CL-Bench is the first expert-validated benchmark for continual learning in frontier LLMs across six real-world domains, showing limited gains and that naive in-context learning outperforms dedicated memory systems.
hub Canonical reference
Agentic Context Engineering: Evolving Contexts for Self-Improving Language Models
Canonical reference. 91% of citing Pith papers cite this work as background.
76
Pith papers citing it
Background
91% of classified citations
abstract
Large language model (LLM) applications such as agents and domain-specific reasoning increasingly rely on context adaptation: modifying inputs with instructions, strategies, or evidence, rather than weight updates. Prior approaches improve usability but often suffer from brevity bias, which drops domain insights for concise summaries, and from context collapse, where iterative rewriting erodes details over time. We introduce ACE (Agentic Context Engineering), a framework that treats contexts as evolving playbooks that accumulate, refine, and organize strategies through a modular process of generation, reflection, and curation. ACE prevents collapse with structured, incremental updates that preserve detailed knowledge and scale with long-context models. Across agent and domain-specific benchmarks, ACE optimizes contexts both offline (e.g., system prompts) and online (e.g., agent memory), consistently outperforming strong baselines: +10.6% on agents and +8.6% on finance, while significantly reducing adaptation latency and rollout cost. Notably, ACE could adapt effectively without labeled supervision and instead by leveraging natural execution feedback. On the AppWorld leaderboard, ACE matches the top-ranked production-level agent on the overall average and surpasses it on the harder test-challenge split, despite using a smaller open-source model. These results show that comprehensive, evolving contexts enable scalable, efficient, and self-improving LLM systems with low overhead.
hub tools
citation-role summary
background 10
method 1
citation-polarity summary
representative citing papers
SimWorld Studio deploys an evolving coding agent to create adaptive 3D environments that co-evolve with embodied learners, delivering 18-point success-rate gains over fixed environments in navigation benchmarks.
LLM agents often fail to abstain at the right time in uncertain multi-turn tasks, and the CONVOLVE context engineering method raises timely abstention rates on WebShop from 26.7 to 57.4 without parameter updates.
SkillHarness models skill learning as a safety-constrained process with skill boundaries and selective reuse, reducing unsafe skills by 57.1% and improving stability in dynamic settings.
CyberEvolver introduces a four-layer self-evolving agent architecture with trace-to-diagnosis and population beam search that raises seed agent success rates by 13.6% on CTF, exploitation, and penetration tasks across four LLMs.
Life-Harness evolves reusable interventions from training trajectories to enhance frozen LLM agents on unseen tasks across seven deterministic environments, yielding 88.5% average relative improvement in 116 of 126 model-environment settings.
EXG is an experience graph framework for self-evolving LLM agents that supports online real-time growth and offline reuse to enhance solution quality and efficiency on code generation and reasoning benchmarks.
Fast-Slow Training uses context optimization as fast weights alongside parameter updates as slow weights to achieve up to 3x better sample efficiency, higher performance, and less catastrophic forgetting than standard RL in continual LLM learning.
RewardHarness self-evolves a tool-and-skill library from 100 preference examples to reach 47.4% accuracy on image-edit evaluation, beating GPT-5, and yields stronger RL-tuned models.
Springdrift provides an auditable persistent runtime for long-lived LLM agents with case-based memory, normative safety gating, and ambient self-perception, shown in a 23-day single-instance deployment where the agent self-diagnosed bugs and maintained cross-channel context.
Meta-Harness discovers improved harness code for LLMs via agentic search over prior execution traces, yielding 7.7-point gains on text classification with 4x fewer tokens and 4.7-point gains on math reasoning across held-out models.
EvoDiagram uses a coordinated multi-agent system and design knowledge evolution to generate editable diagrams via canvas schema, with a new CanvasBench benchmark showing strong performance over baselines.
VIGA introduces a training-free interleaved multimodal reasoning loop that improves vision-as-inverse-graphics accuracy over one-shot baselines on BlenderGym, SlideBench, and new BlenderBench.
