arxiv arXiv cs.CL · just now · src: 2d ago · research

TaoLive introduces Harness-Aware Training for evolvable digital avatar agents

from English

Researchers present TaoLive, a technical report on training digital-avatar agents that can adapt to changing business strategies in real-time live e-commerce without retraining. The system uses an evolvable Harness to decouple skills and tools from model weights, addressing the challenge of models memorizing outdated configurations through Harness-Aware Training (HAT).

HAT employs Harness-State Augmentation across three stages: supervised fine-tuning, general on-policy distillation, and agentic reinforcement learning.
The compact 35B model scores 94.8 on Live-Stream QA and 94.6 on Harness-Variant QA, outperforming the base model (80.3) and strongest general LLM (93.0).
The system achieves 3.407 s P50 and 8.114 s P95 latency on a single NVIDIA H20 GPU with MTP enabled.

This approach produces a latency-feasible compact agent that remains effective under evaluated Harness changes without sacrificing general instruction following.

Importance 2/3 New harness with differentiators arXiv cs.CL AI agents Training methods

Benchmarks

Benchmark	Model	Score
IFEval	compact 35B model	83.5%

Read original

arxiv arXiv cs.CL · 1d ago · 7 views

SocialRL trains 4B model to match GPT-5 in negotiation via social reasoning

Researchers present SocialRL, a training recipe that directly reinforces social reasoning in small language models to improve their performance as strategic negotiators. Applied to a 4B model across six domains including Deal-or-No-Deal and Job Interview, the approach enables the model to match or exceed GPT-5 family performance on held-out scenarios.

media MarkTechPost · 16d ago · 17 views

NVIDIA NeMo releases Molt, a compact PyT-native agentic RL framework

NVIDIA's NeMo team has released Molt, an open-source agentic reinforcement learning framework designed to reduce the complexity of algorithm iteration for researchers. The codebase is approximately 8.6K lines of RL code, significantly smaller than comparable frameworks like verl (62K lines) and slime (25K lines).

media Hugging Face Forums · 16d ago · 15 views

Investigating whether failed agent traces are usable as fine-tuning data

The author proposes a hypothesis that failed tool-using agent traces, such as wrong tool selections or malformed arguments, can serve as valuable fine-tuning data to train models away from these errors. The post seeks community feedback on whether training on corrected failure traces is effective or detrimental.

lab Microsoft Research Blog · 18d ago · 18 views

Microsoft Research introduces EvoLib for test-time learning with an evolving library

Microsoft Research has introduced EvoLib, a framework that enables large language models to learn from their own experience during inference without requiring ground-truth labels or external feedback. Rather than storing raw experiences as static memories, EvoLib transforms them into reusable skills and reflective insights that are continually refined, consolidated, and reweighted.

lab Microsoft Research Blog · 18d ago · 23 views

Microsoft releases Echoverse for training computer-use agents

Microsoft Research has released Echoverse, a collection of twelve deep synthetic environments designed to train computer-use agents. The initiative includes ten domain-specific worlds and two capability-focused worlds, all built to maintain coherent state and behavioral fidelity.