Pith. sign in

REVIEW 3 major objections 5 minor 62 references

Language-model agents can improve how they improve by co-evolving task skills and a five-part rewrite procedure under the same frozen model and pipeline.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-11 07:23 UTC pith:N2TK4QJX

load-bearing objection Solid systems paper on recursive skill evolution; the main Single-Level baseline confounds meta-updates with adaptive K and cross-branch, so the headline +6–8 pts cannot yet be pinned on evolving m alone. the 3 major comments →

arxiv 2607.05297 v1 pith:N2TK4QJX submitted 2026-07-06 cs.AI

MetaSkill-Evolve: Recursive Self-Improvement of LLM Agents via Two-Timescale Meta-Skill Evolution

Zefeng Wang , Minxi Yan , Jinhe Bi , Sikuan Yan , Volker Tresp , Yunpu Ma This is my paper
classification cs.AI
keywords recursive self-improvementLLM agentsskill evolutionmeta-skilltwo-timescale evolutionagentic pipelinesfrontier selectionself-improving agents
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Self-improving agents already rewrite task skills from failure traces, but the rewrite procedure itself is usually fixed once and for all. This paper argues that procedure can be treated as a first-class, branch-local object: a meta-skill with five Markdown components that parameterise how failures are diagnosed, experience is shared, search budget is allocated, edits are proposed, and edits are applied. Task skills update on a fast loop every iteration; the meta-skill updates on a slower loop every few iterations by feeding a summary of recent productivity back into the same five-agent pipeline. With one frozen backbone shared by all agents, the two-timescale system beats no-skill, static-skill, and single-level evolution baselines on three agent benchmarks, with the largest lifts on the harder question-answering tasks. A reader who cares about agent reliability would care because the work separates what the agent does from how it improves, and shows both can be searched with the same machinery and no extra training.

Core claim

The paper claims that agent skill improvement can be made recursive in a bounded, practical form: every search branch carries a task skill s and a meta-skill m that parameterises Analyzer, Retriever, Allocator, Proposer, and Evolver; the same pipeline that rewrites s on the fast loop rewrites m on the slow loop from meta-productivity of recent children, and this co-evolution raises held-out accuracy over single-level evolution that freezes the meta-process.

What carries the argument

Two-timescale meta-skill evolution: branch state b=(s,m,h) with m=(ψ,σ,α,π,ε) as Markdown skill files; frontier selection by utility, meta-productivity, and novelty; a slow loop that builds a synthetic meta-failure trace every H iterations and rewrites all five meta-skill files with the same five-agent pipeline.

Load-bearing premise

That turning the last few children’s tags, diagnoses, outcomes, and average gain into one synthetic failure example gives the same pipeline a stable, informative signal for rewriting the improvement procedure, rather than mostly noise or extra search under another name.

What would settle it

Match total child evaluations and cross-branch retrieval of the full system while permanently freezing the meta-skill files; if held-out accuracy on OfficeQA and SealQA then equals MetaSkill-Evolve, the claim that recursive meta-skill updates are the operative gain collapses.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Fixing the improvement procedure leaves accuracy on the table once a diagnosis or edit style plateaus, even if task skills keep changing.
  • Branch-local rewrite policies can diverge and still transfer useful patterns through meta-level retrieval without a shared global policy.
  • Adaptive child budgets and cross-branch inspiration each contribute measurable gains on at least some domains.
  • On near-ceiling tasks, meta-skill updates can still supply residual improvement after task-skill evolution saturates.
  • Gains can be attributed to evolved skills rather than added model capacity when all pipeline agents share one frozen backbone.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • An already-evolved meta-skill might warm-start a new but related domain better than a hand-written task skill alone.
  • Because the five agent roles and their wiring stay fixed, a natural next recursion would be to let the pipeline structure itself evolve—something the paper deliberately bounds to one level.
  • The synthetic meta-failure construction could be reused in other agent loops that currently hard-code reflection, memory update, or search-allocation rules.
  • Strong sensitivity of one benchmark to a stale meta-skill suggests an adaptive horizon may matter more than any fixed schedule when the task skill drifts quickly.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes MetaSkill-Evolve, a two-timescale framework in which each search branch carries a task skill s and a branch-local meta-skill m=(ψ,σ,α,π,ε) that parameterises a five-agent improvement pipeline (Analyzer, Retriever, Allocator, Proposer, Evolver). Task skills are rewritten on a fast loop; every H iterations the same pipeline is applied to a synthetic meta-failure trace built from recent child gains, so that m itself is evolved. Frontier selection scores parents by utility U(s), estimated meta-productivity P(m|s), and novelty. With a single frozen Gemma-4 31B backbone, the system is evaluated on OfficeQA, SealQA, and ALFWorld against No-Skill, Static-Skill, and Single-Level Evolution baselines, reporting held-out gains of +23.54 / +16.09 / +1.92 over No-Skill and +6.38 / +8.05 / +1.92 over Single-Level, with component ablations and an H-horizon sweep.

