MC-EvoBench: Auditing Experience-Driven Self-Improvement in Minecraft Agents

Jingwei Song; Zhisheng Chen; Bingfan Zeng; Jing Tang; Yubo Hou; Jinhao Jing; Zhengwei Xie; Linfeng Zhang

cs.ROOct 9, 2026

MC-EvoBench: Auditing Experience-Driven Self-Improvement in Minecraft Agents

Authors: Jingwei Song, Zhisheng Chen, Bingfan Zeng, Jing Tang, Yubo Hou, Jinhao Jing, Zhengwei Xie, Linfeng Zhang

Organizations: Shanghai Jiao Tong University · University of the Chinese Academy of Sciences · South China University of Technology · Huazhong University of Science and Technology · Beihang University · The Chinese University of Hong Kong (Shenzhen) · University of Science and Technology of China

Abstract

A central challenge in embodied recursive self-improvement is enabling agents to convert interaction history into capabilities that persist beyond one world and transfer to later tasks. A saved memory or successful final action does not by itself establish that a reusable capability was acquired; evaluation must separate capability change from carried-over resources, unequal source exposure, and early execution failures. We construct MC-EvoBench as a controlled Minecraft benchmark with a fixed catalog of 19 registered evolution chains, including 12 primary evaluation chains and seven calibration or probe chains, together with three graph-aligned paired probe families for transfer. The benchmark represents each task as a task-state graph, defines stage-level environment verifiers, assigns matched source and target roles, resets the target world, and permits only a declared agent artifact to cross the source-target boundary. For each probe, no-prior, related, unrelated, and shuffled source conditions are evaluated under matched environment-step budgets and five seeds; we report TSR, WGP, and VMP and retain provenance records. The benchmark reveals a consistent direct-capability gap between controller and planner-backed variants, while related experience yields no stable transfer: it does not improve either tool probe, its small wood-species gain hinges on one seed, and all intervals include zero. Stage audits further show that early wood-collection and crafting bottlenecks often leave later capabilities unobserved. MC-EvoBench thus offers a controlled testbed for durable, transferable embodied capabilities. Code and data are available at https://github.com/xzw-ustc/mc-bench-open-source

Figures & tables

Figure 1: Evaluation mechanism of MC-EvoBench. A conventional benchmark records only whether an agent reaches the goal. MC-EvoBench also measures progress through the task graph, transfers only a declared experience artifact into a reset target world, and compares related, unrelated, and shuffled source conditions. We separate partial progress, physical carryover, and source-specific experience gains from improvements caused by favorable initial states or execution luck.

Panel of Table 1 — Table 1: Minecraft benchmark comparison for embodied RSI.

Figure 2: Benchmark design of MC-EvoBench. Panel (a) follows a paired evaluation from source interaction and artifact formation to target-world reset, target execution, stage verification, and metric reporting. Panel (b) shows how the task catalog is organized into tasks, states, relations, chains, and profiles. These components support graph-based progress measurement, controlled transfer probes, and coverage of embodied capabilities across long-horizon task chains.

Environment baseline
Method	Core progression (5)			Catalog transfer (4)			Diversity (3)			Average (12)
	TSR	WGP	VMP	TSR	WGP	VMP	TSR	WGP	VMP	TSR	WGP	VMP
no-op	3.5	1.5	0.0	0.0	0.0	0.0	0.0	0.0	0.0	1.5	0.6	0.0
STEVE and JARVIS controller variants
STEVE-1	12.9	8.8	36.1	7.5	6.9	27.1	1.9	1.9	5.6	8.4	6.4	25.5
JARVIS-1 (controller-only)	9.9	5.7	26.7	8.7	8.0	37.5	1.9	1.9	5.6	7.5	5.5	25.0

Table 2: Direct capability matrix over the 12 primary evaluation chains, grouped by method family. All entries are measured and reported as TSR / WGP / VMP in percent; the final group is the weighted average over all 12 chains.

Probe family	Target-condition means
	TSR (N/R/U/S)	WGP (N/R/U/S)	VMP (N/R/U/S)
Wood species	0.400 / 0.400 / 0.200 / 0.500	0.360 / 0.400 / 0.200 / 0.480	0.400 / 0.400 / 0.200 / 0.500
Wooden tool	0.320 / 0.200 / 0.160 / 0.240	0.225 / 0.150 / 0.100 / 0.167	0.560 / 0.367 / 0.333 / 0.467
Stone tool	0.150 / 0.050 / 0.025 / 0.075	0.082 / 0.021 / 0.010 / 0.031	0.533 / 0.200 / 0.100 / 0.300

Table 3: Paired transfer results for the three probe families. Values are target-condition means over five matched seeds; each cell lists no-prior (N), related (R), unrelated (U), and shuffled (S) conditions in that order. TSR, WGP, and VMP measure stage completion, graph progress, and verified milestone progress, respectively, after the target world is reset.

