Kintsugi: Learning Policies by Repairing Executable Knowledge Bases

• URL: http://arxiv.org/abs/2605.09487v1
• Date: Sun, 10 May 2026 11:51:18 GMT
• Title: Kintsugi: Learning Policies by Repairing Executable Knowledge Bases
• Abstract summary: Kintsugi represents task-level policy knowledge as composable typed entries.<n>Between rollouts, a tool-constrained editing loop diagnoses trajectory failures, localizes them to editable KB layers, and proposes candidate edits.<n>A deterministic verification gate admits an edit only when the candidate type-checks, the resulting KB executes, and focused validation success or trajectory-health metrics improve without violating protected-regression checks.
• Score: 31.289004484222847
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Modern embodied agents achieve impressive performance, but their task knowledge is often stored in neural weights, latent state, or prompt-bound memory, making individual policy knowledge difficult to inspect, validate, recombine, and reuse. We introduce \textbf{Kintsugi}, a white-box policy-learning framework that treats embodied policy improvement as verifier-gated construction of a typed executable Knowledge Base (KB). Kintsugi represents task-level policy knowledge as composable typed entries -- predicates, operators, policy schemas, monitors, recovery rules, experience records, and goals -- and improves this artifact through localized typed edits induced from rollout evidence, rather than relying on test-time language-model reasoning. Between rollouts, a tool-constrained agentic editing loop diagnoses trajectory failures, localizes them to editable KB layers, and proposes candidate edits. A deterministic verification gate admits an edit only when the candidate type-checks, the resulting KB executes, and focused validation success or trajectory-health metrics improve without violating protected-regression checks. At inference, the accepted KB is executed by a deterministic symbolic executor with zero LLM calls. Across long-horizon text-agent benchmarks and representative object-centric manipulation settings, Kintsugi achieves strong endpoint performance while preserving inspectability, local editability, and verifier-gated deployment. These results suggest that embodied policy improvement can be organized around executable task knowledge.

