Skip to content
Projects

Tencent-Hunyuan/Hyra-results has 254 GitHub stars and 15 forks. GitHot measured +1 stars in 24 hours. Last checked 2026-11-15 23:18 UTC. Last push 2026-11-14.

How these counts are measured

Measured star growth

in 24h
+1
in 7 days
Update delayed
in 15 days
Update delayed
in 1 month
Update delayed

Net star changes over overlapping periods, not daily totals. Missing measurements are left blank. Stars measure attention, not software quality.

About

Research artifacts from Hyra (/ˈhaɪ.rɑː/)

AI

README.md

Original

Hyra Results

Companion artifacts for Hyra: Hunyuan Research Agent.

This repository collects concrete solutions that Hyra produced across a range of open problems in science, mathematics, engineering, and creative design, released alongside the Hyra launch post. Each folder holds the final solution artifact and, where relevant, the self-contained scripts that reproduce it.

📝 Launch post: hy.tencent.com/research/hyra

🚀 Hyra 2.0 blog: hy.tencent.ai/research/hyra2

News

  • 2026-11-04: 🌊 HyPDE, a 2D foundation model for PDEs, released with weights and code. HyPDE, a 323M-parameter fused-history transformer trained from scratch on thirteen 2D physics datasets, forecasts simulated fields from their last six frames. Zero-shot on the test split of thirteen datasets of The Well and FlowBench, against Walrus (1.29B parameters, released weights, no fine-tuning on either side), it has the lower median VRMSE on 9 of 13 datasets in one-step forecasting, 10 of 13 over rollout frames 1-20 and 8 of 13 over frames 21-60. Weights, the hypde package and the full table: AI4Science/hypde/.

  • 2026-11-03: 🔐 Certified distances of non-CSS quantum codes. The qldpc-challenge board labels all 184 of its general (non-CSS) stabilizer codes as upper bounds only. hyqec, a prover plus an independently written certificate checker, proves the exact Pauli-weight distance of 165 of them: 122 with SAT proofs checked by drat-trim, 42 with re-executable branch-and-bound records, and the board's [[452,115,18]] (d = 18) with a coverage certificate of 124.6 billion search nodes that the independent checker re-executed in full. Code: AI4Science/hyqec/.

  • 2026-11-02: 🚀 Hyra 2.0 blog released. Read Hyra 2.0: The Next Step for Auto Research for an overview of Hyra's progress in developing reusable scientific tools, models, and workflows.

  • 2026-11-02: 🧫 HyCell, a single-cell embedding model, released with weights and code. HyCell, a 16.1M-parameter model, maps the raw counts of a human cell to a 512-dimensional embedding and decodes embeddings back to gene expression. On seven atlases held out from pretraining, its frozen embeddings reach an overall score of 0.735, level with Stack (0.736, 218M parameters). Weights, the hycell package and the technical report: AI4Science/hycell/.

  • 2026-11-02: ⚛️ HyFF force fields released with weights and code. Five machine-learning force fields for molecular simulation, from 2.3M to 1.2B parameters, covering 83 elements with total charge and spin and trained on OMol25. On the full OMol25 validation set, HyFF Andante reaches a force error comparable to UMA-S-1.2 (7.65 vs 7.55 meV/Å) at 29× its MD speed, and HyFF Largo has the lowest energy error of the evaluated models (0.364 meV/atom; DPA4-Pro: 0.380) at 9.7× DPA4-Pro's speed (MD of a 2,586-atom system on one H20). The hyff package runs NVE, NVT and NPT molecular dynamics and batched prediction: AI4Science/hyff/.

  • 2026-11-02: 🧬 Two protein models released with weights and code. HyInv, a 1.96M-parameter inverse-folding model, and HyPLM-Flash, a 650M-parameter protein language model with a fast INT8 inference engine, are published under AI4Science/ together with their weights, inference code and technical reports: AI4Science/hyinv/, AI4Science/hyplm_flash/.

  • 2026-11-02: 📐 Two results on Petty's and Yau's conjectures.

    • Petty's equilateral-set conjecture in dimension five. The highest settled dimension rises from 4 (Makeev, 2005) to 5: every five-dimensional real normed space contains six equidistant points. Paper: AI4Science/petty_dimension_five/.
    • Yau's circle-action conjecture (unrestricted form). A closed seven-manifold with strictly positive sectional curvature admits no effective smooth circle action, giving a counterexample. The manifold is not simply connected; the simply connected version remains open. Paper: AI4Science/yau_circle_action_conjecture/.
  • 2026-08-19: 🧾 Three new results, each with a Lean 4 formalization of its finite and numerical core.

    • Beurling–Ahlfors transform. Iwaniec's 1982 conjecture predicts the sharp Lᵖ operator norm p*−1; the best proven uniform coefficient drops from 1.575 (Bañuelos and Janakiraman, 2008) to 1.523958. Paper: AI4Science/beurling_ahlfors_bellman/.
    • Partial Hadamard matrices. Their asymptotic count is now known throughout every power-law range above n², not just the cubic range m ≫ n³, lowering the critical exponent from 3 (arXiv:2603.30013) to 2. Paper: AI4Science/partial_hadamard_counting/.
    • Commutators close to the identity. The identity is never a commutator, but comes within ε at a price: the least ‖D‖·‖X‖ needed falls from Tao's O(log⁵(1/ε)), since refined to O(log⁴), to O(log³(1/ε)), against Popa's log(1/ε) lower bound. Paper: AI4Science/commutator_log_cubed/.
  • 2026-08-17: 📐 Blaschke–Lebesgue in three dimensions. The least volume of a convex body of constant width w is conjectured to be Meissner's ≈ 0.419860 w³; the best certified lower bound rises from 4π/33 · w³ ≈ 0.380799 w³ (Nishioka, arXiv:2606.01754) to > 0.411040 w³, closing 77.4% of the remaining gap. Paper: AI4Science/3d_blaschke_lebesgue/.

