BREAKING
Tencent Unveils Hyra-1.0 Research Agent
Bilevel Loop: Improve Solution and Grader
1
Context Agent + Experience Bank
↓
2
Proposal Agents pipeline
↓
3
Sandbox scoring via solve.sh
↓
4
Outer loop upgrades evaluator
0
new math records of 55
0
trainable parameters
0
%
parameter reduction
SOL-ExecBench Mean SOL
Hyra-1.0
0.771
Recursive
0.754
0
%
routing gain over SABRE
0
R-squared, held out-of-sample
Robustness Still Largely Untested
AI NEWS BLITZ
Tencent's Hunyuan team has launched Hyra-1.0, its first autonomous research agent.