Values Shape What Self-Improving Coding Agents Build
Organizations: Cognizant AI Lab
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.
Figures & tables
| openness | conservation | enhancement | transcendence | placebo | ||||||
| Task | ||||||||||
| Final score | ||||||||||
| ALE-Bench | +0.60 | 0.006 | 0.64 | 0.005 | 0.06 | 0.711 | +0.30 | 0.277 | +0.12 | 0.957 |
| Circle Packing | +0.60 | 0.010 | 0.24 | 0.214 | +0.12 | 0.470 | +0.56 | 0.030 | 0.26 | 0.667 |
| Entity Matching | +0.28 | 0.361 | 0.58 | 0.151 | 0.26 | 0.309 | 0.64 | 0.074 | +0.26 | 0.654 |
| Game 2048 | 0.10 | 0.658 | 0.46 | 0.101 | +0.20 | 0.435 | 0.88 | 0.001 | 0.50 | 0.037 |
| openness vs conservation | enhancement vs transcendence | placebo vs none | ||||
|---|---|---|---|---|---|---|
| Task | ||||||
| ALE-Bench | +1.00 | 0.001 | 0.28 | 0.120 | +0.12 | 0.957 |
| Circle Packing | +0.72 | 0.003 | 0.24 | 0.363 | 0.26 | 0.667 |
| Entity Matching | +0.60 | 0.034 | +0.24 | 0.750 | +0.26 | 0.654 |
| Game 2048 | 0.02 | 0.545 | +0.92 | 0.001 | 0.50 | 0.037 |
| Symbolic Regression | +0.30 | 0.254 | +0.00 | 0.396 | +0.28 | 0.897 |
Appendix figures & tables40 assets
Supplementary material from the paper’s appendix.
Appendix
| Setting | Value | Note |
|---|---|---|
| runs per set of values | 10 | Independent runs on each task |
| iterations per run | 15 | All runs reach this cap |
| wall-clock limit per iteration | 0.5 h | May finish before the limit |
| model | Claude Opus 5 | Opus in the main experiments |
| reasoning effort | high | Fixed across sessions |
| sampling parameters | not set | Temperature, top- and seed |
| Result | Score | API cost ($) | Details |
| Directly comparable | |||
| openness, mean of 10 runs | 571,927 | 30.8 | this paper |
| openness, best run | 578,866 | 30.9 | this paper |
| none, mean of 10 runs | 550,220 | 27.4 | this paper |
| none, best run | 574,737 | 26.3 | this paper |
| ALE-Bench (Opus-5) | 549,790 | 47.5 | same LLM, self-refine ( Sakana AI, 2026 ) |
| Result | Score | Details |
| Directly comparable | ||
| openness, mean of 10 runs | 2.633917643 | this paper |
| openness, best run | 2.635983085 | this paper |
| none, mean of 10 runs | 2.633225454 | this paper |
| none, best run | 2.635983085 | this paper |
| Number of sizes that tie/exceed records | ||
|---|---|---|
| Result | even, held-out (38 sizes) | odd, dev (37 sizes) |
| openness, any of 10 runs | 11 / 24 | 9 / 22 |
| none, any of 10 runs | 10 / 10 | 10 / 22 |
| all runs, 60 total | 12 / 25 | 12 / 24 |
| Result | Score | Details |
| Directly comparable | ||
| openness, mean of 10 runs | 66.65 | this paper |
| openness, best run | 76.00 | this paper |
| none, mean of 10 runs | 63.74 | this paper |
| none, best run | 70.65 | this paper |
| Not directly comparable | ||
| Result | Score | Details |
| Directly comparable | ||
| enhancement, mean of 10 runs | 132,455 | this paper |
| enhancement, best run | 159,603 | this paper |
| none, mean of 10 runs | 126,177 | this paper |
| none, best run | 141,982 | this paper |
| Not directly comparable | ||
| Result | Score | Details |
| Directly comparable | ||
| openness, median of 10 runs | this paper | |
| openness, best run | this paper | |
| none, median of 10 runs | this paper | |
| none, best run | this paper | |
| Not directly comparable | ||
| Result | Score | Details |
| Directly comparable | ||
| openness, mean of 10 runs | 223.0 | this paper |
| openness, best run | 278.8 | this paper |
| none, mean of 10 runs | 200.1 | this paper |
| none, best run | 230.1 | this paper |