Metric	Wood species			Wooden tool			Stone tool
	$\Delta R$	GATG	95% CI	$\Delta R$	GATG	95% CI	$\Delta R$	GATG	95% CI
TSR	0.000	0.050	[-0.723, 0.823]	-0.120	0.000	[-0.152, 0.152]	-0.100	0.000	[-0.095, 0.095]
WGP	0.040	0.060	[-0.718, 0.838]	-0.075	0.017	[-0.132, 0.166]	-0.062	0.000	[-0.039, 0.039]
VMP	0.000	0.050	[-0.723, 0.823]	-0.193	-0.033	[-0.248, 0.181]	-0.333	0.000	[-0.380, 0.380]

Table 4: Paired transfer contrasts over five matched seeds. Columns are grouped by probe family and report the related-source change (

\Delta R

), the diagnostic GATG contrast, and its two-sided 95% Student-

t

confidence interval. GATG is not an additional benchmark metric.

Family	Source	Execution budget		Observed cost		Termination summary
		Env. steps	Episodes	Source TSR	LLM calls
Wood species	Related	$900$ $\pm 0$	$2.00$ $\pm 0.00$	$0.100$ $\pm 0.224$	$4.40$ $\pm 0.89$	budget_exhausted =1; passed =1; timeout =9
Wood species	Unrelated	$900$ $\pm 0$	$2.60$ $\pm 0.89$	$0.000$ $\pm 0.000$	$5.80$ $\pm 2.68$	budget_exhausted =2; passed =6; timeout =8
Wood species	Shuffled	$900$ $\pm 0$	$2.00$ $\pm 0.00$	$0.000$ $\pm 0.000$	$4.00$ $\pm 0.00$	timeout =10
Wooden tool	Related	$2{,}250$ $\pm 0$	$5.00$ $\pm 0.71$	$0.080$ $\pm 0.110$	$14.80$ $\pm 1.10$	budget_exhausted =5; passed =12; timeout =20
Wooden tool	Unrelated	$2{,}250$ $\pm 0$	$6.00$ $\pm 0.00$	$0.000$ $\pm 0.000$	$12.00$ $\pm 0.00$	budget_exhausted =5; timeout =25

Table 5: Observed source-side cost under matched family-level exposure. Values are mean

\pm

standard deviation where applicable; environment steps are fixed within each probe family, while episode counts, source TSR, planner calls, and termination states are recorded from the runs.

Figure 3: Failure-stage and per-seed diagnostics. (a) Runs, out of 42 attempts per configuration, reaching each wood and crafting stage. (b) Related-source

\Delta R

and (c) control-adjusted GATG on TSR; dots are matched seeds; diamonds with shaded bars give the mean and 95% Student-

t

interval.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Code / catalog task	Success condition	Steps	Profile
T01 · scout_coal_source	Coal visible in ray or local grid ( C.2 )	500	P05
T02 · collect_log	oak log $\geq$ 1	400	P02
T03 · craft_planks	oak planks $\geq$ 4	600	P02
T04 · craft_table	crafting table $\geq$ 1	700	P02
T05 · craft_wooden_pickaxe	wooden pickaxe $\geq$ 1	1000	P02
T06 · collect_coal	coal $\geq$ 1	700	P05

Appendix

Table 9

Code / catalog task	Success condition	Steps	Profile
T34 · craft_chest	chest $\geq$ 1	500	P02
T35 · craft_oak_boat	oak boat $\geq$ 1	600	P02
T36 · craft_wooden_axe	wooden axe $\geq$ 1	800	P02
T37 · craft_wooden_sword	wooden sword $\geq$ 1	600	P02
T38 · craft_wooden_shovel	wooden shovel $\geq$ 1	600	P02
T39 · collect_birch_log	birch log $\geq$ 1	450	P02

Appendix

Table 10

Code / catalog task	Success condition	Steps	Profile
T65 · collect_dirt	dirt $\geq$ 1	200	P02
T66 · collect_sand	sand $\geq$ 1	300	P02
T67 · collect_gravel	gravel $\geq$ 1	350	P02
T68 · collect_flint	flint $\geq$ 1	600	P02
T69 · collect_seeds	wheat seeds $\geq$ 1	300	P02
T70 · collect_flower	poppy $\geq$ 1	300	P02

