arxiv arXiv cs.AI · just now · research

Faraday 27B surpasses Claude Opus 4.8 and GPT-5.5 on paper replication tasks

from English

Researchers have developed Replica, a scalable task space for paper replication, and introduced an auto-generated rubric-based judge to provide low-noise reward signals that align with human assessment.

The team post-trained Faraday, a 27B-parameter "AI Scientist" agent that utilizes coding agents as tools.
Faraday surpasses the performance of Claude Opus 4.8 and GPT-5.5 on held-out replication tasks.
Qualitative analysis indicates that Faraday adopts a more scientifically-principled approach during rollouts.

The authors believe these results provide a stepping stone towards AI agents capable of long-horizon scientific innovation without requiring complex harnesses.

Importance 2/3 New harness with differentiators arXiv cs.AI DeepSeek AI agents Research paper

Read original

arxiv arXiv cs.CL · 40d ago

AutoMem automates learning of memory as a cognitive skill for LLMs

Researchers introduce AutoMem, a framework that treats memory management in large language models as a trainable skill rather than a static component. The system promotes file-system operations to first-class memory actions, allowing the model to autonomously decide how to encode, retrieve, and organize knowledge.

arxiv arXiv cs.CL · 45d ago

Scaling the Horizon, Not the Parameters: Reaching Trillion-Parameter Performance with a 35B Agent

Researchers introduce Agents-A1, a 35B Mixture-of-Experts model that achieves performance comparable to trillion-parameter models by scaling the agent horizon rather than parameter count. The approach focuses on extending long-horizon trajectories and unifying heterogeneous agent abilities through a specialized training infrastructure.

arxiv Moonshot AI (arXiv) · 255d ago · 2 views

DeepSeek-V3.2 introduces Sparse Attention and agentic synthesis, surpassing GPT-5

DeepSeek has released DeepSeek-V3.2, an open large language model that combines high computational efficiency with advanced reasoning capabilities. The release highlights three main technical advancements: a new sparse attention mechanism, a scalable reinforcement learning framework, and a pipeline for synthesizing agentic training data.

arxiv arXiv cs.AI · 1h ago Live

AutoDesign uses meta-harness optimization to improve long-horizon agentic design

The paper introduces AutoDesign, a framework that employs a meta-harness optimizer to guide a code agent in recursively improving its harness based on rollout feedback. This approach aims to align with human design priors and accumulate reusable experience for recursive self-improvement.

arxiv arXiv cs.CL · 2h ago

AutoDesign uses meta-harness optimization to improve long-horizon agentic design

The paper introduces AutoDesign, a framework that employs a meta-harness optimizer to guide a code agent in recursively improving its harness based on rollout feedback. The authors evaluate this approach on the academic paper-to-poster generation task using a new benchmark called PosterBench.