Significance. If the recursive claim holds, the work supplies a concrete, bounded form of recursive self-improvement for skill-based LLM agents: the improvement operator is itself a set of Markdown skill files refined by the same pipeline, without extra models or objectives. The separation of fast task utility from slow meta-productivity, the typed five-agent pipeline with Retriever and Allocator stages, and the meta-aware frontier score are useful design contributions for agent self-evolution. Strengths include a never-seen held-out test partition, monotonic No-Skill → Static → Single-Level → Ours progression on the two QA benchmarks, component ablations (Table 3), and an H sweep. The main significance is therefore empirical and architectural rather than theoretical; it is of clear interest to the agent and self-improvement communities provided the attribution of gains to evolving m is tightened.

major comments (3)
  1. §4.1 and Table 1: Single-Level Evolution is defined as the fast loop with the slow loop frozen and with (Kmax=1, no cross-branch sharing, no meta-skill updates). The headline +6.38 / +8.05 / +1.92 over Single-Level therefore confounds (i) evolution of m via Alg. 2, (ii) adaptive child budgets via α, and (iii) cross-branch retrieval via σ. Table 3 states that “No meta-updates” exactly reproduces the Single-Level row, confirming the multi-factor definition. On ALFWorld the paper itself reports that removing only cross-branch retrieval returns accuracy to 92.31, so the entire +1.92 “meta” gain is cross-branch sharing and does not require evolving m. A pure fixed-m0 control that keeps initial α and σ (and Kmax>1) while only freezing Alg. 2 is needed before the central claim can attribute the QA gains to recursive meta-skill improvement rather than to extra search machinery under a fixed init
  2. §4.2–4.3 and Table 1: All reported accuracies are single-run point estimates with no error bars, multi-seed statistics, or variance over train/val splits. The default budget is only 5 fast iterations (two meta-updates at H=2). Given that the free parameters include η, H, Kmax, and retrieval limits, and that OfficeQA is highly sensitive to H (Fig. 5: 9.1-point drop from H=2 to H=8), the magnitude of the claimed gains cannot be assessed for stability. At least multi-seed means and standard errors on the main table, or an explicit statement that results are single-run, are required for the load-bearing comparisons.
  3. §3.5 / Alg. 2: The slow loop is driven by a synthetic meta-failure trace fm assembled from tags, diagnoses, outcomes, and empirical P̂ over the last H children, then fed to the same Analyzer→…→Evolver pipeline with whole-m rewrites. The paper does not show that this signal is stable or that the resulting m changes improve subsequent task-skill search beyond the extra compute of the slow loop. Qualitative examples of how ψ/σ/α/π/ε actually change, or a comparison of child-gain trajectories under fixed vs evolved m with matched search budget, would make the recursive mechanism falsifiable rather than assumed.
minor comments (5)
  1. Fig. 1 and abstract: “Porposer” is a typo for Proposer.
  2. §3.3 Eq. (4): The novelty term Nv=1/(1+times_selected) is clear, but the paper never reports the realised distribution of selection counts or whether η3=0.25 actually prevents monopolies under the default KF=3.
  3. App. D / Table 2 vs Table 1: The H=2 (6-iteration) operating point in the horizon sweep is not the same as the 5-iteration default behind Table 1; the text notes this but the figure caption could state it more prominently to avoid misreading absolute numbers.
  4. Related Work: Several concurrent skill-evolution and recursive self-improvement systems are cited; a short table contrasting which of them evolve the improvement operator itself (vs only the task skill or a global prompt) would sharpen the novelty claim.
  5. Limitations: The fixed five-agent wiring and fixed H are acknowledged; adding that Single-Level currently bundles multiple ablations would help readers interpret Table 1.

