Pith. sign in

REVIEW 4 major objections 5 minor 42 references

MEGA: Self-Evolving Agent Optimization Infrastructure via Wisdom Graph

T0 review · 4 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read MEGA proposes a single infrastructure where optimizing an agent and evolving the knowledge that guides optimization are the same process, and reports a 7.03-point aggregate gain over GEPA on four benchmarks.

desk verdict A genuinely thoughtful architecture paper whose core self-evolution claim is not actually tested by its own experiments; the headline gains come from a setup that seeds the Wisdom Graph with golden skills and never closes the loop. read the letter →

arxiv 2608.10504 v1 pith:LI6LH7Y6 submitted 2026-08-11 cs.AI

Jung Hwan Lee , Kyu Ho Lee , Gwang Hoon Yoo This is my paper · ORCID
classification cs.AI
keywords agentoptimizationinfrastructureWisdomGraphPCRdecompositioncompositionalretrievalself-evolvingknowledgeSeed-Epochattributionskillcurationmulti-agent
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

MEGA tries to establish that the bottleneck in agent development is not any single agent or optimizer but the absence of an infrastructure that both optimizes agent systems and accumulates the resulting knowledge. It proposes a closed three-layer cycle: distill durable wisdom from sessions, reason over that wisdom in a typed graph to assemble execution plans, and run evaluation-driven optimization whose evidence feeds back to refine the graph. The paper reports that the full loop outperforms existing skill-curation pipelines on SkillsBench and existing prompt and workflow optimizers on four benchmarks. If the paper is right, the significance is that each optimization run leaves behind validated strategies, curation patterns, and optimization trajectories, so future projects begin from prior evidence rather than from scratch.

What carries the argument

The central object is the typed Wisdom Graph (WG-DB), a directed multi-graph whose nodes are atomic PCR triplets and whose edges carry sufficiency/necessity scores; a role-fluid node pool lets one concept act as the action in one triplet and as the condition in another. Layer 2 reasons over the graph with deduction, abduction, and induction to add unrecorded relations, then solves a Prize-Collecting Steiner Tree formulation to retrieve a connected subgraph that includes low-similarity bridging knowledge. The Seed-Epoch fixed-seed evaluation procedure in Layer 3 is the attribution mechanism that turns optimization outcomes into trustworthy evidence. Together these pieces make the graph self-correcting: verified plans raise evidence confidence, failed combinations are demoted, and contradictions are resolved through graph maintenance.

What would settle it

Run MEGA's Layer 1 and Layer 2 on SkillsBench with a skill pool generated only from ordinary agent sessions, withholding the benchmark's golden skills: if pass rate does not exceed the No Skills baseline or stays far below the pre-seeded-pool result, the claim that operational evidence alone grows the graph is unsupported. Alternatively, optimize a project, then start a second project with a cold graph versus the warmed graph; equal performance would refute the compounding claim.

Watch

Extended reading notes

Core claim

The central claim is that one infrastructure can unify three capabilities that current systems keep separate: optimizing agent workflows, accumulating transferable knowledge, and evolving that knowledge from operational evidence. The mechanism is a typed Wisdom Graph in which every wisdom asset is decomposed into atomic PCR (Primary-Context-Resultant) units with typed dependencies and sufficiency/necessity scores; logical inference expands implicit relations among these units, and Prize-Collecting Steiner Tree retrieval assembles role-differentiated execution plans. In Layer 3, a Seed-Epoch regime fixes the evaluation set within each epoch so performance deltas are attributable to strategy changes, and each verdict updates evidence confidence and transfer rates in the graph. On the paper's measurements, the integrated system reaches 46.5 percent pass rate on SkillsBench with the best token efficiency, and an aggregate score of 76.55 on GPT-4.1 Mini across four benchmarks, 7.03 points above GEPA.

Load-bearing premise

The load-bearing premise is that a pre-seeded skill pool containing the benchmark's own golden skills can stand in for a Wisdom Graph that MEGA's distillation and feedback loop builds and refines itself, so the promised self-evolution is assumed rather than demonstrated.