Different patterns of environmental change (space vs time) require different LLM update behaviors; no single family of methods—prompts, distillation, RL, or compression—handles all regimes.
The paper proposes Agentic Data Environments that amplify agent capabilities (via information management, retrieval, and elicitation) while bounding failure consequences (via branching and data flow control).
Introduces loop engineering as a distinct practice layer for coding agents, supplies a taxonomy and verification ladder, and analyzes a hand-coded corpus of fifty real loops.
Metis combines text and code memory hierarchically for self-evolving agents, claiming up to 20.6% higher accuracy and 22.8% lower cost than ReAct on the AppWorld benchmark.
MetaPS trains models via simulation rollouts to select from programmatic strategy libraries for market agents, yielding better performance than fixed or direct LLM baselines across model sizes.
LEDGER improves consistency in agentic document editing from 56% to 76% across six models by using a dependency graph for targeted context retrieval while cutting token usage.
RSEA adds a strict held-out keep-better gate to recursive self-evolution of agent artifacts, yielding monotone-safe gains or parity with the base ReAct agent on ALFWorld, GAIA, τ-bench, and WebShop.
HOI-Edit benchmarks human–object interaction image editing with a VLM-based metric, and SCPE steers I2V models via agentic self-correcting prompts to produce competitive HOI edits.
Agent-guided prompt search with diagnostic code execution does not dominate simpler search on benchmarks, but in a production mental-health chatbot it chained eight noisy A/B tests into a claimed +29.4% next-day-retention gain.
An LLM agent self-evolves a set of query-rewriting rules that raise BM25 performance on the LeCaRD-v2 legal retrieval benchmark above human-designed and greedy baselines.
ReSum's contrastive RL branching on self-summarization points improves LLM math reasoning accuracy by about 4% and shortens rollouts by about 18.6% across tested backbones.
citing papers explorer
-
Continual Learning Bench: Evaluating Frontier AI Systems in Real-World Stateful Environments
CL-Bench is the first expert-validated benchmark for continual learning in frontier LLMs across six real-world domains, showing limited gains and that naive in-context learning outperforms dedicated memory systems.
-
SimWorld Studio: Automatic Environment Generation with Evolving Coding Agent for Embodied Agent Learning
SimWorld Studio deploys an evolving coding agent to create adaptive 3D environments that co-evolve with embodied learners, delivering 18-point success-rate gains over fixed environments in navigation benchmarks.
-
Agentic Abstention: Do Agents Know When to Stop Instead of Act?
LLM agents often fail to abstain at the right time in uncertain multi-turn tasks, and the CONVOLVE context engineering method raises timely abstention rates on WebShop from 26.7 to 57.4 without parameter updates.
-
SkillHarness: Harnessing Safe Skills for Computer-Use Agents
SkillHarness models skill learning as a safety-constrained process with skill boundaries and selective reuse, reducing unsafe skills by 57.1% and improving stability in dynamic settings.
-
CyberEvolver: Structured Self-Evolution for Cybersecurity Agents On the Fly
CyberEvolver introduces a four-layer self-evolving agent architecture with trace-to-diagnosis and population beam search that raises seed agent success rates by 13.6% on CTF, exploitation, and penetration tasks across four LLMs.
-
Adapting the Interface, Not the Model: Runtime Harness Adaptation for Deterministic LLM Agents
Life-Harness evolves reusable interventions from training trajectories to enhance frozen LLM agents on unseen tasks across seven deterministic environments, yielding 88.5% average relative improvement in 116 of 126 model-environment settings.
-
EXG: Self-Evolving Agents with Experience Graphs
EXG is an experience graph framework for self-evolving LLM agents that supports online real-time growth and offline reuse to enhance solution quality and efficiency on code generation and reasoning benchmarks.