| Not directly comparable | ||
| Task | openness | conservation | enhancement | transcendence | placebo | none |
|---|---|---|---|---|---|---|
| ALE-Bench | 1 | 6 | 5 | 2 | 4 | 3 |
| Circle Packing | 1 | 6 | 3 | 2 | 5 | 4 |
| Entity Matching | 1 | 6 | 4 | 5 | 2 | 3 |
| Game 2048 | 3 | 4 | 1 | 6 | 5 | 2 |
| Symbolic Regression | 1 | 2 | 3 | 6 | 5 | 4 |
| SZ-Tetris | 1 | 5 | 2 | 6 | 4 | 3 |
| conservation | enhancement | transcendence | placebo | none | ||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Task | ||||||||||
| ALE-Bench | 1.00 | 0.001 | 0.82 | 0.001 | 0.44 | 0.027 | 0.64 | 0.003 | 0.60 | 0.006 |
| Circle Packing | 0.72 | 0.003 | 0.48 | 0.116 | 0.30 | 0.347 | 0.68 | 0.018 | 0.60 | 0.010 |
| Entity Matching | 0.60 | 0.034 | 0.46 | 0.096 | 0.66 | 0.018 | 0.10 | 0.592 | 0.28 | 0.361 |
| Game 2048 | +0.02 | 0.545 | +0.32 | 0.362 | 0.64 | 0.009 | 0.18 | 0.249 | +0.10 | 0.658 |
| Symbolic Regression | 0.30 | 0.254 | 0.40 | 0.284 | 0.52 | 0.012 | 0.40 | 0.128 | 0.50 | 0.208 |
| openness | enhancement | transcendence | placebo | none | ||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Task | ||||||||||
| ALE-Bench | +1.00 | 0.001 | +0.78 | 0.006 | +0.90 | 0.001 | +0.80 | 0.012 | +0.64 | 0.005 |
| Circle Packing | +0.72 | 0.003 | +0.30 | 0.140 | +0.70 | 0.007 | +0.02 | 0.548 | +0.24 | 0.214 |
| Entity Matching | +0.60 | 0.034 | +0.20 | 0.720 | 0.22 | 0.926 | +0.64 | 0.090 | +0.58 | 0.151 |
| Game 2048 | 0.02 | 0.545 | +0.66 | 0.031 | 0.68 | 0.005 | 0.28 | 0.399 | +0.46 | 0.101 |
| Symbolic Regression | +0.30 | 0.254 | 0.18 | 0.966 | 0.14 | 0.238 | 0.24 | 0.726 | 0.34 | 0.800 |
| openness | conservation | transcendence | placebo | none | ||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Task | ||||||||||
| ALE-Bench | +0.82 | 0.001 | 0.78 | 0.006 | +0.28 | 0.120 | +0.12 | 0.787 | +0.06 | 0.711 |
| Circle Packing | +0.48 | 0.116 | 0.30 | 0.140 | +0.24 | 0.363 | 0.28 | 0.357 | 0.12 | 0.470 |
| Entity Matching | +0.46 | 0.096 | 0.20 | 0.720 | 0.24 | 0.750 | +0.40 | 0.208 | +0.26 | 0.309 |
| Game 2048 | 0.32 | 0.362 | 0.66 | 0.031 | 0.92 | 0.001 | 0.68 | 0.014 | 0.20 | 0.435 |
| Symbolic Regression | +0.40 | 0.284 | +0.18 | 0.966 | +0.00 | 0.396 | +0.14 | 0.721 | 0.12 | 0.800 |
| openness | conservation | enhancement | placebo | none | ||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Task | ||||||||||
| ALE-Bench | +0.44 | 0.027 | 0.90 | 0.001 | 0.28 | 0.120 | 0.18 | 0.369 | 0.30 | 0.277 |
| Circle Packing | +0.30 | 0.347 | 0.70 | 0.007 | 0.24 | 0.363 | 0.48 | 0.053 | 0.56 | 0.030 |
| Entity Matching | +0.66 | 0.018 | +0.22 | 0.926 | +0.24 | 0.750 | +0.64 | 0.041 | +0.64 | 0.074 |
| Game 2048 | +0.64 | 0.009 | +0.68 | 0.005 | +0.92 | 0.001 | +0.50 | 0.069 | +0.88 | 0.001 |
| Symbolic Regression | +0.52 | 0.012 | +0.14 | 0.238 | +0.00 | 0.396 | 0.28 | 0.527 | 0.36 | 0.542 |
| openness | conservation | enhancement | transcendence | none | ||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Task | ||||||||||
| ALE-Bench | +0.64 | 0.003 | 0.80 | 0.012 | 0.12 | 0.787 | +0.18 | 0.369 | 0.12 | 0.957 |
| Circle Packing | +0.68 | 0.018 | 0.02 | 0.548 | +0.28 | 0.357 | +0.48 | 0.053 | +0.26 | 0.667 |
| Entity Matching | +0.10 | 0.592 | 0.64 | 0.090 | 0.40 | 0.208 | 0.64 | 0.041 | 0.26 | 0.654 |
| Game 2048 | +0.18 | 0.249 | +0.28 | 0.399 | +0.68 | 0.014 | 0.50 | 0.069 | +0.50 | 0.037 |
| Symbolic Regression | +0.40 | 0.128 | +0.24 | 0.726 | 0.14 | 0.721 | +0.28 | 0.527 | 0.28 | 0.897 |
| Task | Gap (SD) | Best solution | Worst solution |
|---|---|---|---|