Editorial extensions

If this is right

  • New agent projects could begin with previously validated skills, strategies, curation patterns, and optimization trajectories instead of restarting strategy search from scratch.
  • Optimization would target the entire heterogeneous workflow, including code nodes, LLM calls, and tool-using agents, rather than prompts alone, because Layer 3 jointly optimizes all node types.
  • Retrieval quality should improve with use: every attributed verdict updates evidence confidence and historical transfer rates, and maintenance merges duplicates, resolves contradictions, and updates stale content.
  • The reported results imply that even with smaller validation sets than baseline optimizers use, the closed loop can converge to higher scores by composing wisdom rather than only rewriting prompts.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A testable extension the paper leaves implicit: if the loop truly compounds, warmed graphs should require fewer optimization epochs or less validation data per new project; a longitudinal study tracking convergence time against cumulative verdict count would test this directly.
  • Because Layer 1 distills wisdom from sessions before any model-specific optimization, the same Wisdom Graph might transfer strategic guidance across different base LLMs; the paper does not test cross-model transfer.
  • Explicit accumulation of negative evidence, encoded as 'when not to invoke,' suggests a stronger claim than ordinary skill libraries: the graph could steer future retrieval away from known failure combinations, which would show up as reduced error rates on repeated task families.
  • Layer 3's ability to synthesize evaluation data when none exists implies that the infrastructure also shifts the human role from dataset builder to objective-setter and data-strategy approver; the paper keeps this at the level of pipeline design and marks text-only generation as the current limit.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes MEGA, a three-layer infrastructure for agent-system optimization. Layer 1 distills agent sessions into validated 'wisdom assets' via clustering and behavioral A/B testing; Layer 2 decomposes these assets into atomic PCR (Primary-Context-Resultant) units stored in a typed Wisdom Graph, performs deductive, abductive, and inductive inference to expand the graph, and retrieves execution plans via a Prize-Collecting Steiner Tree formulation; Layer 3 runs a multi-agent optimization loop over heterogeneous workflows, attributes performance deltas to specific changes via a fixed-seed 'Seed-Epoch' regime, and feeds verdicts back into the graph. The central claim is that this closed loop lets optimization knowledge compound across projects, so new projects start from previously validated wisdom. The empirical sections report a SkillsBench curation study (Table 4) and a four-benchmark optimization study against prior optimizers (Table 5), with an aggregate score of 76.55 versus GEPA's 69.52 on GPT-4.1 Mini.

Significance. If the self-evolving loop were actually demonstrated, the paper would address a real gap in the agent-optimization literature: current optimizers do not accumulate transferable knowledge, and current skill libraries do not reason compositionally over that knowledge. The architecture is specified in unusual detail, including the PCR formalism (Definition 1), the Seed-Epoch attribution scheme (Definition 2), verdict feedback (Definition 3), and the PCST retrieval objective (Eq. 5). The paper also ships a repository link for benchmark configuration and results, which is a genuine strength, and it compares against published baselines rather than only internal ablations. However, the empirical sections do not test the mechanism that defines the paper's contribution: the cross-project, evidence-driven self-evolution of the Wisdom Graph. The evaluation is set up so that the graph already contains the golden skills for the test tasks, and the optimization study measures a single pass over one static pool, so the reported gains cannot be attributed to the claimed loop.