-
Learning, Fast and Slow: Towards LLMs That Adapt Continually
Fast-Slow Training uses context optimization as fast weights alongside parameter updates as slow weights to achieve up to 3x better sample efficiency, higher performance, and less catastrophic forgetting than standard RL in continual LLM learning.
-
RewardHarness: Self-Evolving Agentic Post-Training
RewardHarness self-evolves a tool-and-skill library from 100 preference examples to reach 47.4% accuracy on image-edit evaluation, beating GPT-5, and yields stronger RL-tuned models.
-
Springdrift: An Auditable Persistent Runtime for LLM Agents with Case-Based Memory, Normative Safety, and Ambient Self-Perception
Springdrift provides an auditable persistent runtime for long-lived LLM agents with case-based memory, normative safety gating, and ambient self-perception, shown in a 23-day single-instance deployment where the agent self-diagnosed bugs and maintained cross-channel context.
-
Meta-Harness: End-to-End Optimization of Model Harnesses
Meta-Harness discovers improved harness code for LLMs via agentic search over prior execution traces, yielding 7.7-point gains on text classification with 4x fewer tokens and 4.7-point gains on math reasoning across held-out models.
-
EvoDiagram: Agentic Editable Diagram Creation via Design Expertise Evolution
EvoDiagram uses a coordinated multi-agent system and design knowledge evolution to generate editable diagrams via canvas schema, with a new CanvasBench benchmark showing strong performance over baselines.
-
Vision-as-Inverse-Graphics Agent via Interleaved Multimodal Reasoning
VIGA introduces a training-free interleaved multimodal reasoning loop that improves vision-as-inverse-graphics accuracy over one-shot baselines on BlenderGym, SlideBench, and new BlenderBench.
-
When Does Continual Learning Require Learning
Different patterns of environmental change (space vs time) require different LLM update behaviors; no single family of methods—prompts, distillation, RL, or compression—handles all regimes.
-
Agentic Data Environments
The paper proposes Agentic Data Environments that amplify agent capabilities (via information management, retrieval, and elicitation) while bounding failure consequences (via branching and data flow control).
-
Stop Hand-Holding Your Coding Agent: Engineering the Loops that Replace Step-by-Step Prompting
Introduces loop engineering as a distinct practice layer for coding agents, supplies a taxonomy and verification ladder, and analyzes a hand-coded corpus of fifty real loops.
-
Metis: Bridging Text and Code Memory for Self-Evolving Agents
Metis combines text and code memory hierarchically for self-evolving agents, claiming up to 20.6% higher accuracy and 22.8% lower cost than ReAct on the AppWorld benchmark.
-
MetaPS: Adaptive Programmatic Strategy Selection for Market Agents
MetaPS trains models via simulation rollouts to select from programmatic strategy libraries for market agents, yielding better performance than fixed or direct LLM baselines across model sizes.
-
LEDGER: Scaling Agentic Document Editing with Dependency-aware Graph Retrieval
LEDGER improves consistency in agentic document editing from 56% to 76% across six models by using a dependency graph for targeted context retrieval while cutting token usage.
-
Recursive Self-Evolving Agents via Held-Out Selection
RSEA adds a strict held-out keep-better gate to recursive self-evolution of agent artifacts, yielding monotone-safe gains or parity with the base ReAct agent on ALFWorld, GAIA, τ-bench, and WebShop.
-
Taming I2V models for Image HOI Editing: A Cognitive Benchmark and Agentic Self-Correcting Framework
HOI-Edit benchmarks human–object interaction image editing with a VLM-based metric, and SCPE steers I2V models via agentic self-correcting prompts to produce competitive HOI edits.
-
SAGE: Stochastic Prompt Optimization via Agent-Guided Exploration
Agent-guided prompt search with diagnostic code execution does not dominate simpler search on benchmarks, but in a production mental-health chatbot it chained eight noisy A/B tests into a claimed +29.4% next-day-retention gain.
