Engineering
Training Agents to Evolve with Their Harness: The TaoLive Digital Avatar System
A technical report from Alibaba’s Taobao Live team describes Harness-Aware Training (HAT), a method for training compact models to adapt…
Engineering
PILOT: A Supervisor-Worker Harness for Live Agent Self-Improvement
This paper introduces PILOT, a supervisor-worker agent harness that performs self-improvement live during a run rather than only after it…
Engineering
Understanding Evolution Strategies for LLM Reasoning: Broader Coverage Than GRPO
This study systematically compares Evolution Strategies (ES) against Group Relative Policy Optimization (GRPO) as post-training paradigms for LLM reasoning. It…
Engineering
WikiSkill: Co-Evolving Agent Skills with a Persistent Knowledge Base
WikiSkill is a framework that co-evolves an LLM agent’s reusable skill library alongside a persistent wiki-style knowledge base, separating raw…
Engineering
CaSKG: Counterfactual-Causal Skill Graphs for Agent Skill Retrieval
CaSKG is a retrieval framework that calibrates the reliability of a skill graph’s edges before using it to retrieve procedural…
Engineering
What Does an Evaluation License? A Reproducibility Audit of Inspect Evals
This paper formalizes what is required to legitimately ‘replay’ a claim attached to an LLM evaluation metric, then applies that…
Engineering
VoiceMem: Streaming Dual-Brain Memory for Real-Time Speech Agents
VoiceMem is a memory architecture for duplex speech language models built around a parallel ‘informational left brain’ and ’emotional right…
Engineering
FrontierChallenge: A Benchmark for End-to-End Scientific Agent Workflows
FrontierChallenge is a new cross-domain benchmark of 300 end-to-end scientific workflows spanning quantum chemistry, molecular dynamics, materials science, and other…
Engineering
WarpSAC: Regime-Aware Off-Policy Reinforcement Learning for Massively Parallel Simulation
This paper shows that common off-policy RL stabilizers behave differently depending on data regime: parameter normalization helps under narrow replay…
Engineering
JIT-Agent: Just-in-Time Synthesis of Agent Harnesses
JIT-Agent is a harness intelligence model that generates task-adaptive agent harnesses on the fly for arbitrary off-the-shelf LLMs, formalizing the…
Engineering
D3-MOPD: Dynamic Domain Scheduling for Multi-Teacher Distillation
D3-MOPD is a zero-overhead scheduler for multi-teacher on-policy distillation that adjusts per-domain training mixture ratios online, reusing the reverse-KL signal…
Engineering
Agent-G2: Gaussian Guidance for Hint-Based Agentic Reinforcement Learning
Agent-G2 addresses reward sparsity in long-horizon agentic RL by replacing the common practice of treating expert-trajectory guidance depth as a…