Circularity Check

0 steps flagged

Empirical systems paper with held-out benchmarks; no derivation that reduces by construction to its inputs.

full rationale

MetaSkill-Evolve is an empirical agent-systems paper, not a closed-form derivation. Task utility U(s) is validation accuracy and meta-productivity P(m|s) is mean child utility gain; both are estimated from loop-visible batches, while reported numbers are accuracy on a never-seen held-out test partition (Table 1, §4.1, App. F). Recursive application of the same five-agent pipeline to m is an intentional design choice (bounded one-level recursion), not a tautology that forces the headline gains. Gains over No-Skill / Static / Single-Level are measured against external benchmarks (OfficeQA, SealQA, ALFWorld) with a frozen shared backbone, so the central claim is falsifiable and not equivalent to its inputs by definition. Related-work self-citations exist but are not load-bearing uniqueness theorems or ansatz smuggling that force the result. Experimental confounding in the Single-Level baseline (jointly freezing meta-updates, Kmax, and cross-branch) is a design/attribution issue, not circularity of a derivation chain. No self-definitional equations, fitted-as-prediction steps, or uniqueness-from-authors reductions apply.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 3 invented entities

The central empirical claim rests on design choices and hyperparameters of the evolutionary loop, plus the modeling assumption that Markdown meta-skill files and a synthetic meta-failure trace are adequate objects for recursive improvement. No physical constants; free parameters are search/scoring knobs. Invented entities are the typed meta-skill and meta-productivity used as first-class search objects.

free parameters (5)
  • frontier weights η1, η2, η3 = 1.0, 0.5, 0.25
    Default η1=1.0, η2=0.5, η3=0.25 directly control parent selection among utility, meta-productivity, and novelty; claim depends on this scoring rule.
  • meta-update horizon H = 2
    Default H=2 sets how often m is rewritten; sweep shows OfficeQA is highly sensitive (large drop as H grows).
  • child budget bounds Kmax and initial K = Kmax=3, initial K=2
    K∈[1,Kmax] with Kmax=3 and initial K=2 (frozen under −α) controls search width per parent.
  • iteration budget and early-stopping = 5 iterations
    Default 5 fast iterations with early stop after 5 iterations without frontier improvement; main numbers are from this short budget.
  • cross-branch retrieval probability and limits = pcross=0.2, Lsame=3, Lcross=2
    pcross=0.2, Lsame=3, Lcross=2 (3× over-fetch) affect inspiration quality and the ALFWorld meta-gain attribution.
axioms (5)
  • domain assumption Task and meta skills are Markdown LLM-agent programs consumable by the same five ToolCallingAgent roles sharing one frozen backbone.
    Stated in §3.2 and experimental setup; enables recursion without a new model.
  • domain assumption Utility U(s) estimated as accuracy/success on a validation batch is a sufficient objective for skill selection and reporting generalization on a held-out test partition.
    §3.1 and §4.1 evaluation protocol.
  • ad hoc to paper Empirical mean child utility gain ˆP is a valid estimate of meta-productivity P(m|s) for frontier scoring and slow-loop drive.
    Eq. 3 and §3.5; defines the slow objective.
  • ad hoc to paper A synthetic meta-failure trace fm built from H descendants can be diagnosed and edited by the same pipeline as a task failure.
    Algorithm 2; load-bearing for recursive meta-updates.
  • ad hoc to paper Archive only admits children with ΔU>0 as parents while still allowing non-improving nodes as inspirations.
    §3.3 evolution graph rules shape search dynamics.
invented entities (3)
  • branch-local meta-skill m=(ψ,σ,α,π,ε) no independent evidence
    purpose: Parameterize Analyzer, Retriever, Allocator, Proposer, Evolver so the improvement procedure can co-evolve with task skills.
    Core postulated object of the framework; independent_evidence is only the empirical ablations in this paper.
  • meta-productivity P(m|s) no independent evidence
    purpose: Slow-timescale objective and frontier term measuring expected per-child utility gain under policy m.
    Eq. 3; used both to drive meta-updates and to select parents.
  • synthetic meta-failure trace fm no independent evidence
    purpose: Reshape recent improvement history into a failure-like example so the same Analyzer prompt serves both timescales.
    §3.5 / Alg. 2 construction unique to this system.