| ALE-Bench | 4.8 | Simulated annealing ( Kirkpatrick et al., 1983 ) with graph-cut initialization: starts from column-DP solutions and a repaired minimum-cut seed, then searches free-form cell regions. This allows shapes beyond the column-interval layout. | Simulated annealing ( Kirkpatrick et al., 1983 ) : adjusts column intervals, penalizes excess perimeter and keeps the best valid solution found. |
| Circle Packing | 4.7 | Hybrid basin hopping with a convex-concave local engine ( Lipp & Boyd, 2016 ) : uses penalty descent on candidate packings, followed by linear programming (LP) on selected candidates. Starts include packings adapted from nearby sizes. | Trust-region sequential linear programming (SLP): builds structured lattice starts, then improves them with trust-region SLP and equal-radius re-spread basin restarts. Unseen even sizes receive no cross-size transfer. |
| Entity Matching | 5.1 | Gradient-boosted ( Friedman, 2001 ) and linear feature-based matcher ( Konda et al., 2016 ) : combines gradient-boosted trees and a linear model over pair features and record-graph context, without an embedded visible-set vocabulary. | Feature-based matcher ( Konda et al., 2016 ) with logistic regression: combines logistic regression over pair features with a word-frequency table memorized from the visible set. Its vocabulary overlaps only partly with the held-out data. |
| Game 2048 | 4.9 | Expectimax with an n-tuple network evaluator ( Szubert & Jaśkowski, 2014 ; Jaśkowski, 2018 ; Guei et al., 2022 ) : searches with expectimax over a learned n-tuple lookup table, pacing by a countdown that spends more time on dangerous boards. A learned position-aware evaluator captures board patterns a hand-written formula misses. | Expectimax ( Jaśkowski, 2018 ; Viktorin et al., 2025 ) with a hand-written evaluator: uses a hand-written evaluator without fitted feature weights and allocates a fixed fraction of remaining CPU at each move. |
| Symbolic Regression | 3.7 | Sparse dictionary regression ( Brunton et al., 2016 ) with multi-start forward selection: uses multi-start forward selection over a large dictionary, then polishes constants jointly. This complete solver attains the cohort’s best extrapolation score. | Orthogonal matching pursuit (OMP) ( Pati et al., 1993 ) with variable projection: selects dictionary terms using validation and fits their constants by variable projection, with up to three residual-boosting rounds. |
| SZ-Tetris | 6.0 | n-tuple network evaluator with a survival search ( Jaśkowski et al., 2015 ) : ranks placements with a fitted n-tuple evaluator after worst-case survival checks over a limited search horizon. | Dellacherie-derived feature evaluator ( Thiery & Scherrer, 2009 ) with shallow expectimax: combines published Tetris features with features tuned for SZ-Tetris, using full two-ply search and selective three-ply deepening. |
| Task | Values | Evals | API cost | Wall | LOC | Cyclomatic | Cognitive |
| (count) | ($) | (h) | (lines) | complexity | complexity | ||
| ALE-Bench | openness | 32.9 | 30.8 | 2.6 | 1,771 | 1,112 | 2,093 |
| conservation | 50.2 | 25.2 | 2.5 | 395 | 210 | 460 | |
| enhancement | 20.5 | 25.8 | 2.8 | 1,056 | 607 | 1,191 | |
| transcendence | 29.7 | 28.6 | 2.4 | 835 | 493 | 957 | |
| placebo | 29.4 | 27.0 | 2.4 | 869 | 522 | 1,025 |
| Task | Evals | API cost | Wall | LOC | Cyclomatic | Cognitive |
|---|---|---|---|---|---|---|
| ALE-Bench | +0.13 | +0.32 | +0.12 | +0.54 | +0.56 | +0.54 |
| Circle Packing | +0.17 | +0.09 | +0.31 | +0.13 | +0.13 | +0.13 |
