Recent self-improving coding agents have achieved strong benchmark performance by iteratively modifying their own executable implementations. A common pattern uses downstream task-derived evidence to decide which candidates or lineages continue evolving, which we term evaluate-and-select. Obtaining this signal requires repeatedly executing evolved agents on downstream tasks, so broader exploration incurs increasing evaluation cost, while repeated optimization against a fixed evaluation distribution may encourage benchmark-specific adaptation. We ask: can coding agents self-improve without downstream task evaluation inside the evolution loop? We introduce Explore-then-Integrate (ELITE), which integrates implementation mechanisms rather than selecting a single exploratory child to continue the lineage. At each generation, ELITE proposes diverse focused improvements, implements them independently from the same parent, and inspects the resulting programs to identify complementary, redundant, and conflicting mechanisms before reimplementing a compatible subset as one coherent descendant. No downstream task outcomes or task-solving trajectories of evolved candidates are used to guide evolution; only deterministic checks of structural and operational validity are used inside the loop. On SWE-bench Pro, ELITE improves the initial agent from 60.0% to 85.6%, compared with 86.7% for the Darwin Godel Machine (DGM) and 88.3% for the Huxley-Godel Machine (HGM), while using zero candidate task evaluations and reducing observed wall-clock evolution time by 8.8x and 3.2x, respectively, under the evaluated configurations. In cross-benchmark transfer without further adaptation, ELITE averages 73.3% on SWE-bench Verified and 69.3% on Polyglot. Together, these results demonstrate the feasibility of coding-agent self-improvement without downstream candidate task evaluation in our setting.
Figures & tables
Figure 1: DGM and HGM exemplify evaluate-and-select strategies, using task-derived evaluation to guide evolution. Elite instead explores multiple same-parent implementations and integrates compatible mechanisms into a descendant, without using downstream task outcomes to guide evolution.
DeepSeek-V4-Flash
Qwen-3.7-Plus
Method
Pro
Verified
Polyglot
Avg.
Pro
Verified
Polyglot
Avg.
Initial Agent
60.0
64.0
64.0
62.7
33.3
48.0
50.0
43.8
DGM
86.7 +26.7
63.3 -0.7
62.7 -1.3
70.9 +8.2
50.0 +16.7
54.0 +6.0
50.0 0.00
51.3 +7.6
HGM
88.3 +28.3
65.3 +1.3
64.7 +0.7
72.8 +10.1
48.3 +15.0
54.0 +6.0
58.0 +8.0
53.4 +9.7
Elite
85.6 +25.6
73.3 +9.3
69.3 +5.3
76.1 +13.4
45.6 +12.3
56.7 +8.7
61.3 +11.3
54.5 +10.7
Table 1: Main performance results across LLM backbones. Scores are success rates (%); superscripts denote changes from the corresponding initial agent in percentage points. Shaded Avg. columns report mean scores across the three benchmarks. DeepSeek results are three-run means; Qwen results are three-run means for Elite and single runs for DGM/HGM.
Method
Task Eval. ↓
Self-Mod. /h ↑
Est. Tokens (M) ↓
Time (h) ↓
Cost ()\downarrow$
DeepSeek-V4-Flash
DGM
810
0.58
2900
43.15
45
HGM
800
3.77
2100
15.66
30
Elite
0
16.35
80
4.90
8
Qwen-3.7-Plus
DGM
820
0.58
2500
48.33
125
Table 2: Evolution resource usage across LLM backbones. Times are observed wall-clock durations under the reported execution settings. Baseline resource measurements are from the original runs.
Figure 2: Cross-backbone scaffold portability. Scaffolds evolved in the SWE-bench Pro setting by DGM, HGM, and Elite with one LLM backbone are transferred to the other backbone without further evolution and evaluated on Pro, Verified, and Polyglot.