pith-pipeline@v1.1.0-grok45 · 22484 in / 3893 out tokens · 33154 ms · 2026-07-11T07:23:05.397087+00:00 · methodology

0 comments
read the original abstract

Recent LLM agents tackle increasingly long-horizon, open-ended tasks, and external skills, reusable procedural knowledge supplied to the agent, further extend this capability. However, a fixed, hand-authored skill is rarely optimal, and cannot adapt to the diversity of tasks an agent encounters. Self-improving agents address this by rewriting their own skill files from execution traces, yielding meaningful gains on challenging benchmarks. Yet such self-evolution remains non-recursive: it improves only the task skill (what the agent does) while the improvement procedure (how it improves) is authored once and held fixed. We introduce MetaSkill-Evolve, a two-timescale framework that makes agentic skill improvement recursive: every branch carries both a task skill $s$ and a branch-local meta-skill $m=(\psi,\sigma,\alpha,\pi,\varepsilon)$ whose five components parameterise the Analyzer, Retriever, Allocator, Proposer, and Evolver agents of the improvement pipeline. Task skills evolve on a fast loop while the meta-skill evolves on a slower one under the same pipeline applied to itself, with no additional model or objective. With all five pipeline agents sharing a single frozen backbone, MetaSkill-Evolve outperforms no-skill, static-skill, and single-level evolution baselines on three agentic benchmarks (OfficeQA, SealQA, ALFWorld), improving held-out test accuracy over the raw backbone by +23.54, +16.09, and +1.92 points respectively.

Figures

Figures reproduced from arXiv: 2607.05297 by Jinhe Bi, Minxi Yan, Sikuan Yan, Volker Tresp, Yunpu Ma, Zefeng Wang.

Figure 1
Figure 1. Figure 1: Four regimes of agent skill improvement. No-Skill: no reusable skill memory. Static Skill: a hand￾authored s0 held fixed (padlocked). Single Level Evolve: the task skill evolves s0 → s1 → s2, but the driving meta-process stays padlocked. MetaSkill-Evolve (ours): a branch-level meta-skill m = (ψ, σ, α, π, ε) co-evolves on a slower outer ring via the same five-agent pipeline that rewrites s, with no extra mo… view at source ↗
Figure 2
Figure 2. Figure 2: System overview. The branch state b = (s, m, h) (left) feeds the five-agent pipeline (centre), whose output is appended to the SQLite node graph (right). Frontier selection (Eq. 4) draws the next parent from the graph. and Nv = 1/(1 + times_selectedv). Each term targets a distinct failure mode of greedy search: • Uv – exploitation: prevents chasing volatile gain estimates on weak parents. • Pˆ v – trajecto… view at source ↗
Figure 3
Figure 3. Figure 3: End-to-end held-out test accuracy on the three benchmarks. All conditions share the Gemma-4 31B backbone; only the skill-evolution strategy differs. Red annotations mark MetaSkill-Evolve’s gain over No￾Skill; the No-Skill→Ours ordering improves on every benchmark, with the largest margins on the two QA tasks where the backbone has the most headroom. slightly regresses (−1.93), and single-level evo￾lution o… view at source ↗
Figure 5
Figure 5. Figure 5: Meta-update horizon sweep. Held-out test accuracy as the horizon H (fast iterations between con￾secutive meta-skill evolutions) widens from 2 to 8, with the meta-update count held fixed at three, so total itera￾tions = 3H (shown in parentheses). The broken y-axis separates near-ceiling ALFWorld (top) from the two QA accuracies (bottom). The default H=2 (shaded band) is best on every benchmark; OfficeQA is … view at source ↗
Figure 4
Figure 4. Figure 4: Component ablations on the two QA bench￾marks. Blue polygon: accuracy with one meta-skill component removed; dashed rings: full MetaSkill￾Evolve and Static-Skill references. Dominant compo￾nent differs by domain: α on OfficeQA, π on SealQA. ALFWorld in [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗

Review history (2 revisions) →

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

62 extracted references · 1 canonical work pages

  1. [1]

    2025 , eprint=

    SkillWeaver: Web Agents can Self-Improve by Discovering and Honing Skills , author=. 2025 , eprint=

    2025

  2. [2]

    2026 , eprint=

    SkillRL: Evolving Agents via Recursive Skill-Augmented Reinforcement Learning , author=. 2026 , eprint=

    2026

  3. [3]

    2026 , eprint=

    MemSkill: Learning and Evolving Memory Skills for Self-Evolving Agents , author=. 2026 , eprint=

    2026

  4. [4]

    2025 , eprint=

    ShinkaEvolve: Towards Open-Ended And Sample-Efficient Program Evolution , author=. 2025 , eprint=

    2025

  5. [5]

    2025 , eprint=

    AlphaEvolve: A coding agent for scientific and algorithmic discovery , author=. 2025 , eprint=

    2025

  6. [6]

    2026 , eprint=

    ReasoningBank: Scaling Agent Self-Evolving with Reasoning Memory , author=. 2026 , eprint=

    2026

  7. [7]

    2025 , eprint=

    SE-Agent: Self-Evolution Trajectory Optimization in Multi-Step Reasoning with LLM-Based Agents , author=. 2025 , eprint=

    2025

  8. [8]

    2026 , eprint=

    CoEvoSkills: Self-Evolving Agent Skills via Co-Evolutionary Verification , author=. 2026 , eprint=

    2026

  9. [9]

    2025 , eprint=

    ThetaEvolve: Test-time Learning on Open Problems , author=. 2025 , eprint=

    2025

  10. [10]

    2026 , eprint=

    Meta Context Engineering via Agentic Skill Evolution , author=. 2026 , eprint=

    2026

  11. [11]

    2026 , eprint=

    Trace2Skill: Distill Trajectory-Local Lessons into Transferable Agent Skills , author=. 2026 , eprint=

    2026

  12. [12]

    2026 , eprint=

    Skill1: Unified Evolution of Skill-Augmented Agents via Reinforcement Learning , author=. 2026 , eprint=

    2026

  13. [13]

    2026 , howpublished=

    google/gemma-4-31B , author=. 2026 , howpublished=

    2026

  14. [14]

    Gradient Descent

    Automatic Prompt Optimization with "Gradient Descent" and Beam Search , author=. 2023 , eprint=

    2023

  15. [15]

    2023 , eprint=

    Large Language Models Are Human-Level Prompt Engineers , author=. 2023 , eprint=

    2023

  16. [16]

    2023 , eprint=

    Reflexion: Language Agents with Verbal Reinforcement Learning , author=. 2023 , eprint=

    2023

  17. [17]

    2023 , eprint=

    Voyager: An Open-Ended Embodied Agent with Large Language Models , author=. 2023 , eprint=

    2023

  18. [18]

    2024 , eprint=

    ExpeL: LLM Agents Are Experiential Learners , author=. 2024 , eprint=

    2024

  19. [19]

    2015 , eprint=

    Illuminating search spaces by mapping elites , author=. 2015 , eprint=

    2015

  20. [20]