major comments (4)
  1. [Section 7.1, Table 4] The skill pool for the curation evaluation 'comprises 4,207 assets including SkillsBench's golden skills for all curation systems.' This means MEGA's WG-DB retrieval is evaluated under an oracle condition in which the target task's reference skills are present in the pool, and the same is true for the comparison systems. As a result, the 46.5% pass rate measures retrieval and plan assembly from a pool that already contains the correct solutions; it does not measure Layer 1's distillation of reusable wisdom from raw sessions, nor does it test the cross-project accumulation that the paper's central claim requires. An ablation with the golden skills withheld, or with a pool built only from MEGA's own Layer 1 output, would be needed to attribute the result to the Wisdom Graph rather than to the oracle pool.
  2. [Section 7.2 and Eq. (1)] The optimization experiment does not exercise the self-evolving loop that Eq. (1) formalizes. MEGA's Layer 3 is run on the same static 4,207-asset graph evaluated in Section 7.1, and the verdicts produced during a single benchmark are fed back into that same pool; there is no measurement over multiple projects, no sequence of graph updates, and no evidence that cross-project optimization trajectories change the starting point of a later optimization. The abstract's claim that 'optimizing an agent system and evolving the knowledge that guides optimization are one and the same process' is therefore assumed rather than demonstrated. A minimal test would run MEGA on a first project, then a second project with the updated graph, and compare against a graph that received no feedback; the paper reports no such comparison.
  3. [Table 5, Section 7.2] The headline +7.03 aggregate improvement over GEPA is reported as a single set of point values with no error bars, no multiple-seed variance, and no significance tests. Baseline scores for MIPROv2, TextGrad, GEPA, and Feedback Descent are cited from the respective prior papers rather than reproduced under identical conditions, while MEGA is run with smaller validation sets (100 vs. 300 for HotpotQA, as stated in the setup). These differences make the comparison unmatched: the reported gain could result from evaluation-protocol differences, validation-set size, or random seed rather than from MEGA's mechanisms. The claim that 'MEGA's optimization loop requires less validation data to converge' is not supported by any controlled comparison that varies validation size.
  4. [Sections 1.2 and 5.6] The Scope of Disclosure states that 'algorithmic details are selectively disclosed at the level of design objectives and formal properties,' and Section 5.6 explicitly says that 'detailed parameter settings and statistical assumptions are omitted.' The architecture relies on a large set of free parameters and unspecified functions: the BIRCH radius threshold T (Eq. 3), the cold/warm gating thresholds θ_cold and θ_warm (Eq. 7), the blending threshold k (Eq. 8), the promotion thresholds n_min and r_min (Eq. 9), the evolution threshold θ_evo (Eq. 10), and the prize/cost functions of the PCST objective (Eq. 5). Because none of these are given concrete values or estimation procedures, the central derivation cannot be independently checked, and the reported experimental results cannot be reproduced from the manuscript alone.
minor comments (5)
  1. [General] The manuscript contains repeated OCR-like artifacts, including 'with-out' in the abstract, 'relations;' in the abstract, and missing spaces in Section 1; these should be corrected throughout.
  2. [Table 4 and Section 7.1] The efficiency metric is defined in the text as 'pass-rate improvement per additional megatoken consumed relative to the NoSkills baseline' but the table reports 'Efficiency (score/Mtok)' and the caption defines it as 'pass rate per megatoken consumed'; these definitions are inconsistent and should be unified.
  3. [Figure 10 and Table 4] Pass rates are computed from 5 attempts per task across 84 tasks with a stochastic agent, yet no error bars or confidence intervals are reported; at minimum, a binomial confidence interval would clarify whether the differences among the three curation methods are meaningful.
  4. [Section 7.2] The statement that HotpotQA and PUPA 'approach the ceiling reachable through workflow and prompt optimization' is presented without evidence; the paper does not provide an oracle or ceiling analysis to justify this interpretation.
  5. [References] Several references list only partial author lists or use placeholder years (e.g., some 2026 arXiv preprints), which is acceptable for a preprint but should be checked for consistency with the journal's citation requirements.

Circularity Check

0 steps flagged · score 0.0 of 10

No derivation-level circularity: Eq. (1) is an architectural definition, Tables 4-5 are direct measurements, and the untested cross-project self-evolution claim is a validity gap rather than a circular reduction.

full rationale

The paper's formal backbone is an architecture specification, not a theorem whose conclusions are preloaded into its premises. Equation (1) defines W(t), Pi(t), E(t), and T3(t) as the outputs of the three layers and then asserts that T3(t-1) and E(t-1) feed back into Layer 1 and Layer 2; the abstract's statement that "optimizing an agent system and evolving the knowledge that guides optimization are one and the same process" restates this design definition rather than deriving a non-trivial consequence from independent premises. The quantitative claims are direct measurements: Section 7.1 compares curation methods on a shared 4,207-asset pool, and Section 7.2 compares optimizers on common ground-truth datasets with baseline scores cited from GEPA and Feedback Descent. Including SkillsBench's golden skills in the common pool is a deliberate control for curation quality, and since all curation conditions share the same pool and MEGA still fails more than half of the tasks, the 46.5% pass rate is not forced by construction. Similarly, Section 7.2's use of the same Wisdom Graph is an intended within-run optimization condition, not a fitted parameter renamed as a prediction. The genuine problems are external-validity and disclosure gaps: Section 7 never exercises cross-project evidence accumulation, and the "Scope of Disclosure" states that algorithmic details are selectively disclosed, so the operationalized self-evolution loop cannot be independently checked from the report. These are substantive scientific limitations, but they are not circularity of the kind where a prediction is equivalent to a fitted input, a cited uniqueness theorem forces the choice, or a result is identical to its assumptions by definition.