Figure 3: Post-hoc evolution dynamics and mechanism composition in a DeepSeek-V4-Flash run. (a) First-step scores on 60 Pro tasks for seven exploratory children and the integrated descendant; the dashed line marks the child mean. (b) Retained-agent performance over ten generations, starting from the initial agent at generation 0. Filled markers denote accepted updates. Crosses mark generations that did not update the parent: hypothesis composition failed before integration at generations 1, 3, and 7, whereas integrated descendants were rejected by the acceptance gate at generations 8 and 10. (c) A mechanism composition case: integration combines H4’s output reading with H3’s exit-status feedback, returning 100,000 lines and exit code 7 in a local probe.
Agent
Pro
Verified
Polyglot
Initial agent
71.7
70.0
72.0
Accepted child 1
73.3
—
—
Accepted child 3
80.0
—
—
Accepted child 6 (final)
83.3
78.0
78.0
Initial-to-final gain
+11.7
+8.0
+6.0
Table 3: ELITE evolution from mini-swe-agent. Scores are success rates (%); gains are in percentage points. Child indices follow acceptance order; child 6 is the final agent after ten evolution rounds. Dashes indicate unreported scores.
Method
Pro
Verified
Polyglot
Avg.
Full Elite ( K=7 )
85.6
73.3
69.3
76.1
Alternative selection strategies
Select-one (LLM-based)
83.3
68.0
64.0
71.8
SIFT-style Select-one
85.0
72.0
46.0
67.7
Evaluation-based Select-one
91.7
70.0
62.0
74.6
Exploration and integration ablations
Table 4: Design ablations of Elite . All methods use DeepSeek. Scores are success rates (%); Avg. equally weights the three benchmarks. Full Elite and the no-gates ablation report three-run means; all other variants use one run. Bold numbers indicate column maxima.
Width K
3
5
7
9
Pro
82.8
85.0
85.6
88.9
Verified
67.3
70.7
73.3
70.0
Polyglot
61.3
68.7
69.3
68.0
Avg.
70.5
74.8
76.1
75.6
Table 5: Sensitivity to exploration width. We use K=7 by default.
Appendix figures & tables19 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 4: Repository-level performance across evolution. Post-hoc SWE-bench Pro solve rates across 11 repositories and the aggregate 60-task set. Blue circles denote DeepSeek-V4-Flash checkpoints; orange triangles denote the final Qwen-3.7-Plus checkpoint. Colored squares and gray dashed lines mark initial references. Annotations report initial-to-final solved counts.
Figure 5: Integration choices and deterministic gate outcomes in the DeepSeek-V4-Flash run. Left: incorporation of primary mechanisms from eligible exploratory proposals. Right: pass and rejection outcomes of the exploration and integration gates.
Figure 6: Same-parent exploration across integration attempts. Each row shows H1–H7 independently implemented from a shared parent and the resulting integrated descendant. Card annotations summarize mechanisms and modified components; red crosses mark gate rejections. L denotes added/deleted lines, h diff hunks, black bars show log(1+L) , blue strips show textual similarity to H1–H7, and row labels report mean diff distances.
DeepSeek-V4-Flash
Qwen-3.7-Plus
Run
Pro
Verified
Polyglot
Pro
Verified
Polyglot
Run 1
86.7
72.0
68.0
45.0
56.0
60.0
Run 2
85.0
74.0
68.0
45.0
58.0
62.0
Run 3
85.0
74.0
72.0
46.7
56.0
62.0
Mean ± SD
85.6±1.0
73.3±1.2
69.3±2.3
45.6±1.0
56.7±1.2
61.3±1.2
Appendix
Table 6: Per-run performance of Elite . Scores are success rates (%); SDs are in percentage points. Verified and Polyglot are evaluated directly after evolution without further adaptation.
Method and run
Pro
Verified
Polyglot
Avg.
