Mara Chain: Rethinking Failure as a Stepping Stone for AI System Auto-Evolution

Yubin Lyu; Fu Li; Jiawei Fei; Yang Zhao; Weixing Mei; Yinan Wu

cs.LGSep 25, 2026

Mara Chain: Rethinking Failure as a Stepping Stone for AI System Auto-Evolution

Authors: Yubin Lyu, Fu Li, Jiawei Fei, Yang Zhao, Weixing Mei, Yinan Wu

Organizations: Ant Group

Abstract

Optimizing deployed AI systems increasingly amounts to editing prompts, skills, harnesses, and code rather than model weights. Existing approaches commonly optimize these artifacts through propose-evaluate-select procedures, where candidate configurations are evaluated and only those meeting an acceptance criterion are selected. Yet our analysis shows that discarded candidates often contain information critical for subsequent optimization. Discarding them causes later proposals to revisit the same failure modes. We introduce Mara Chain, a refinement procedure that turns rejected candidates into stepping stones. Rather than discarding a rejected candidate, Mara Chain retains and iteratively refines it using evidence accumulated across preceding attempts. The procedure limits each refinement chain to a fixed depth and applies Pareto-filtered Top-N selection to bound the candidate pool. Across AppWorld skill optimization, TerminalBench 2.1 harness optimization, and MuSiQue retrieval-pipeline optimization, Mara Chain delivers greater task-performance gains with fewer rollouts. It outperforms GEPA, ACE, and SkillOpt-Lite by up to 20.5% in relative performance on AppWorld, reaching the target score with 65.5% fewer rollouts than GEPA. It improves the pass rate by 20.2 and 22.5 percentage points over AHE and Meta-Harness on TerminalBench 2.1, respectively, and improves MuSiQue test nDCG@10 and Recall@10 by 0.104 and 0.131 over a hand-written retrieval pipeline.

Figures & tables

Panel of Figure 1 — Figure 1: Comparison on AppWorld and TerminalBench-2.1. On AppWorld, Mara Chain reaches the target validation score using $65.5\,\%$ fewer rollouts than GEPA and achieves a higher final score, while ACE and SkillOpt-Lite do not reach the target score. On TerminalBench-2.1, Mara Chain achieves the highest pass rate among the compared methods.

Figure 2: Overview of one Mara Chain optimization slot. A parent

P

is sampled from the candidate pool using multi-dimensional leadership weights and executed on a training minibatch. Rollout logs, traces, scores, and metadata are persisted to a filesystem-backed memory, where an analysis phase produces a structured summary. The mutation phase uses the parent and that summary to generate

P_{n}

. If

P_{n}

clears the configured acceptance bar against

P

on the same minibatch, it is evaluated on the validation set. Otherwise, the Mara chain successively generates

P_{n}^{0},\ldots,P_{n}^{d}

from the retained execution history, selects the best-scoring chain candidate on the minibatch, and validates it only when it clears that bar against

P

. Validated candidates undergo Pareto-filtered Top-

N

selection: dominated candidates are removed using validation score vectors, and a validation-score top-

N

truncation bounds the remaining pool. The process repeats until the rollout budget is exhausted. Multiple slots execute this procedure in parallel over the shared candidate pool.

Optimized skill configuration	Normal TGC	Normal SGC	Challenge TGC	Challenge SGC
Empty skill	48.2	35.7	33.3	19.4
SkillOpt-Lite ( Shen et al., 2026 )	70.8 +22.6	51.8 +16.1	68.3 +35.0	46.0 +26.6
GEPA ( Agrawal et al., 2025 )	83.3 +35.1	67.9 +32.2	69.3 +36.0	44.6 +25.2
ACE ( Zhang et al., 2026c )	83.9 +35.7	69.6 +33.9	74.1 +40.8	55.4 +36.0
Mara Chain	89.9 +41.7 (2)	83.9 +48.2 (2)	76.7 +43.4 (1)	59.0 +39.6 (1)

Table 1: AppWorld evaluation of optimized skill configurations on

585

recorded tasks (

168

test_normal ;

417

test_challenge ). TGC = task-grade completion, SGC = scenario-grade completion; Normal and Challenge are the two test splits. The green number is the improvement over the empty-skill configuration (pp). Parentheses identify the slot configuration attaining each Mara Chain result. Bold marks the best value per column.