Assumptions & free parameters 6 free parameters · 5 assumptions · 3 invented entities

The central claim rests on a large set of undisclosed thresholds, scoring functions, and inference rules, plus the unverified assumption that a static pool containing golden skills can validate a self-evolving loop. The free parameters are not fitted to data in a reproducible way; they are simply omitted, making the architecture impossible to reconstruct from the paper alone.

free parameters (6)
  • BIRCH radius threshold T = not disclosed
    Eq. (3) controls cluster granularity in Layer 1; the value affects the number of behavioral patterns and therefore wisdom quality, but it is not reported.
  • Gate thresholds theta_cold and theta_warm = not disclosed
    Eq. (7) switches between exploration and exploitation in curation; these thresholds determine which wisdom enters plans, yet no values are given.
  • Blending threshold k in transfer rate = not disclosed
    Eq. (8) controls how quickly historical transfer estimates converge; the value of k is not specified.
  • Promotion thresholds n_min and r_min = not disclosed
    Eq. (9) gates whether a combination becomes a curation strategy; both thresholds are left unspecified.
  • Evolution threshold theta_evo = not disclosed
    Eq. (10) triggers content evolution when revision suggestions accumulate; the threshold is not given.
  • PCST prize and cost functions = not disclosed
    Eq. (5) requires defining query relevance prizes and causal-weakness costs; no concrete functional form or parameter values are provided.
assumptions (5)
  • domain assumption Inference rules for deduction, abduction, and induction over PCR triplets produce valid new relations.
    Section 5.3 asserts these inferences expand implicit relations, but no soundness proof or empirical validation of the inferred edges is provided.
  • domain assumption Fixed-seed evaluation within an epoch eliminates data variance and attributes deltas to strategy changes.
    Section 6.4 (Definition 2) claims identical evaluation data makes deltas attributable, but this ignores stochasticity in model decoding and possible overfitting within the epoch.
  • domain assumption Behavioral A/B validation with an LLM teacher-student protocol provides a reliable performance-lift estimate.
    Section 4.4 relies on generated test cases and model-based comparison; no evidence is given that these estimates are unbiased or statistically stable.
  • ad hoc to paper The 4,207-asset pool containing SkillsBench's golden skills is an appropriate evaluation substrate for wisdom curation.
    Section 7.1 includes golden skills for the test tasks in the retrieval pool, conflating retrieval of reference solutions with curation quality.
  • domain assumption PCST over the undirected projection of the Wisdom Graph yields plans that improve task performance.
    Section 5.4 adopts PCST as the retrieval objective; no empirical comparison against simpler retrieval or justification of prize and cost definitions is given.
invented entities (3)
  • Wisdom Graph (WG-DB)
    purpose: Typed multi-graph of PCR units that supports reasoning, retrieval, and self-evolution
    Introduced as the core substrate; its benefits are only evaluated in a static, potentially leaked setup, not as an independently validated artifact.
  • PCR triplet (Primary-Context-Resultant)
    purpose: Atomic decomposition unit for wisdom assets
    No independent evidence that decomposing skills into PCR units improves composition; the benchmark pools already contain golden skills.
  • Curation patterns and optimization trajectories
    purpose: Higher-order wisdom types that encode which combinations and trajectories worked
    Their value is asserted in the design sections but not demonstrated in the experiments.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MEGA: Self-Evolving Agent Optimization Infrastructure via Wisdom Graph." pith.science (2026). https://pith.science/paper/LI6LH7Y6

@misc{pith2026260810504,
  author       = {Pith},
  title        = {Pith review of: MEGA: Self-Evolving Agent Optimization Infrastructure via Wisdom Graph},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LI6LH7Y6}},
  note         = {Machine review of arXiv:2608.10504}
}
read the original abstract