    Differentiation

    TextGrad: Automatic "Differentiation" via Text , author=. 2024 , eprint=

    2024

  21. [21]

    2026 , eprint=

    SealQA: Raising the Bar for Reasoning in Search-Augmented Language Models , author=. 2026 , eprint=

    2026

  22. [22]

    2026 , eprint=

    EvoSkill: Automated Skill Discovery for Multi-Agent Systems , author=. 2026 , eprint=

    2026

  23. [23]

    2026 , eprint=

    OfficeQA Pro: An Enterprise Benchmark for End-to-End Grounded Reasoning , author=. 2026 , eprint=

    2026

  24. [24]

    2025 , eprint=

    EvoPrompt: Connecting LLMs with Evolutionary Algorithms Yields Powerful Prompt Optimizers , author=. 2025 , eprint=

    2025

  25. [25]

    2021 , eprint=

    ALFWorld: Aligning Text and Embodied Environments for Interactive Learning , author=. 2021 , eprint=

    2021

  26. [26]

    2026 , eprint=

    Skills-Coach: A Self-Evolving Skill Optimizer via Training-Free GRPO , author=. 2026 , eprint=

    2026

  27. [27]

    2026 , eprint=

    From Context to Skills: Can Language Models Learn from Context Skillfully? , author=. 2026 , eprint=

    2026

  28. [28]

    2026 , eprint=

    GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning , author=. 2026 , eprint=

    2026

  29. [29]

    2023 , eprint=

    Self-Refine: Iterative Refinement with Self-Feedback , author=. 2023 , eprint=

    2023

  30. [30]

    2025 , eprint=

    Feedback Descent: Open-Ended Text Optimization via Pairwise Comparison , author=. 2025 , eprint=

    2025

  31. [31]

    2023 , eprint=

    ReAct: Synergizing Reasoning and Acting in Language Models , author=. 2023 , eprint=

    2023

  32. [32]

    2024 , eprint=

    Large Language Models as Optimizers , author=. 2024 , eprint=

    2024

  33. [33]

    2023 , eprint=

    DSPy: Compiling Declarative Language Model Calls into Self-Improving Pipelines , author=. 2023 , eprint=

    2023

  34. [34]

    2026 , eprint=

    SkillsBench: Benchmarking How Well Agent Skills Work Across Diverse Tasks , author=. 2026 , eprint=

    2026

  35. [35]

    2026 , eprint=

    SoK: Agentic Skills -- Beyond Tool Use in LLM Agents , author=. 2026 , eprint=

    2026

  36. [36]

    2026 , eprint=

    AutoSkill: Experience-Driven Lifelong Learning via Skill Self-Evolution , author=. 2026 , eprint=

    2026

  37. [37]

    2026 , eprint=

    EvoTest: Evolutionary Test-Time Learning for Self-Improving Agentic Systems , author=. 2026 , eprint=

    2026

  38. [38]

    2026 , eprint=

    SKILLFOUNDRY: Building Self-Evolving Agent Skill Libraries from Heterogeneous Scientific Resources , author=. 2026 , eprint=

    2026

  39. [39]

    ArXiv , year=

    SkillForge: Forging Domain-Specific, Self-Evolving Agent Skills in Cloud Technical Support , author=. ArXiv , year=

  40. [40]

    2026 , eprint=

    SkillX: Automatically Constructing Skill Knowledge Bases for Agents , author=. 2026 , eprint=

    2026

  41. [41]

    2026 , eprint=

    SkillClaw: Let Skills Evolve Collectively with Agentic Evolver , author=. 2026 , eprint=

    2026

  42. [42]

    2026 , eprint=

    Memp: Exploring Agent Procedural Memory , author=. 2026 , eprint=

    2026

  43. [43]

    Speculations Concerning the First Ultraintelligent Machine , author=. Adv. Comput. , year=

  44. [44]

    2006 , eprint=

    Goedel Machines: Self-Referential Universal Problem Solvers Making Provably Optimal Self-Improvements , author=. 2006 , eprint=

    2006

  45. [45]