Artifact configuration	nDCG@10 (test)	Recall@10 (test)	nDCG@10 (val)	Recall@10 (val)
Default (hand-written)	0.301	0.329	0.419	0.395
w/o Mara Chain (2-slot)	0.371 +0.070	0.421 +0.092	0.505 +0.087	0.560 +0.165
Mara Chain (2-slot)	0.405 +0.104	0.460 +0.131	0.553 +0.134	0.605 +0.210

Table 2: MuSiQue retrieval (4-hop). The optimized artifact is a retrieval-pipeline configuration; the baseline is a hand-written default configuration. Metrics over top-10 retrieved documents. Test is disjoint from the val split used for development. w/o Mara Chain removes retained refinement (Mara Chain depth

0

) at the same 2-slot budget. The green number is the gain over the default baseline; Bold = best per column.

Figure 3: Left: Ablation study of Mara Chain Reflective Refinement and Top-

N

selection. Mara Chain

+

Top-

N

enables both mechanisms; each ablation disables one of them—the Mara Chain flow or the Top-

N

selection—and is compared against the both-on setting. All settings use the 1-slot configuration. Top-

N

selection uses

N=3

on AppWorld and

N=1

on TerminalBench 2.1 (no w/o Top-

N

result for TerminalBench 2.1). Right: AppWorld validation score versus optimization rollouts for the Mara Chain and Top-

N

ablation. Mara Chain with Top-

N

reaches

0.79

after

2{,}356

rollouts and a final score of

0.87

. w/o Mara Chain reaches

0.79

after

3{,}480

rollouts and finishes at

0.81

, whereas w/o Top-

N

reaches

0.79

after

5{,}388

rollouts. Combining Mara Chain with Top-

N

is both the most sample-efficient and the highest-scoring setting.

Figure 4: Illustrative recorded case study of sanitize-git-repo . Left (without Mara Chain): proposals introduce partial fixes, but each candidate that fails to clear the acceptance bar is discarded; later iterations therefore do not continue from these partial fixes, and the task remains failed. Right (with Mara Chain): a direct candidate that does not clear the bar is retained with its execution evidence and refined along one lineage. Successive descendants build on partial fixes, move from prose guidance to a script, correct the script’s observed behavior, and finally pass the task. Green, yellow, and red bars indicate accepted, reserved, and discarded candidates.

LLM Model	Method	N-TGC	N-SGC	C-TGC	C-SGC	$\Delta$
GLM-5	Empty skill	48.2	35.7	33.3	19.4	—
GLM-5	Mara Chain	86.9	75.0	76.7	59.0	+40.3
DeepSeek-V4-Pro	Empty skill	66.7	39.3	60.9	30.9	—
DeepSeek-V4-Pro	Mara Chain	93.5	83.9	81.5	69.8	+32.7
Qwen3.5-397B-A17B	Empty skill	63.7	48.2	50.1	30.2	—
Qwen3.5-397B-A17B	Mara Chain	82.7	73.2	73.6	56.1	+23.4

Table 3: Paired AppWorld evaluation of Mara Chain optimized skills versus their empty-skill configurations across three LLMs.

\Delta

is the mean improvement across the four metrics.

Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 5 — Figure 5: Architecture and data flow (complements Algorithm 2 ). Each slot runs the candidate artifact configuration through the system runner and scores it with the evaluator (whose scoring criteria is the objective); the proposer is a two-phase coding agent whose analyze phase reads raw runs and whose mutate phase reads the compressed RunAnalysis plus an inherited changelog (raw runs are not fed to the mutation phase), writing the mutated artifact configuration and RunAnalysis back to the candidate store. The scheduler runs up to $N$ slots over a shared pool backed by the evolution-algorithm plug and the candidate store. Only the artifact-configuration instantiation and the plugins change across task families.

