Auto-Research Arena: 6,300 runs, agents rediscover MQA/MLA, memory layers and more
qixing_huang · x · 2026-09-25
Auto-Research Arena shows what auto-research agents can discover: across 6,300 research runs, agents with no internet access independently found mechanisms closely echoing recent frontier work — looped-Transformer-like designs, depth-specific residuals, MQA/MLA, memory layers, cross-layer KV sharing, fused inference kernels, and more, collectively touching nearly every major direction in LLM architecture and efficiency.
The key design choice: Arena expands research beyond a single loss objective, letting agents explore a much broader frontier of benchmark-validated ideas.
More from coding & agent
- Stripe engineer answers 'what's it like at Stripe' with demos over memos culture — schwentker · 2026-09-25
- Perplexity's Fast Search costs $1 per 1,000 requests at 200 ms, 5x cheaper than standard search — AravSrinivas · 2026-09-25
- Wasmer runs a real PostgreSQL 18.4 server on iOS and in the browser via WebAssembly — jedisct1 · 2026-09-25
- Google Cloud details 4 patterns for real-time AI voice agents that see, talk, think and code — rseroter · 2026-09-25
- Opus 5.5 coded a polished 30-second hype video from one prompt in 3 hours of autonomous work — DaveRogenmoser · 2026-09-25
- NeurIPS Oral: IDS co-evolves code with formal proofs, hitting 3x Claude Code's success rate — adityagp · 2026-09-25