As coding agents increasingly handle implementation, the central challenge shifts from building individual agents to building an infrastructure that systematically improves them. Current approaches optimize agent systems without accumulating transferable knowledge, accumulate knowledge without compositional reasoning over it, and lack a mechanism for that knowledge to self-evolve through operational evidence. MEGA (Meta Evaluation-Grounded Adaptation) addresses these gaps as a self-evolving infrastructure: each optimization cycle produces durable assets, compositional reasoning over those assets guides subsequent optimization, and operational evidence refines both the accumulated wisdom and the reasoning that governs it. Layer 1 distills reusable wisdom from agent sessions through behavioral-pattern clustering and empirical A/B validation, transforming each process into a durable asset. Layer 2 decomposes these assets into atomic PCR (Primary-Context-Resultant) units within a typed Wisdom Graph and performs deductive, abductive, and inductive reasoning to expand implicit relations; it then assembles context-specific execution plans through compositional retrieval that surfaces bridging knowledge unreachable by embedding similarity alone. Layer 3 performs multi-agent collaborative optimization over heterogeneous agent workflows (code nodes, LLM calls, and tool-using agents), attributing improvement effects to specific strategy changes through controlled evaluation that eliminates data variance. Evidence fed back from Layer 3 drives the self-evolution of both the curation strategies that govern wisdom composition and the optimization trajectories accumulated across runs. The result is an infrastructure in which optimizing an agent system and evolving the knowledge that guides optimization are one and the same process.

Figures

Figures reproduced from arXiv: 2608.10504 by the authors.

Figure 1
Figure 1. MEGA architecture. W comprises four wisdom types (skill, strategy, curation pattern, optimiza￾tion trajectory). Π(q) is a context-specific execution plan. Evidence E (red) includes verdicts, curation feedback, and optimization trajectories that Layer 2 uses to generate curation pattern and optimization tra￾jectory wisdom within WG-DB. Execution sessions T3 (purple) feed into Layer 1 as new extraction input. 3.1 Inte… view at source ↗
Figure 2
Figure 2. Layer 1 extraction pipeline. Client-side privacy filtering precedes all processing. BIRCH discov [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. PCR Reasoning. The directional P→C→R structure of PCR provides the structural basis for three classical modes of reasoning. Solid lines denote observed edges; red dashed lines denote inferred edges. 5.4 PCST-Based Compositional Retrieval Embedding similarity search returns only knowledge directly similar to the query. It cannot discover “bridging knowledge”—knowledge essential for connecting multiple skills but exhi… view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: PCST-based retrieval. Seed nodes have high query similarity, while Bridge nodes have low [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]
Figure 5
Figure 5. Figure 5: Self-evolution feedback loop. Per-wisdom attributions from each curation session update three [PITH_FULL_IMAGE:figures/full_fig_p015_5.png]
Figure 6
Figure 6. Figure 6: Three graph hygiene operations. Deduplication merges semantically equivalent PCR triplets via [PITH_FULL_IMAGE:figures/full_fig_p016_6.png]
Figure 7
Figure 7. Figure 7: Prompt-centric vs. heterogeneous workflow optimization. Prompt-centric optimizers treat each [PITH_FULL_IMAGE:figures/full_fig_p017_7.png]
Figure 8
Figure 8. Figure 8: The Adapt pipeline of Layer 3. Asset Scan inventories existing system state; Adaptive Entry [PITH_FULL_IMAGE:figures/full_fig_p019_8.png]
Figure 9
Figure 9. Figure 9: Coverage-driven data augmentation. The system classifies the evaluation structure, discovers [PITH_FULL_IMAGE:figures/full_fig_p021_9.png]
Figure 10
Figure 10. Figure 10: Pass rate vs. efficiency on SkillsBench. Higher and further right is better. [PITH_FULL_IMAGE:figures/full_fig_p024_10.png]
Figure 11
Figure 11. Figure 11: Per-benchmark comparison on GPT-4.1 Mini across four benchmarks. Each panel uses a [PITH_FULL_IMAGE:figures/full_fig_p026_11.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

42 extracted references · 22 canonical work pages

  1. [1]

    Fromdatatowisdom

    RussellL.Ackoff. Fromdatatowisdom. Journal of Applied Systems Analysis,16:3–9,1989

    work page 1989

  2. [2]

    GEPA:Reflectivepromptevolutioncanoutperformreinforcementlearning

    Lakshya A Agrawal, Shangyin Tan, Dilara Soylu, Noah Ziems, Rishi Khare, Krista Opsahl-Ong, ArnavSinghvi,etal. GEPA:Reflectivepromptevolutioncanoutperformreinforcementlearning. In International Conference on Learning Representations (ICLR),2026. OralPresentation

    work page 2026

  3. [3]

