled openness advantage over no values and the openness–conservation contrast remain supported under these within-table Bonferroni corrections.

  3. Unfortunately, running these additional experiments requires significant amounts of additional work, time, and budget, but we do hope to show additional model and harness results in future work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.