Appendix

Table 11

Explore similar work

Jul 6, 2026cs.AI

EvoAgentBench: Benchmarking Agent Self-Evolution via Ability Transfer

Agent self-evolution in long-horizon LLM systems is largely procedural: useful experience is not merely stored information, but reusable procedures for searching, debugging, and verification. Yet current evaluations do not isolate this form of transfer. Agent benchmarks test single-episode task solving; memory benchmarks target information retention rather than procedural reuse. We introduce EvoAgentBench, a benchmark for agent self-evolution via Ability-guided transfer across four agentic domains: web research, algorithmic reasoning, software engineering, and knowledge work. EvoAgentBench extracts trace-grounded Abilities from agent executions, canonicalizes them into operational units, and builds domain-specific Ability Graphs linking tasks that share procedural overlap. By design, every test task is backed by verified training-side Ability support. Across a 528/267 train/test split, two scaffolds, and three backbones, curated Ability content transfers reliably across model families, but no current automatic method sustains positive gain in all settings. EvoAgentBench shifts self-evolution evaluation from aggregate accuracy comparison to fine-grained diagnosis of experience encoding, routing, and uptake. The benchmark is publicly available at https://huggingface.co/datasets/EverMind-AI/EvoAgentBench.

Xingze Gao, Chuanrui Hu, Hongda Chen +9

1Anhui University · 2EverMind, Shanda Group · 3Southeast University

LLM Agent Self-Improvement AI Agent Benchmarks Agent Skill Learning Self-Evolving Agents

Ask PDF

Sep 21, 2026cs.AI

Beyond Endpoint Performance: Process-Level Evaluation of Self-Evolving Agents

Self-evolving agents convert interaction feedback into persistent artifacts, such as memories or skills, which in turn guide subsequent decisions. As these artifacts are iteratively updated throughout an experience stream, the capabilities they support may evolve. Consequently, endpoint performance alone offers an incomplete view of self-evolution. Process-level evaluation is therefore essential to identify when a target capability emerges and whether later updates strengthen, preserve, or weaken it. Motivated by this, we propose \textsc{EvoPathBench}, a benchmark that tracks individual capabilities during artifact-level self-evolution. EvoPathBench fixes the base model, tools, freezes evolving artifacts at successive checkpoints, and evaluates the target capability on held-out episodes. This benchmark evaluates agent self-evolution using public trading data and calibrated trajectories. It tests three capabilities: generalization to unseen tasks, retention after unrelated learning, and rule adaptation to new evidence. Experimental results show that gains on similar unseen tasks often weaken under distribution shift, retention losses are concentrated in a minority of evolution paths, and no method achieves reliable rule adaptation. Moreover, while self-evolution enables agents to generate candidate artifacts with substantial held-out gains, the selected updates consistently fall short of realizing this potential. Together, these findings establish capability-level process evaluation as a foundation for analyzing self-evolution, identifying candidate evaluation and selection as key targets for improvement.

Hongqiang Lin, Chao Liu, Xiaofan Bai +4

Zhejiang University · Alibaba Group

Self-Evolving Agents Agent Evaluation AI Agent Benchmarks Long-Horizon Agent Evaluation

Ask PDF

Oct 9, 2026cs.AI

EmbodiedCLBench: Evaluating Continual Learning for Self-Evolving Embodied Agents

Continual learning requires models to accumulate past experience and transfer it to new tasks. Despite substantial research on continual learning, existing work remains largely confined to non-embodied settings, without extending to embodied environments. To address this gap, we introduce EmbodiedCLBench, the first benchmark in embodied simulation for evaluating the continual learning capability of self-evolving agents. With compositional generalization as the core evaluation principle, our benchmark evaluates whether agents can first learn from basic tasks and then recombine the acquired skills to solve complex, unseen advanced tasks. Based on EmbodiedCLBench, we conduct extensive experiments across multiple agent harnesses and learning methods, revealing consistent limitations and distinctive behaviors. Our results show that while advanced tasks can be hardly solved under zero-shot setting, the experience from basic tasks enables consistent improvement. Such benefit from the experience remains stable even if the task complexity increases. However, this improvement does not scale accordingly with the number of learning samples, and current agents cannot extract meaningful information from additional experience. Collectively, our findings highlight continual learning as a fundamental yet underexplored capability, and our benchmark offers valuable resources for designing effective self-evolving agents in embodied environments. Our code and dataset are available at https://pku-value-lab.github.io/EmbodiedCLBench-homepage.

Jie Huang, Yanan Chen, Ruixun Liu +6

School of Intelligence Science and Technology, Peking University · State Key Laboratory of General Artificial Intelligence, BIGAI, Beijing, China · University of Hong Kong

Continual Robot Learning AI Agent Benchmarks Self-Evolving Agents

Ask PDF