Hyra 2.0 Agent Now Builds Tools, Not Just Answers

The AI lab at Chinese tech giant Tencent has released an updated version of its research agent, Hyra 2.0. Its predecessor, Hyra 1.0, solved a math conjecture on its own three months ago. The conjecture was in additive combinatorics, the field that studies sums and differences of sets of numbers, and it had stood unsolved for more than fifty years. That was a first for a Chinese open model. Hyra 2.0 goes further. It does not just solve problems. It builds general-purpose tools too: models, pipelines, and validators meant to be useful to the whole research community.

For example, Hyra 2.0 built a model for designing protein sequences called HyInv. It has just 1.96 million parameters. The agent trained it entirely on its own, through pretraining, fine-tuning, reinforcement learning, and acceleration. The resulting model generates protein sequences that match a target structure better than ProteinMPNN, today's best tool for the job, and it runs 57 times faster. Another model, HyPLM-Flash, has 650 million parameters and was trained on 726 billion tokens, a unit of text data. That is 0.73 of the data volume used to train the popular protein model ESM-2 650M. It also used between a fifth and an eighth of the data used to train two other protein models, E1-600M and ESM C 600M respectively. Tested against 12 other protein-analysis models, HyPLM-Flash came out the fastest and placed third on average across 35 trials.

Hyra 2.0 also came up with a new error-correction code for quantum computers on its own. An error-correction code is a method for catching and fixing errors during quantum computation. With parameters [488, 123, 22], this code beats the current best entry on the open qLDPC Challenge leaderboard for quantum error-correction codes by about 48%. It also recalculated the distances for 178 of the 187 codes on that leaderboard, running 8 to 55 times faster than the original CPU implementation.

The research agent also contributed to the open-source XiangShan processor project, which is built on the open RISC-V instruction set architecture. Thanks to the agent's changes, the processor gained a 5.79% performance boost per GHz on the SPEC CPU2006 benchmark, with no change to clock speed and a chip area increase of no more than 0.3%. The Hyra team says a gain like this is roughly equal to two or three months of work by the project's human developers. The agent found ten bugs through fuzzing, a method of automated testing with random inputs, and one more serious bug through formal verification, a formal method for proving code correctness. That bug turned out to be a cache inconsistency between different processor cores, and it had not yet been logged in the project's bug tracker. The fix was accepted into OpenLLC, an open-source chip design project.

Finally, Hyra 2.0 proved two open math conjectures and had the proofs checked alongside human experts. One conjecture concerns circle actions and was proposed by the mathematician Yau. The other, about equidistant points in five dimensions, was proposed by the mathematician Petty. In robotics, the agent designed an odd-looking five-fingered robot hand whose fingers grow longer as they spiral outward. With extra training using the same amount of compute, it can spin objects 28.6% faster than the best solution described in the literature so far. That result has only been checked in simulation, not on physical hardware.