AI Tech Daily - 2026-09-03 | Recsys Frontier

type

Post

status

Published

date

Sep 3, 2026 05:00

slug

ai-daily-en-2026-09-03

summary

📊 Today's Overview

Autonomous software development took a big step forward today. Shanghai AI Lab's Harness-of-Harness framework lets coding agents run multi-day, self-improving development cycles — it built a complete FPS game across 70+ iterations with a 52% average gain over standalone harnesses. AMD open-sourced Instella-MoE, a fully transparent 16B-parameter MoE model trained from scratch on MI300X GPUs, complete with weights, data mixtures, and training code. On the RL front, Alibaba's CANOPY shows outcome-only reinforcement learning can top leaderboards on long-horizon tasks — no dense rewards, no skill libraries, just scaled exploration and on-policy discipline. Microsoft also contributed two evaluation frameworks: SAGE delivers state-grounded dialogue judging at zero LLM cost, and ReDeck brings step-level render feedback to slide generation.

🔥 Trend Insights

Multi-day autonomous agents arrive: Harness-of-Harness runs 70+ iteration development loops that build complete games, shifting coding agents from single-session tasks to sustained, self-improving projects.

Outcome-only RL gets a second look: Alibaba's CANOPY tops AppWorld with pure end-state rewards, challenging the assumption that small open models need dense signals or skill libraries for long-horizon tasks.

Fully open MoE models gain ground: AMD's Instella-MoE ships complete training stacks — weights, data, code — outperforming prior open models and lowering the barrier for reproducible MoE research.

📄 Paper Highlights

Harness-of-Harness: Multi-Day Autonomous Software Development with Continual Improvement

Shanghai AI Laboratory ｜ 🏷️ Agent Framework, Code Agent, Multi-Agent

Framework that wraps existing coding-agent harnesses into iterative planning-coding-testing loops. Achieves 52% average relative gains and autonomously develops a complete FPS game over 70+ iterations — evidence that agents can sustain multi-day engineering projects.

Instella-MoE Technical Report

AMD ｜ 🏷️ MoE, Architecture, Training

Fully open 16B-parameter MoE model (2.8B active) trained from scratch on AMD MI300X GPUs. Releases weights, data mixtures, and training code — a strong reproducible foundation for efficient MoE research beyond the NVIDIA ecosystem.

Explore More, Drift Less: Outcome-Only Reinforcement Learning Can Suffice for Long-Horizon Interactive Agents

Alibaba Research ｜ 🏷️ RLHF/DPO, Agent Framework, Code Agent

CANOPY protocol shows outcome-only RL tops AppWorld's leaderboard on a Qwen3-14B policy — no dense rewards or scaffolding needed. Diagnoses signal starvation and policy drift as the real ceilings, then fixes both with scaled exploration and KL-anchored updates.

MemoryWalker: Stop Training Agents on Contexts They Never Saw

Alibaba Group ｜ 🏷️ Agent Framework, Training, RLHF/DPO

Fixes the conditioning problem when training agents under context compression — evictions turn the learning object into a tree, not a sequence. Proposes exact gradient-equivalent corrections plus SDCC, a single-backward-pass relaxation that works with black-box harnesses.

From Production Traffic to Post-Training: Building a Self-Hosted LLM That Covers the Corporate Request Mix

T-Tech ｜ 🏷️ Fine-tuning, RLHF/DPO, Tool Use

Consolidates 200+ internal apps onto one self-hosted model by training separate GRPO experts per failure axis and merging via two-stage SLERP. Handles 116M requests monthly at a fraction of prior serving cost — a practical playbook for data-residency-constrained enterprises.

ReDeck: Step-Level Render-Grounded Refinement for Document-to-Slide Generation

Microsoft Research ｜ 🏷️ Agent Framework, Multimodal, Reasoning

Shifts slide-refinement feedback from turn-level to step-level — "one edit, one observation" — with renderer-derived signals catching overflow, overlap, and clipping as they happen. Consistently beats existing slide agents across GPT-5.4, Claude-4.6, and Gemini-3.1.

SAGE: State-Grounded, Abstention-Aware Evaluation of Task-Oriented Dialogue Agents

Microsoft ｜ 🏷️ Evaluation, Agent Framework, NLP Task

Evaluates dialogue agents by checking per-turn workflow state transitions rather than holistic reply quality. SAGE-Core decides 81-91% of criteria at zero LLM cost and ties state-aware GPT-4.1 judges — a serious challenger to LLM-as-a-judge economics.

Verifiable Disaster Storylines and Causal Knowledge Graphs: A Citation-Grounded Pipeline from Heterogeneous Humanitarian Sources

European Commission JRC ｜ 🏷️ RAG, Knowledge Graph, Safety

Pipeline fusing EM-DAT, ReliefWeb, and media monitoring into citation-grounded disaster storylines and causal knowledge graphs. Human evaluation with domain experts confirms strong faithfulness and clear preference for grounded outputs — built for crisis-response situational awareness.

Topic Matching in the Wild: Benchmark and Lessons from Real-World ASR Transcripts

Dialpad Inc. ｜ 🏷️ NLP Task, RAG, Inference

Benchmark for real-time agent-assist topic matching on noisy, unpunctuated call-center ASR transcripts. Lightweight Gemini-based LLM matchers with natural language descriptions beat regex and embedding baselines — practical signal for contact-center tooling.