AI Repo & Hardness

Self-Improving Agent Harnesses

• 🔥 Self-Harness: Introduces harnesses that rewrite themselves from run data rather than remaining fixed...

Static task-level skill retrieval falls short for web agents as page states evolve during execution. SGDR introduces state-grounded dynamic retrieval...

New arXiv paper introduces SearchSwarm to tackle finite context limits in agentic LLMs via smart task delegation to subagents.

• Defines delegation...

Two recent papers strengthen the foundations of RL post-training for LLMs by refining trust-region control and credit assignment.

• Divergence...

Latent and graph-based memories are unlocking efficient long-context multimodal reasoning.

• Latent Memory replaces each evidence item with one...

Agent scaffolds are shifting from static wrappers to learnable artifacts that rewrite themselves based on run outcomes. This turns manual upkeep into compounding gains for long-horizon systems.

Role-Agent turns a single LLM into both agent and environment, using World-In-Agent state-prediction rewards and Agent-In-World failure-driven task retrieval to drive bootstrapped co-evolution and deliver >4% gains on benchmarks.

EEVEE is the first multi-dataset test-time prompt learning framework for LLM agents, deploying a router to cluster heterogeneous inputs and...

Diffusion models can move beyond greedy per-step decisions with Learned Relay Representations (LRRs), enabling them to plan for the future.

Two fresh benchmarks emphasize long-horizon rigor for agents and world models.

• Workflow-GYM tests professional GUI workflows; even top models reach...

RHO enables agents to self-improve their harness using only prior trajectories: it picks challenging tasks, generates rollouts, applies...

Inference Optimization Releases

• 🔥 End-to-End Context Compression at Scale: Presents Latent Context Language Models (LCLMs) with...

Two new arXiv papers highlight complementary techniques for boosting agent performance via skill and trajectory refinement.

• Bayesian-Agent models...

Two fresh benchmarks move beyond binary task success to probe deeper agent capabilities.

• SWE-Explore isolates repository navigation, requiring...

Memory mechanisms are rapidly evolving from simple agent stores to sophisticated world-model and context-compression systems.

• Agentmemory delivers...

Live GitHub data reveals clear leaders among agent frameworks.

• LangChain tops at 138.8K stars; CrewAI follows with 53K stars and 14.9M monthly PyPI...

• Human questionnaires mischaracterize LLMs, as lexical cues trigger desirable responses unlike realistic queries
• Demographic prompts shift...

Evaluation Optimization Failure Modes

• 🔥 Gradient Dilution in LLM Judges: arXiv:2605.26046 documents 59% task-focus drop and Spearman rho...

Trending AI GitHub Repositories

• 🔥 Weekly Explosive Growth Lists: 5 GitHub Repos That Exploded This Week and Top 10 AI GitHub Repos This Week...

Textual gradient methods for LLM judges face unique challenges when optimizing across multiple criteria simultaneously, unlike numerical multi-task...