-
When Rules Learn: A Self-Evolving Agent for Legal Case Retrieval
An LLM agent self-evolves a set of query-rewriting rules that raise BM25 performance on the LeCaRD-v2 legal retrieval benchmark above human-designed and greedy baselines.
-
ReSum: Synergizing LLM Reasoning and Summarization with Reinforcement Learning
ReSum's contrastive RL branching on self-summarization points improves LLM math reasoning accuracy by about 4% and shortens rollouts by about 18.6% across tested backbones.
-
Toward Generalist Autonomous Research via Hypothesis-Tree Refinement
Arbor combines a coordinator, executors, and a hypothesis tree to enable cumulative autonomous research, outperforming Codex and Claude Code by over 2.5x on six real tasks and reaching 86.36% Any Medal on MLE-Bench Lite.
-
From Player to Master: Enhancing Test-Time Learning of LLM Agents via Reinforcement Learning over Memory
MemoPilot trains memory updates for LLM agents via multi-turn GRPO on RPS and poker, achieving top Elo scores and outperforming baselines including DeepSeek-V3.2.
-
AURA: Intent-Directed Probing for Implicit-Need Surfacing in Situated LLM Agents
AURA improves implicit-need coverage by 0.07 over ReAct baselines on a 100-query benchmark by inserting an intent inference step controlled by a gap score, while cutting probes 82% on factual tasks.
-
RAMPART: Registry-based Agentic Memory with Priority-Aware Runtime Transformation
RAMPART is a registry-based memory system for LLM agents with priority-aware primitives that experimentally demonstrates position-dependent performance cliffs and benefits from block grouping and relevance gating.
-
BenchEvolver: Frontier Task Synthesis via Solution-Centric Evolution
BenchEvolver evolves coding problem solutions to generate harder, valid tasks, producing LiveCodeBench-Plus where frontier models score 27.5-62.6% and enabling RL gains on held-out tests.
-
PACE: Two-Timescale Self-Evolution for Small Language Model Agents
PACE coordinates low-risk prompt evolution with validated higher-risk control-logic updates to improve frozen SLM agents on benchmarks without model retraining.
-
Towards Direct Evaluation of Harness Optimizers via Priority Ranking
Priority ranking offers a low-cost direct evaluation for harness optimizers that correlates with their real multi-step optimization performance, supported by the Shor dataset of 182 scenarios.
-
PEEK: Context Map as an Orientation Cache for Long-Context LLM Agents
PEEK maintains a constant-sized context map via a programmable cache policy to give LLM agents persistent orientation knowledge about recurring external contexts, yielding 6-34% gains and lower cost than prior prompt-learning methods.
-
PQR: A Framework to Generate Diverse and Realistic User Queries that Elicit QA Agent Failures
PQR framework generates diverse realistic queries to elicit QA agent failures, uncovering 23-78% more unhelpful responses than prior methods in e-commerce agent tests.
-
REALISTA: Realistic Latent Adversarial Attacks that Elicit LLM Hallucinations
REALISTA generates semantically coherent adversarial prompts via latent-space optimization over input-dependent editing directions, achieving stronger hallucination elicitation than prior realistic attacks on open-source and reasoning LLMs.
-
Learning with Rare Success but Rich Feedback via Reflection-Enhanced Self-Distillation
RESD turns failure trajectories into token-level supervision via retrospective reflections and a persistent global playbook, enabling faster improvement than standard self-distillation or GRPO with only one rollout per prompt.
-
Shepherd: Enabling Programmable Meta-Agents via Reversible Agentic Execution Traces
Shepherd provides a reversible execution trace substrate for LLM agents that enables meta-agents to inspect and transform runs, yielding reported gains on coding and terminal benchmarks via supervision, counterfactual repair, and RL credit assignment.
-
SkillEvolver: Skill Learning as a Meta-Skill
A meta-skill authors and refines prose-and-code skills for agents by learning from post-deployment failures with an overfit audit, achieving 56.8% accuracy on SkillsBench tasks versus 43.6% for human-curated skills.
