Tencent Hunyuan releases Accelerate Bench: 32 CPU tasks to grade coding agents' optimization
_akhaliq · x · 2026-11-18
Tencent Hunyuan released Accelerate Bench (arXiv, Bowen Liu et al.), a benchmark evaluating coding agents on application-level program optimization.
- 32 CPU optimization tasks adapted from scientific applications and supercomputing competitions; each provides a working program with specified workload, correctness requirements, timing boundaries, and permitted edits — agents choose where and how to optimize.
- On benchmark v1.1, 16 coding agents (model + harness pairs) scored 25.5%–67.4% at Score@1; speed ceiling separates the field — the top two agents hit at least 2x on 22+ of 32 tasks, while the other fourteen reached 2x on only 8–17.
- Controlled replays show why cross-application coordination matters: identical source code runs 0.29x–2.12x depending only on build/runtime settings, and one change is worth 1.01x–6.4x depending on context.
- The benchmark pairs executable workloads with agents' programs and trajectories to evaluate both gains and process.
Related event: Tencent Hunyuan Releases Accelerate Bench for Evaluating Coding Agents(2 posts)→
More from coding & agent
- YC-backed Ample launches one-stop cloud for agents, hits 5,000 users two weeks after launch — ycombinator · 2026-11-18
- OmniLoRA: train any model on any GPU inside your ComfyUI graph — pente5 · 2026-11-18
- findatasets ships /feedback endpoint so AI agents can file bug reports, following Coinbase's lead — virattt · 2026-11-18
- Put rules in code, not prose: what actually makes an AI instruction a Rule — gerardsans · 2026-11-18
- Dev shares a tiered file-permission policy to stop agents stalling on approvals — gaganghotra_ · 2026-11-18
- Museo, a whisperable AI ring for phone-free walks, ships next month with open dev platform — Scobleizer · 2026-11-18