    Agent KB: Leveraging cross-domain experience for agentic problem solving,

    Zhuoqun Chen et al. Agent KB: Leveraging cross-domain experience for agentic problem solving,

    work page

  4. [4]

    arXiv preprint arXiv:2404.16130,2024

    Darren Edge, Ha Trinh, Newman Cheng, Joshua Bradley, Alex Chao, Apurva Mody, Steven Truitt, andJonathanLarson.Fromlocaltoglobal: AgraphRAGapproachtoquery-focusedsummarization. arXiv preprint arXiv:2404.16130,2024

    arXiv 2024

  5. [5]

    A general approximation technique for constrained forestproblems

    Michel X Goemans and David P Williamson. A general approximation technique for constrained forestproblems. SIAM Journal on Computing,24(2):296–317,1995

    work page 1995

  6. [6]

    Flow: Modularized agentic workflow automation

    Boye Gu et al. Flow: Modularized agentic workflow automation. InInternational Conference on Learning Representations (ICLR),2025

    work page 2025

  7. [7]

    DS-Agent: Automated data science by empowering large language models with case-basedreasoning

    Siyuan Guo et al. DS-Agent: Automated data science by empowering large language models with case-basedreasoning. In International Conference on Machine Learning (ICML),2024

    work page 2024

  8. [8]

    LightRAG: Simple and fast retrieval-augmented generation, 2024

    Zirui Guo et al. LightRAG: Simple and fast retrieval-augmented generation, 2024. arXiv:2410.05779

    arXiv 2024

Show all 42 references
  1. [9]

    FromRAGtomemory: Non-parametriccontinuallearningforlarge languagemodels

    BernalJiménezGutierrezetal. FromRAGtomemory: Non-parametriccontinuallearningforlarge languagemodels. In International Conference on Machine Learning (ICML),2025

    2025

  2. [10]

    Upskill: Agent skill generation and evaluation

    Hugging Face. Upskill: Agent skill generation and evaluation. https://github.com/ huggingface/upskill,2025

    2025

  3. [11]

    https://ghuntley.com/loop/,2025.Autonomous agentlooppatternforiterativetaskcompletion

    GeoffreyHuntley.Everythingisaralphloop. https://ghuntley.com/loop/,2025.Autonomous agentlooppatternforiterativetaskcompletion

    2025

  4. [12]

    arXiv:2507.00014

    ThomasJoshi,ShayanChowdhury,andFarukUysal.SWE-Bench-CL:Continuallearningforcoding agents,2025. arXiv:2507.00014

    2025 arXiv

  5. [13]

    Joshi, Hanna Mober, et al

    Omar Khattab, Arnav Singhvi, Paridhi Maheshwari, Zhiyuan Zhang, Keshav Santhanam, Sri Vard- hamanan, Saiful Haq, Ashutosh Sharma, Thomas T. Joshi, Hanna Mober, et al. DSPy: Compil- ing declarative language model calls into state-of-the-art pipelines. InInternational Conference...

    2024

  6. [14]

    Feedback descent: Open-ended text optimization via pairwisecomparison

    Yoonho Lee, Joseph Boen, and Chelsea Finn. Feedback descent: Open-ended text optimization via pairwisecomparison. arXiv preprint arXiv:2511.07919,2025

    2025

  7. [15]

    Organizing, orchestrating, and benchmarking agent skills at ecosystem scale.arXiv preprint arXiv:2603.02176,2026

    Hao Li, Chunjiang Mu, Jianhao Chen, Siyue Ren, Zhiyao Cui, Yiqun Zhang, Lei Bai, and Shuyue Hu. Organizing, orchestrating, and benchmarking agent skills at ecosystem scale.arXiv preprint arXiv:2603.02176,2026

    2026

  8. [16]

    SkillNet: Create, evaluate, and connect AI skills.arXiv preprint arXiv:2603.04448, 2026

    Hao Li et al. SkillNet: Create, evaluate, and connect AI skills.arXiv preprint arXiv:2603.04448, 2026. Mind AI Tech Report 28

    2026

  9. [17]