-
Empowering VLMs for Few-Shot Multimodal Time Series Classification via Tailored Agentic Reasoning
MarsTSC is a VLM agentic system with generator, reflector, and modifier roles that iteratively refines a knowledge bank to improve few-shot multimodal time series classification and produce human-readable explanations.
-
FlashEvolve: Accelerating Agent Self-Evolution with Asynchronous Stage Orchestration
FlashEvolve accelerates LLM agent self-evolution via asynchronous stage orchestration and inspectable language-space staleness handling, reporting 3.5-4.9x proposal throughput gains over synchronous baselines on GEPA workloads.
-
From Context to Skills: Can Language Models Learn from Context Skillfully?
Ctx2Skill automatically produces natural-language skill files via self-play between a challenger, a reasoner, and a judge, improving LLM performance on context-learning tasks.
-
AutoPyVerifier: Learning Compact Executable Verifiers for Large Language Model Outputs
AutoPyVerifier learns compact sets of executable Python verifiers from labeled LLM outputs via LLM synthesis and DAG search, improving objective prediction by up to 55 F1 points and downstream LLM accuracy by up to 17 points.
-
How Far Are Video Models from True Multimodal Reasoning?
Current video models succeed on basic understanding but achieve under 25% success on logically grounded generation and near 0% on interactive generation, exposing gaps in multimodal reasoning.
-
Training LLM Agents for Spontaneous, Reward-Free Self-Evolution via World Knowledge Exploration
LLM agents trained with a task-success reward on self-generated knowledge can spontaneously explore and adapt to new environments without any rewards or instructions at inference, yielding 20% gains on web tasks and allowing a 14B model to beat Gemini-2.5-Flash.
-
ContraPrompt: Contrastive Prompt Optimization via Dyadic Reasoning Trace Analysis
ContraPrompt extracts optimization rules from dyadic differences in reasoning traces on identical inputs and organizes them into input-aware decision trees, outperforming GEPA on four benchmarks with gains up to 8.29 pp.
-
SOCIA-EVO: Automated Simulator Construction via Dual-Anchored Bi-Level Optimization
SOCIA-EVO generates statistically consistent simulators by separating structural refinement from parameter calibration via bi-level optimization and falsifying strategies through execution feedback in a Bayesian-weighted playbook.
-
AdaExplore: Failure-Driven Adaptation and Diversity-Preserving Search for Efficient Kernel Generation
AdaExplore improves correctness and speed of Triton kernel generation by converting recurring failures into a memory of rules and organizing search as a tree that mixes local refinements with larger regenerations, yielding 3.12x and 1.72x speedups on KernelBench Level-2 and Level-3 within 100 steps.
-
EvoRAG: Making Knowledge Graph-based RAG Automatically Evolve through Feedback-driven Backpropagation
EvoRAG adds a feedback-driven backpropagation step that attributes response quality to individual knowledge-graph triplets and updates the graph to raise reasoning accuracy by 7.34 percent over prior KG-RAG methods.
-
Dive into Claude Code: The Design Space of Today's and Future AI Agent Systems
Claude Code answers recurring agent design questions with a thin model loop wrapped in dense safety, context, extensibility, and persistence harnesses, and those same questions get different answers in OpenClaw and Hermes.
-
Harness Engineering for Agentic AI Coding Tools: An Exploratory Study
Across 2,923 GitHub repositories, context files - especially AGENTS.md and CLAUDE.md - dominate agentic-tool configuration, while advanced mechanisms like Skills and Subagents are rarely used and mostly static.
-
Agentic Learner with Grow-and-Refine Multimodal Semantic Memory
ViLoMem is a dual-stream grow-and-refine memory system that separates visual and logical error patterns in MLLMs to improve pass@1 accuracy and reduce repeated mistakes across six multimodal benchmarks.