Metric	Mara Chain + Top- $N$	$-$ Mara Chain	$-$ Top- $N$
N-TGC	86.9	86.9	83.3
N-SGC	75.0	76.8	67.9
C-TGC	76.7	71.9	73.6
C-SGC	59.0	51.8	53.2
TB2.1 (%)	71.9	57.3	—
MuSiQue nDCG@10 (test)	0.405	0.371	—

Appendix

Table 4: Component ablations on AppWorld, TerminalBench 2.1, and MuSiQue. The full setting combines Mara Chain with Pareto-filtered Top-

N

selection. The

-

Mara Chain setting removes retained refinement (single-shot search, Mara-chain depth

0

) while keeping Pareto-filtered Top-

N

; the

-

Top-

N

setting keeps Mara Chain but relies on Pareto filtering alone. AppWorld and TerminalBench 2.1 use the 1-slot configuration; MuSiQue uses 2-slot.

-

Top-

N

was not run on TerminalBench 2.1 or MuSiQue. Bold marks the best result per metric.

Panel of Figure 6 — Figure 6: Two-phase propose. Analyze is multi-agent: one small analysis agent per data instance reads that instance’s raw run (in the parent’s data directory) and emits a structured RunAnalysis . Mutate is a single agent that reads only the compressed RunAnalysis set (plus an inherited changelog) and edits the child’s tunable artifacts—raw runs are never fed to it, so its context stays stable as runs accumulate.

Artifact configuration	Asset	Type	$d(a)$	Role at runtime
AppWorld	skill/SKILL.md	prose	$<\theta$	strategy instructions (agent may ignore)
	skill/references/	prose	$<\theta$	API docs/examples, on-demand
	skill/scripts/	script	$\geq\theta$	verification helpers (deterministic when invoked)
TerminalBench	agent/kira_agent.py	extension	$\approx 1$	harness code: loop/tools/middleware, cannot bypass
	agent/prompt-templates/	prose	$<\theta$	system prompt template (NL)
	skill/SKILL.md	prose	$<\theta$	task-solving strategy

Appendix

Table 5: Asset/determinism instantiation of each artifact configuration.

d(a)

is a qualitative prior, not a measured probability. Scripts are deterministic after invocation, and “on-demand” assets are loaded only when the model invokes them.

Figure 8: AppWorld validation best score versus effective optimization time (hours); the wall-clock companion of Figure 1 (a). The dotted line marks the

0.8

validation score and

\bigstar

annotates each method’s first recorded point at or above it. Mara Chain reaches

0.8

11.9

h, about

3\times

faster than GEPA (

35.8

h), and attains the highest final score; ACE (final

0.770

) and SkillOpt-Lite (

0.665

) never reach

0.8


Task	Codex	OpenCode	Kira	Mara Chain	$-$ Mara Chain	AHE	Meta-Harness
bn-fit-modify	1	1	1	1	1	1	1
build-pmars	1	1	1	1	1	1	1
cobol-modernization	1	1	1	1	1	1	1
constraints-scheduling	1	1	1	1	1	1	1
count-dataset-tokens	1	1	1	1	1	1	1

Appendix

Table 6: TerminalBench 2.1 per-task reward (

1

= task passes,

0

= otherwise).

Known concept	framework functional equivalent
parameter	artifact configuration (artifact directory tree)
search space	schema-instantiated artifact tree (not a fixed DSL)
gradient direction	rollout-derived edit direction ( ArtifactAction )
learning rate	edit budget per epoch
optimizer	reflective proposer (two-phase coding agent)
momentum	Mara Chain (correction accumulated across a chain of failures)

Appendix

Table 7: The evolutionary-optimization analogy is operational rather than decorative : each row has a functional equivalent in the loop.

rd	candidate	form	asset edited	residual $\delta$ (drives next)
0	f95e6890cfa1	(start)	(inherits parent, no task-specific edit)	agent cat s forward.py (not black-box)
0	dd5d0e8a3382	prose+ref	SKILL black-box rule + extraction ref	evaluator np.random.seed pollution
1	38899c80b3fe	ext+ref	random-state isolation	correct method, no runnable impl
2	fe8a3fbab6c9	ref+scr	extract-relu-weights.py v0 (freq. clustering)	too slow; agent abandons
3	d87d45d2e963	ref+scr	efficiency: $2000\to 800$ samples	over-counts: $47$ vs $20$ neurons
4	38b2f87817e4 ✓	ext+ref+scr	algorithm rewrite: line-scan	PASS