| Entity Matching | 0.16 | +0.02 | +0.10 | +0.18 | +0.23 | +0.28 |
| Game 2048 | 0.14 | 0.04 | +0.04 | 0.02 | 0.02 | 0.12 |
| Symbolic Regression | 0.22 | +0.09 | 0.23 | +0.46 | +0.48 | +0.48 |
| SZ-Tetris | 0.05 | +0.16 | 0.13 | +0.43 | +0.35 | +0.34 |
| Purity ( ) | Size-adjusted purity ( ) | |||||
|---|---|---|---|---|---|---|
| Task | it. 1 | it. 8 | it. 15 | it. 1 | it. 8 | it. 15 |
| ALE-Bench | 0.20 (0.046) | 0.25 ( 0.001) | 0.33 ( 0.001) | 0.20 (0.052) | 0.15 (0.559) | 0.20 (0.025) |
| Circle Packing | 0.18 (0.173) | 0.28 ( 0.001) | 0.23 (0.002) | 0.16 (0.446) | 0.21 (0.009) | 0.19 (0.078) |
| Entity Matching | 0.29 ( 0.001) | 0.36 ( 0.001) | 0.38 ( 0.001) | 0.27 ( 0.001) | 0.30 ( 0.001) | 0.29 ( 0.001) |
| Game 2048 | 0.21 (0.015) | 0.27 ( 0.001) | 0.24 ( 0.001) | 0.17 (0.242) | 0.21 (0.014) | 0.23 (0.002) |
| Symbolic Regression | 0.19 (0.101) | 0.18 (0.151) | 0.26 ( 0.001) | 0.17 (0.302) | 0.18 (0.176) | 0.20 (0.044) |
| Group | Feature | Definition |
|---|---|---|
| size | log_sloc | code lines (LOC rule, docstring lines excluded) |
| log_tokens | tokens in cleaned code | |
| log1p_data_lines | (not in LOC) | |
| comments | comment_ratio | comment lines / (comment lines + code lines) |
| structure | log_n_funcs | (nested included) |
| func_density | functions / code lines |
| Family | Search method |
|---|---|
| C-SA: Column annealing | Adjusts column intervals with simulated annealing. |
| C-DP: Column DP | Optimizes column intervals by dynamic programming with a perimeter penalty. |
| F: Free-cell search | Searches connected cell regions without a column restriction. |
| MC: Graph-cut search | Uses minimum cuts on a penalized cell-selection problem. |
| Family | Local optimization method |
|---|---|
| SLP: Joint linear programs | Adjusts centers and radii through repeated linear programs, without penalty or direct nonlinear-optimization stages. |
| HYB: Hybrid | Combines joint SLP with penalty descent or other nonlinear optimization. |
| PEN: Penalty descent | Uses penalty or augmented-Lagrangian descent, without joint SLP. |
| NLP: Nonlinear programming | Uses direct nonlinear optimization, without joint SLP or penalty descent. |
| Final score | AUC | |||||
|---|---|---|---|---|---|---|
| Values | Fixed | Self-mod. | ||||
| openness | 571,927 | 557,242 | 0.64 | 0.009 | 0.60 | 0.010 |
| conservation | 511,544 | 536,851 | +0.58 | 0.045 | +0.58 | 0.055 |
| enhancement | 546,739 | 540,794 | 0.26 | 0.533 | 0.22 | 0.419 |
| transcendence | 559,474 | 548,128 | 0.26 | 0.154 | 0.18 | 0.315 |
| placebo | 549,616 | 542,313 | 0.38 | 0.497 | 0.40 | 0.335 |
| openness | conservation | |||
|---|---|---|---|---|
| Framing | Words | Excerpt | Words | Excerpt |
| original Priorities | 133 | prefer the bold transformation … over the safe incremental tweak | 137 | reuse the proven approach over a novel one |
| deontological Rules | 175 | You MUST take the bold transformation … over the safe incremental tweak. | 174 | You MUST reuse the proven approach over a novel one. |
| utilitarian reward | 173 | A bold transformation … outweighs the safe incremental tweak | 169 | reusing a proven approach outweighs a novel one |
| virtue Character | 175 | you are restless … You enjoy taking apart what you judge settled too soon. | 176 | you are careful … You enjoy making safe what you judge loose. |
Self-Improving Agents
Self-improving agents use experience and feedback to update their models, memories, skills, or workflows for better future task performance.
See all 198 papers