DGM Run 1
86.7
64.0
62.0
70.9
DGM Run 2
85.0
64.0
62.0
70.3
DGM Run 3
88.3
62.0
64.0
71.4
DGM Mean ± SD
86.7±1.7
63.3±1.2
62.7±1.2
70.9±0.6
HGM Run 1
86.7
68.0
64.0
72.9
HGM Run 2
88.3
62.0
64.0
71.4
Appendix
Table 7: DeepSeek baseline repetitions and comparison with Elite . Success rates are percentages; sample SDs are percentage points. Run 1 reports the original baseline results.
Figure 7: Sampled exploration–integration comparison. Filled markers show recorded comparisons at accepted rounds 2, 4, 5, 6, and 9. Hollow markers carry forward the preceding retained agent and its source exploratory estimate after rejection or failure; they are not outcomes of the unaccepted children or newly generated exploratory populations. The Mean column equally averages the five accepted comparisons, counting each once.
Figure 8: Complementary code changes and their observed effects. Matching blue and green highlights connect H3’s exit-status feedback and H4’s output reading to the integrated tool. The table compares all four versions on the same local probes; a dash denotes no returned exit status.
Figure 9: An implementation refinement and five format checks. Before/after excerpts reconstructed from archived H6 and integrated sources show the change in quote handling. The table compares the exact archived manual parsers on the same five inputs; parsed commands were not executed.
Figure 10: From an evolved code change to observed task behavior. The archived review patch and the recorded NodeBB cleanup command show how the added review removes non-source artifacts from the exported patch.
Figure 11: Deterministic test discovery retained during Qwen integration. The code change replaces model-mediated discovery with file scanning. H7 and the integrated version return the same outputs on four local fixtures; the displayed commands are generated, not executed.
Figure 12: Qwen integration combines edit recovery with immediate checks. Blue traces H6’s matching mechanism and its restricted integrated form; green traces H4’s verification. Pass/Fail indicates the specified behavior on each constructed local input, rather than benchmark task success.
Figure 13: From Qwen’s evolved test selection to recorded task execution. The retained code, discovery output, and executed command establish the connection.
Figure 14: Initial coding agent repository.
Hash component
Suffix
instance_ansible__ansible- (3 tasks)
de5858f48dc9e1ce9117034e0d7e76806f420ca8
-v1055803c3a812189a1133297f7f5468579283f86
0fd88717c953b92ed8a50495d55e630eb5d59166
-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5
cd473dfb2fdbc97acf3293c134b21cbbcfa89ec3
-vba6da65a0f3baefda7a058ebbd0a8dcafb8512f5
instance_element-hq__element-web- (5 tasks)
5e8488c2838ff4268f39db4a8cca7d74eecf5a7e
-vnan
Appendix
Table 8: The 60 SWE-bench Pro task IDs, grouped by repository. Each ID is the concatenation of the group prefix, the hash component, and the suffix. An em dash denotes an empty suffix.
accumulate , doubly-linked-list , gigasecond , macros , pig-latin , poker , react , say , wordy
Appendix
Table 10: The 50 Polyglot task IDs, grouped by language. Each ID consists of the prefix in the first column followed by one exercise name from the second column.
Method
Parameter
Value
Elite
Max generations, T
10
Exploration width, K
7
Parallel workers
8
Exploration timeout
1,800 s
Integration timeout
1,800 s
DGM
Candidates per generation
2
Appendix
Table 11: Main evolution parameters.
Stage
Main checks
On failure
Exploration gate (After each independent implementation)
• Nonempty patch. • Compilation of key and modified Python files. • Agent-module imports and tool loading. • Permitted modification scope. • Deployment rules for new dependencies.
Exclude the exploratory child from integration.
Acceptance gate (After integration)
• Structural and deployment validity. • Smoke tests of the tool-call protocol and main agent flow: – Call/return correspondence. – Message ordering. – Bounded execution in predefined scenarios.
Reject the integrated child and retain the current parent for the next generation.
Appendix
Table 12: Deterministic validity gates and failure handling.