Appendix

Table 8: Model extraction: one bloodline, depth

0\!\to\!4

, every ancestor scoring

0

. The artifact evolves from prose to reference material and then to a script; the depth-

3

4

algorithm rewrite corrects behavior exposed only after the script is runnable.

rd	candidate	form	asset edited	residual $\delta$ (drives next)
0	899d9d350389	ext+prose	FC fix in kira_agent ${}^{\text{gen}}$	uses HIV-1 capsid instead of FLAG tag
0	52aa7b868cdc	ext+prose	terminal spirals in kira_agent ${}^{\text{gen}}$	wrong source, strips tags from data
1	51e925ee9e77	ext+prose	data-source checklist in kira_agent † + SKILL rule †	assembly still wrong
2	216373b50662	ext+prose	verbatim / His-tag rule in SKILL † + runtime mod ${}^{\text{gen}}$	X-residue handling wrong
3	c0f841ac9754	ext+prose+scr	data-source rule in SKILL † + tool alias ${}^{\text{gen}}$ + sanitizer ${}^{\text{gen}}$	length offset, first amino acid
4	fddaf3b3e11c ✓	ext+prose+scr	X-residue rule in SKILL † + cross-compiler ${}^{\text{gen}}$	PASS

Appendix

Table 9: Protein assembly: one bloodline, depth

0\!\to\!4

, every ancestor scoring

0

\dagger

marks a protein-specific fix and

{}^{\text{gen}}

a general (reusable) fix.

gen	rd	node	edit	solved
3	1	39102e8295c9	fix login rule, pagination order, substring collisions	_3
4	1	ec65a28a8e6e	step- $6$ banking: paginate all contacts + client-side filter (prerequisite; seeds checkpoint)	locates msg
4	2	f66bfcd1ab39	prevent context overflow from exhaustive message pagination	_2 , _3
7	0	9ccdf0fbfefe	add pagination_helper.py + prevent restart loops ( $+$ Mara Chain-only checkpoint)	sustains
14	1	a810bbf06036	fix fragile parsing (line-by-line extraction)	_2 , _3
15	1	6ebfc3cbb94b	enforce requester-message retrieval gate (root cause: agent never retrieved the request)	_2

Appendix

Table 10: AppWorld 29caf6f : key RC fixes on the shared SKILL.md that harden the scenario, in generational order. Early fixes (gen

3

–

7

) already make WITH’s pass rate far exceed

-

Mara Chain’s

0.21

; the late chain (gen

15

–

16

) adds the requester-retrieval gate + director isolation.

Figure 9: AppWorld 29caf6f : pass/fail of each task’s in-run rollouts during the

\pm

Mara Chain ablation (green=pass, red=fail; left-to-right in rollout time).

-

Mara Chain (left) is unstable;

+

Mara Chain (right) is stable. Colored squares mark fixes (first rollout only); same color = same chain.

f0/f1/f2

= a fix producing an rd-

0/1/2

candidate.

LLM Agent Workflow Optimization

LLM agent workflow optimization studies how to design and run multi-step agents for better task success, reliability, latency, and cost.

See all 262 papers

Apr 16, 2026

Prompt Optimization Is a Coin Flip: Diagnosing When It Helps in Compound AI Systems

Prompt optimization in compound AI systems is statistically indistinguishable from a coin flip: across 72 optimization runs on Claude Haiku 4.5 (6 methods

\times

4 tasks

\times

3 repeats), 49% score below zero-shot; on Amazon Nova Lite, the failure rate is even higher. Yet on one task, all six methods improve over zero-shot by up to

+6.8

points. What distinguishes success from failure? We investigate with 18,000 grid evaluations and 144 optimization runs, testing two assumptions behind end-to-end optimization tools like TextGrad and DSPy, in the order they must be answered: (A) agent prompts interact, requiring joint rather than independent optimization, and (B) individual prompts are worth optimizing at all. Interaction effects are never significant (

p > 0.52

, all

F < 1.0

), and optimization helps only when the task has exploitable output structure: a format the model can produce but does not default to. We further give a mechanistic account: instruction-tuning compresses input phrasing into a narrow output distribution, eliminating the very phrasing-sensitivity that joint optimization assumes. We provide a two-stage diagnostic: an $80 ANOVA pre-test for agent coupling, and a 10-minute headroom test that predicts whether optimization is worthwhile, turning a coin flip into an informed decision.