    SkillsBench: Benchmarkinghowwellagentskillswork acrossdiversetasks,2026

    XiangyiLi,WenboChen,YiminLiu,etal. SkillsBench: Benchmarkinghowwellagentskillswork acrossdiversetasks,2026. arXiv:2602.12670

    2026 arXiv

  10. [18]

    SWE- Next: Scalable real-world software engineering tasks for agents.arXiv preprint arXiv:2603.20691, 2026

    JiarongLiang,ZhihengLyu,ZijieLiu,XiangchaoChen,PingNie,KaiZou,andWenhuChen. SWE- Next: Scalable real-world software engineering tasks for agents.arXiv preprint arXiv:2603.20691, 2026

    2026

  11. [19]

    Self-refine: Iterativerefinementwith self-feedback

    Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon,NouhaDziri,ShrimaiPrabhumoye,YimingYang,etal. Self-refine: Iterativerefinementwith self-feedback. InAdvances in Neural Information Processing Systems (NeurIPS),2023

    2023

  12. [20]

    ProcMEM: Learning reusable procedural memory from experience via non- parametricPPOforLLMagents,2026

    Haoran Mi et al. ProcMEM: Learning reusable procedural memory from experience via non- parametricPPOforLLMagents,2026. arXiv:2602.01869

    2026 arXiv

  13. [21]

    Optimizinginstructionsanddemonstrationsformulti-stagelanguagemodelprograms

    KristaOpsahl-Ong,MichaelJ.Ryan,JoshHardy,ShwethaPatel,MichaelS.Bernstein,andChristo- pherPotts. Optimizinginstructionsanddemonstrationsformulti-stagelanguagemodelprograms. In Empirical Methods in Natural Language Processing (EMNLP),2024

    2024

  14. [22]

    Le, etal

    Siru Ouyang, Jun Yan, I-Hung Hsu, Yanfei Chen, Ke Jiang, Zifeng Wang, Rujun Han, Long T. Le, etal. ReasoningBank: Scalingagentself-evolvingwithreasoningmemory. In International Confer- ence on Learning Representations (ICLR),2026. arXiv:2509.25140

    2026 arXiv

  15. [23]

    Gonzalez

    CharlesPacker,SarahWooders,KevinLin,VivianFang,ShishirG.Patil,IonStoica,andJosephE. Gonzalez. MemGPT:TowardsLLMsasoperatingsystems,2023. arXiv:2310.08560

    2023 arXiv

  16. [24]

    Synthesizingagenticdataforwebagentswithprogressivedifficultyenhancementmechanisms

    Shrey Pandit, Xuan-Phi Nguyen, Yifei Ming, Austin Xu, Jiayu Wang, Caiming Xiong, and Shafiq Joty. Synthesizingagenticdataforwebagentswithprogressivedifficultyenhancementmechanisms. arXiv preprint arXiv:2510.13913,2025

    2025

  17. [25]

    Causality: Models, Reasoning, and Inference.CambridgeUniversityPress,2ndedition, 2009

    JudeaPearl. Causality: Models, Reasoning, and Inference.CambridgeUniversityPress,2ndedition, 2009

    2009

  18. [26]

    Measuring and narrowing the compositionality gap in language models

    OfirPress,MuruZhang,SewonMin,LudwigSchmidt,NoahASmith,andMikeLewis. Measuring and narrowing the compositionality gap in language models. InFindings of the Association for Computational Linguistics (EMNLP),2023

    2023

  19. [27]

    ATOMIC: An atlas of machine common- senseforif-thenreasoning

    Maarten Sap, Ronan Le Bras, Emily Allaway, Chandra Bhagavatula, Nicholas Lourie, Hannah Rashkin, Brendan Roof, Noah A Smith, and Yejin Choi. ATOMIC: An atlas of machine common- senseforif-thenreasoning. In AAAI Conference on Artificial Intelligence,2019

    2019

  20. [28]

    Reflexion: Languageagentswithverbalreinforcementlearning

    NoahShinn,FedericoCassano,AshwinGopinath,KarthikNarasimhan,andShunyuYao. Reflexion: Languageagentswithverbalreinforcementlearning. In Advances in Neural Information Processing Systems (NeurIPS),2023

    2023

  21. [29]