Related papers

• SkillForge: Compositional Skill Synthesis with Verification-in-the-Loop for Generating Formally Verified Dafny Programs [10.916179875278326]
  We introduce SKILLFORGE, a framework that decomposes formal code synthesis into a library of atomic, reusable skills.<n>A verification-driven harness orchestrates these skills.<n>On a curated benchmark of natural language to Dafny specification pairs, SKILLFORGE substantially outperforms both state-of-the-art agentic approaches.
  arXiv  Detail & Related papers  (2026-08-30T15:15:00Z)
• Interpretable reinforcement learning with decision-tree pruning [0.0]
  Reinforcement learning policies are difficult to inspect, but interpreting them is a prerequisite for trustworthiness.<n>We present a pruning process that simplifies such rule-based policies while preserving task performance.<n>We investigate this approach on classic control and MuJoCo benchmarks, where pruning traces reveal consistent interpretability improvements.
  arXiv  Detail & Related papers  (2026-08-07T12:14:33Z)
• Beyond Perplexity: A Behavioral Evaluation Framework for Deployment-Memory Claims in LLM Test-Time Training [30.22002300048916]
  Large language model test-time training (TTT) is often evaluated through local proxy metrics.<n>We introduce a behavioral evaluation framework that calibrates TTT memory claims to the evidence that supports them.
  arXiv  Detail & Related papers  (2026-07-01T03:07:17Z)
• SkillGenBench: Benchmarking Skill Generation Pipelines for LLM Agents [15.669318649782285]
  We introduce SkillGenBench, a benchmark for evaluating skill generation pipelines under a unified protocol.<n>A generator receives raw corpora and produces standardized skill artifacts, which are then executed under fixed harnesses.<n> Experiments across a range of skill-generation methods and backbones show substantial performance variation.
  arXiv  Detail & Related papers  (2026-05-18T17:28:36Z)
• OLIVIA: Online Learning via Inference-time Action Adaptation for Decision Making in LLM ReAct Agents [74.20327254615854]
  Large language model agents interleave reasoning, action selection, and observation to solve sequential decision-making tasks.<n>Existing inference-time adaptation methods for LLM agents mainly rely on prompting or retrieval.<n>We propose OLIVIA, an inference-time action adaptation framework for ReAct-style agents.
  arXiv  Detail & Related papers  (2026-05-11T19:28:20Z)
• PYTHALAB-MERA: Validation-Grounded Memory, Retrieval, and Acceptance Control for Frozen-LLM Coding Agents [0.0]
  We introduce PYTHALAB-MERA, a lightweight external controller for local validation-conditioned code generation.<n>We evaluate the implementation as a local CLI artifact on reinforcement-learning coding tasks with strict validation gates.
  arXiv  Detail & Related papers  (2026-05-08T20:39:32Z)
• AJ-Bench: Benchmarking Agent-as-a-Judge for Environment-Aware Evaluation [71.49152943451328]
  We introduce a benchmark AJ-Bench to evaluate Agent-as-a-Judge across three domains-search, data systems, and graphical user interfaces-comprising 155 tasks and 516 annotated trajectories.<n>Experiments demonstrate consistent performance gains over LLM-as-a-Judge baselines, while also revealing substantial open challenges in agent-based verification.
  arXiv  Detail & Related papers  (2026-04-20T13:23:38Z)
• PolicyBank: Evolving Policy Understanding for LLM Agents [51.86716874651299]
  PolicyBank maintains structured, tool-level policy insights and iteratively refines them.<n>PolicyBank closes up to 82% of the gap toward a human oracle.
  arXiv  Detail & Related papers  (2026-04-16T20:29:30Z)
• RoboPlayground: Democratizing Robotic Evaluation through Structured Physical Domains [19.460905215536254]
  Evaluation of robotic manipulation systems has largely relied on fixed benchmarks authored by experts.<n>We argue that evaluating modern manipulation policies requires reframing evaluation as a language-driven process.<n>We present RoboPlayground, a framework that enables users to author executable manipulation tasks using natural language.
  arXiv  Detail & Related papers  (2026-04-06T22:42:05Z)
• MAC: Multi-Agent Constitution Learning [13.313180901776816]
  Constitutional AI is a method to oversee and control LLMs based on a set of rules written in natural language.<n>Existing LLM-based prompt methods are ineffective at learning constitutions since they require many labeled examples and lack structure in the optimized prompts.<n>We propose Multi-Agent Constitutional Learning (MAC), which optimize over structured prompts represented as sets of rules.
  arXiv  Detail & Related papers  (2026-03-16T22:42:07Z)
• Are We Evaluating the Edit Locality of LLM Model Editing Properly? [68.441768731381]
  We find that existing specificity evaluation protocols are inadequate for this purpose.<n>Existing specificity metrics are weakly correlated with the strength of specificity regularizers.<n>We also find that current metrics lack sufficient sensitivity, rendering them ineffective at distinguishing the specificity performance of different methods.
  arXiv  Detail & Related papers  (2026-01-24T07:07:21Z)
• Deriving Character Logic from Storyline as Codified Decision Trees [67.01182739162142]
  Role-playing (RP) agents rely on behavioral profiles to act consistently across diverse narrative contexts.<n>We propose Codified Decision Trees (CDT), a data-driven framework that induces an executable and interpretable decision structure from large-scale narrative data.
  arXiv  Detail & Related papers  (2026-01-15T05:12:43Z)
• KBQA-R1: Reinforcing Large Language Models for Knowledge Base Question Answering [64.62317305868264]
  We present textbfKBQA-R1, a framework that shifts the paradigm from text imitation to interaction optimization via Reinforcement Learning.<n>Treating KBQA as a multi-turn decision process, our model learns to navigate the knowledge base using a list of actions.<n>Experiments on WebQSP, GrailQA, and GraphQuestions demonstrate that KBQA-R1 achieves state-of-the-art performance.
  arXiv  Detail & Related papers  (2025-12-10T17:45:42Z)
• Rationale-Augmented Retrieval with Constrained LLM Re-Ranking for Task Discovery [4.061135251278187]
  Head Start programs utilizing GoEngage face significant challenges when new or rotating staff attempt to locate appropriate Tasks on the platform homepage.<n>These difficulties arise from domain-specific jargon, system-specific nomenclature, and the inherent limitations of lexical search in handling typos and varied word ordering.<n>We propose a pragmatic hybrid semantic search system that combines lightweight typo-tolerant lexical retrieval, embedding-based vector similarity, and constrained large language model (LLM) re-ranking.
  arXiv  Detail & Related papers  (2025-10-01T01:28:59Z)
• Learning Verifiable Control Policies Using Relaxed Verification [49.81690518952909]
  This work proposes to perform verification throughout training to aim for policies whose properties can be evaluated throughout runtime.<n>The approach is to use differentiable reachability analysis and incorporate new components into the loss function.
  arXiv  Detail & Related papers  (2025-04-23T16:54:35Z)
• STACKFEED: Structured Textual Actor-Critic Knowledge Base Editing with FeedBack [9.207360247989346]
  FEED is a novel Structured Textual Actor-Critic Knowledge base editing with FEEDback approach.<n>FEED refines the knowledge base based on expert feedback using a multi-actor, centralized critic reinforcement learning framework.<n>FEED significantly improves KB quality and performance of the RAG system.
  arXiv  Detail & Related papers  (2024-10-14T14:56:01Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.