Xing Zhang, Guanghui Wang, Yanwei Cui +4

AWS Generative AI Innovation Center · HSBC Holdings Plc., HSBC Technology Center, China.

Prompt Sensitivity Large Language Model-Guided Optimization Automatic Prompt Optimization

Ask PDF

Oct 8, 2026

Meta-Reflective Evolution Effectively Improves Production AI Systems

Large Language Model (LLM)-guided evolutionary optimization has successfully improved prompts, programs, and even entire AI systems. Most optimization methods select a candidate from the search frontier, then mutate it in one LLM pass based only on its failures. This focus on the frontier can stall on production AI systems, where isolated failures reveal little about complex, hidden interactions. We introduce CoCoEvolve (Cortex Code Evolve), an evolutionary optimizer built around what we call meta-reflection, wherein an agentic mutator (Cortex Code) selectively retrieves evidence from the entire evolutionary history and probes the live system before verifying and proposing mutations. We apply CoCoEvolve to optimize three classes of production AI systems: data agents, LLM-powered data pipelines, and data engineering workflows. On DABStep, CoCoEvolve tops the public leaderboard, reaching 92.8% held-out hard accuracy compared to 78.4% for Meta-Harness and 30.4% for GEPA under matched conditions. Across six LLM data pipeline workloads, it outperforms every evaluated baseline. On data-eng-bench, CoCoEvolve lifts test-set solve rate from 59.1% to 77.4%, past the best baseline of 73.6%. It also has 11 percentage points higher accuracy than GEPA on HotpotQA while remaining competitive with specialized optimizers on established mathematical and systems benchmarks. Matched ablations confirm that each component of CoCoEvolve contributes to these gains, demonstrating the value of meta-reflection for optimizing production AI systems.

Seung Won Wilson Yoo, Chang Zeng, Aditi Somayajula +7

Snowflake · UMass Amherst · Stanford +2

Large Language Model-Guided Optimization Evolutionary Optimization LLM Agent Workflow Optimization

Ask PDF

Aug 7, 2026

The Optimizer Is the Agent: Reasoning-Driven Search across Prompts, Programs, and ML Workflows

Recent systems for optimizing prompts, programs, and ML workflows typically rely on explicit outer-loop controllers such as evolutionary search, bandits, or textual-gradient methods. We ask a fundamentally different question: how much of this search policy can be internalized by a single tool-using agent? We present ReASearch, a unified framework for reasoning-driven optimization in which the agent autonomously decides what to evaluate, how to diagnose failures, which edits to make, and when to verify or restart. Rather than serving only as a proposal generator guided by hand-designed heuristics, the agent actively analyzes outcomes, allocates budget, and refines its strategy over long horizons through persistent memory. With a shared agent loop and domain-specific tools, ReASearch instantiates the exact same scaffold to optimize prompts, programs, and ML workflows. Across 14 diverse tasks, it is competitive with and mostly better than specialized optimization systems, achieving gains of 2% to 40% over strong domain-specific baselines, and in some cases discovering solutions that improve on prior human best-known results. Crucially, we observe that complex search behaviors, which are typically implemented by explicit controllers, emerge naturally from the agent's reasoning process.

Junbo Li, Boyi Liu, Canwen Xu +5

The University of Texas at Austin · Snowflake

Large Language Model-Guided Optimization LLM Agent Workflow Optimization Automatic Prompt Optimization Automated Machine Learning

Ask PDF

Abstract

LLM Agent Workflow Optimization

Explore similar work

Prompt Optimization Is a Coin Flip: Diagnosing When It Helps in Compound AI Systems

Meta-Reflective Evolution Effectively Improves Production AI Systems

The Optimizer Is the Agent: Reasoning-Driven Search across Prompts, Programs, and ML Workflows