    ConceptNet 5.5: An open multilingual graph of generalknowledge

    Robyn Speer, Joshua Chin, and Catherine Havasi. ConceptNet 5.5: An open multilingual graph of generalknowledge. In AAAI Conference on Artificial Intelligence,2017

    2017

  22. [30]

    Dynamic cheatsheet: Adaptive memory for test-time learning.arXiv preprint, 2025

    Mirac Suzgun et al. Dynamic cheatsheet: Adaptive memory for test-time learning.arXiv preprint, 2025

    2025

  23. [31]

    Voyager: An open-ended embodied agent with large language models.arXiv preprint arXiv:2305.16291,2023

    GuanzhiWang,YuqiXie,YunfanJiang,AjayMandlekar,ChaoweiXiao,YukeZhu,LinxiFan,and Anima Anandkumar. Voyager: An open-ended embodied agent with large language models.arXiv preprint arXiv:2305.16291,2023

    2023 arXiv

  24. [32]

    Stateful reflective developer platforms (SRDP): Ai-native platforms for compounding agentproductivity

    Jun Wang. Stateful reflective developer platforms (SRDP): Ai-native platforms for compounding agentproductivity. arXiv preprint arXiv:2512.22716,2025. DefinestheSRDPframeworkonwhich Memento-Skills[41]isbuilt. Mind AI Tech Report 29

    2025

  25. [33]

    arXiv preprint arXiv:2409.07429,2024

    ZoraZhiruoWang,JiayuanMao,DanielFried,andGrahamNeubig.Agentworkflowmemory. arXiv preprint arXiv:2409.07429,2024

    2024 arXiv

  26. [34]

    Optimas: Optimizing compound AI systems with globally aligned local rewards

    Yuxin Wu et al. Optimas: Optimizing compound AI systems with globally aligned local rewards. arXiv preprint,2025

    2025

  27. [35]

    SWE-Exp: Experience-driven software issue resolution, 2025

    Xingyao Xiang et al. SWE-Exp: Experience-driven software issue resolution, 2025. arXiv:2507.23361

    2025

  28. [36]

    UTBoost: Rigorousevaluationofcodingagents onSWE-Bench

    BoxiYu,YuxuanZhu,PinjiaHe,andDanielKang. UTBoost: Rigorousevaluationofcodingagents onSWE-Bench. In Association for Computational Linguistics (ACL),2025. arXiv:2506.09289

    2025 arXiv

  29. [37]

    TextGrad: Automaticdifferentiationviatext

    Mert Yuksekgonul, Federico Bianchi, Joseph Boen, Sheng Liu, Zhi Huang, Carlos Guestrin, and JamesZou. TextGrad: Automaticdifferentiationviatext. Nature,2025

    2025

  30. [38]

    AFlow: Automatingagenticworkflowgeneration

    JiayiZhangetal. AFlow: Automatingagenticworkflowgeneration. In International Conference on Learning Representations (ICLR),2025. OralPresentation

    2025

  31. [39]

    Agentic context engineering: Evolving contexts for self- improvinglanguagemodels.In International Conference on Learning Representations (ICLR),2026

    Qizheng Zhang, Changran Hu, et al. Agentic context engineering: Evolving contexts for self- improvinglanguagemodels.In International Conference on Learning Representations (ICLR),2026. arXiv:2510.04618

    2026 arXiv

  32. [40]

    BIRCH: An efficient data clustering method forverylargedatabases

    Tian Zhang, Raghu Ramakrishnan, and Miron Livny. BIRCH: An efficient data clustering method forverylargedatabases. In ACM SIGMOD International Conference on Management of Data,1996

    1996

  33. [41]

    arXiv preprint arXiv:2603.18743,2026

    HuichiZhou,SiyuanGuo,AnjieLiu,ZhongweiYu,ZiqinGong,BowenZhao,etal.Memento-skills: Letagentsdesignagents. arXiv preprint arXiv:2603.18743,2026

    2026

  34. [42]

    Syn- thetic sandbox for training machine learning engineering agents.arXiv preprint arXiv:2604.04872, 2026

    YuhangZhou,LizhuZhang,YifanWu,JiayiLiu,XiangjunFan,ZhuokaiZhao,andHongYan. Syn- thetic sandbox for training machine learning engineering agents.arXiv preprint arXiv:2604.04872, 2026

    2026 arXiv

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.