Tencent's Hunyuan team on July 21, 2026, released Hyra-1.0, a research agent built to recursively refine its own outputs on performance-driven research and engineering tasks — a system that iteratively improves not just with each prompt, but with each cycle of self-generated feedback.
July 21, 2026 · Tencent Hunyuan
Hyra-1.0: An Agent That Rewrites Its Own Solutions
A research agent built to recursively refine its own outputs on performance-driven research and engineering tasks — improving with every cycle of self-generated feedback, across AI4AI, AI4Science, and AI4Fun.
29
new records on 55 unsolved math problems
44.4%
better qubit routing on IBM's Q20 vs SABRE
58.3%
smaller Transformer for 10-digit addition
0.77
out-of-sample R² recovering sunspot regression
Model size for 10-digit addition
Fewer parameters = a leaner solution. Each block = 3 parameters.
The recursive self-improvement loop
Context Agentsynthesizes from Experience Bank
→
Proposal Agentsgenerate candidates
→
Sandbox & Scoreexecute + evaluate
↺
When no reliable evaluator exists, a bilevel loop improves both the solution and the evaluator itself — with measures to counter reward hacking.
What impresses
NanoChat BPB 0.9015 vs 0.9109 baseline
NanoGPT Speedrun 3.28 loss in 76.4s vs 77.5s
SOL-ExecBench mean 0.771 vs 0.754
Asynchronous multi-agent, self-evolving evaluator
Flagged limitations
Reward hacking: bidirectional attention inflated a BPB score
Cache exploit let empty kernels pass evaluation
Research- and demo-focused only
No pricing, availability, or API disclosed
The stated ambition reaches beyond public leaderboards toward product systems, real R&D pipelines, scientific discovery, and industrial applications — but whether recursive self-improvement holds up outside curated benchmarks is the key test ahead.
Continue reading The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in ✓ Signed in — this article isn’t included in your current plan.Unlocking the full article…