    2024 , eprint=

    Self-Taught Optimizer (STOP): Recursively Self-Improving Code Generation , author=. 2024 , eprint=

    2024

  46. [46]

    2023 , eprint=

    Promptbreeder: Self-Referential Self-Improvement Via Prompt Evolution , author=. 2023 , eprint=

    2023

  47. [47]

    2025 , eprint=

    Automated Design of Agentic Systems , author=. 2025 , eprint=

    2025

  48. [48]

    2026 , eprint=

    Darwin Godel Machine: Open-Ended Evolution of Self-Improving Agents , author=. 2026 , eprint=

    2026

  49. [49]

    ArXiv , year=

    PRISM: Self-Pruning Intrinsic Selection Method for Training-Free Multimodal Data Selection , author=. ArXiv , year=

  50. [50]

    LL a VA Steering: Visual Instruction Tuning with 500x Fewer Parameters through Modality Linear Representation-Steering

    Bi, Jinhe and Wang, Yujun and Chen, Haokun and Xiao, Xun and Hecker, Artur and Tresp, Volker and Ma, Yunpu. LL a VA Steering: Visual Instruction Tuning with 500x Fewer Parameters through Modality Linear Representation-Steering. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025

    2025

  51. [51]

    Jinhe Bi and Aniri and Minglai Yang and Xingcheng Zhou and Wenke Huang and Sikuan Yan and Yujun Wang and Zixuan Cao and Michael Färber and Xun Xiao and Volker Tresp and Yunpu Ma , booktitle=. Echo. 2026 , url=

    2026

  52. [52]

    Forty-third International Conference on Machine Learning , year=

    The Geometry of Reasoning: Self-Evaluation via Layerwise Trajectory Evolution , author =. Forty-third International Conference on Machine Learning , year=

  53. [53]

    ArXiv , year=

    CoT-Kinetics: A Theoretical Modeling Assessing LRM Reasoning Process , author=. ArXiv , year=

  54. [54]

    2025 , eprint=

    Loong: Synthesize Long Chain-of-Thoughts at Scale through Verifiers , author=. 2025 , eprint=

    2025

  55. [55]

    2026 , eprint=

    AlignSAE: Concept-Aligned Sparse Autoencoders , author=. 2026 , eprint=

    2026

  56. [56]

    2025 , eprint=

    Beyond Magic Words: Sharpness-Aware Prompt Evolving for Robust Large Language Models with TARE , author=. 2025 , eprint=

    2025

  57. [57]

    2025 , eprint=

    Reinforcement Mid-Training , author=. 2025 , eprint=

    2025

  58. [58]

    2026 , eprint=

    Graph is a Substrate Across Data Modalities , author=. 2026 , eprint=

    2026

  59. [59]

    2025 , url=

    Guancheng Wan and Xiaoran Shang and Yuxin Wu and Guibin Zhang and Jinhe Bi and Liangtao Zheng and Xin Lin and Yue Liu and Yanbiao Ma and Wenke Huang and Bo Du , booktitle=. 2025 , url=

    2025

  60. [60]

    Self-Evolving Multi-Agent Systems via Textual Backpropagation

    Ma, Xiaowen and Ma, Yunpu and Lin, Chenyang and Yan, Sikuan and Bi, Jinhe and Cao, Zixuan and Tian, Yijun and Tresp, Volker and Schuetze, Hinrich. Self-Evolving Multi-Agent Systems via Textual Backpropagation. Findings of the A ssociation for C omputational L inguistics: ACL 2026. 2026. doi:10.18653/v1/2026.findings-acl.483

    work page doi:10.18653/v1/2026.findings-acl.483 2026

  61. [61]

    2026 , eprint=

    Beyond NL2Code: A Structured Survey of Multimodal Code Intelligence , author=. 2026 , eprint=

    2026

  62. [62]

    DocBench: A Comprehensive Benchmark for Expert-Level and Difficult Document Parsing , author=

    Dr. DocBench: A Comprehensive Benchmark for Expert-Level and Difficult Document Parsing , author=. 2026 , eprint=

    2026