  • 2026-07-29: 🎉 The sum-vs-difference problem is settled. For finite A ⊆ ℤ, the optimal exponent relating |A+A| to |A−A| has supremum exactly 2, approached arbitrarily closely but never attained: a complete, machine-checked resolution going beyond the sums_diffs record below. Proof: sum-diff-proof.

Results

Each row compares the best prior published result or conjecture status (Prev best, from the cited sources) against Hyra. Arrows mark the better direction (↓ lower is better, ↑ higher is better); the Hyra result is in bold.

Track Task Metric Prev best Hyra
AI4AI nanochat_autoresearch val BPB ↓ 0.9109 e 0.9015
nanogpt_speedrun wall-clock ↓ 77.5 s e 76.4 s (mean val loss: 3.280)
sol_execbench score ↑ 0.754 e 0.771
AI4Science autocorrelation_first C₁ ↓ 1.502870 a 1.502850
autocorrelation_second R ↑ 0.962694 b 0.962901
erdos_min_overlap C₅ ↓ 0.380868 b 0.380859
sums_diffs C(A) ↑ 1.14489 b 1.21079
packing_records records broken n/a f 100
smallest_adder params ↓ 36 c 15
parp1_docking objective ↓ −9.77 d −10.60
qubit_routing CNOTs added ↓ 269,037 b 258,369
sunspot_symbolic forecast R² ↑ 0.47 g 0.78
3d_blaschke_lebesgue Vol/w³ bound ↑ 0.380799 h 0.411040
beurling_ahlfors_bellman C_BA ↓ 1.575 i 1.523958
partial_hadamard_counting regime exponent ↓ 3 j 2
commutator_log_cubed log exponent ↓ 4 k 3
petty_dimension_five dimension settled ↑ 4 l 5
yau_circle_action_conjecture conjecture status open m counterexample (dim 7)
hyqec board non-CSS codes with exact d ↑ 0 / 184 n 165 / 184 (122 with drat-trim-checked SAT proofs, 1 independently re-checked certificate)

Prev-best sources.

  • a TTT-Discover: Learning to Discover at Test Time (arXiv:2601.16175).
  • b SimpleTES: Evaluation-driven Scaling for Scientific Discovery (arXiv:2604.19341).
  • c AdderBoard trained-weights leaderboard (github.com/anadim/AdderBoard).
  • d Olaparib, an approved PARP1 inhibitor (drug baseline).
  • e Recursive: First Steps Toward Automated AI Research (github.com/recursive-org/first-steps-toward-automated-ai-research); AI4AI baselines: nanoGPT-speedrun, nanochat, and SOL-ExecBench.
  • f Erich Friedman's Packing Center (erich-friedman.github.io/packing). Hyra's record-improving packings are credited there as "Found by Haowei Lin": 100 across 28 shape-in-shape families, each beating the previously listed best.
  • g Baseline: a "copy last frame" (persistence) forecast. Hyra's score is the forecast R² on a fully-held-out, half-century-long segment of the record.
  • h Nishioka: An improved lower bound for the three-dimensional Blaschke–Lebesgue problem from spectral and dual perspectives (arXiv:2606.01754), which gives 4π/33 ≈ 0.380799109526. Hyra's certified bound is (130838246407123/10¹⁵)·π > 0.411040473721188.
  • i R. Bañuelos and P. Janakiraman, Lᵖ-bounds for the Beurling–Ahlfors transform (2008), which gives the uniform coefficient 1.575. Iwaniec's conjectured sharp value is 1.
  • j D. Davis: Counting Partial Hadamard Matrices in the Cubic Regime (arXiv:2603.30013).
  • k B. Bilich: An O(log⁴(1/ε)) refinement of Tao's construction of commutators close to the identity (github.com/bilichboris/TaoCommutators); the published exponent is Tao's 5, since refined to 4.
  • l V. V. Makeev: Equilateral simplices in normed 4-space, Zap. Nauchn. Sem. POMI 329 (2005), 88–91; English translation in J. Math. Sci. 140(4) (2007), 548–550, establishing the four-dimensional case of Petty's equilateral-set conjecture.
  • m S.-T. Yau: Open problems in geometry, Proc. Sympos. Pure Math. 54, Part 1 (1993), 1–28; see also Structure of manifolds with positive curvature based on geometric analysis, Notices of the ICCM 1(2) (2013), 24–27, for the smooth circle-action conjecture.
  • n The qldpc-challenge board (github.com/unitaryfoundation/qldpc-challenge) labels every one of its general stabilizer (non-CSS) entries as an upper bound (d≤) only, as its certifier does not minimise Pauli weight (TRACKS.md, 2026-10-27). Codes are from the 2026-10-16 snapshot. Exact distances of other non-CSS codes have been computed elsewhere (e.g. Lee et al., arXiv:2609.30069); the count is about this board. Hyra's 165 are made up as follows. 122 have SAT proofs checked by drat-trim; the CNF encoding is hyqec's own. 1 ([[452,115,18]]) has a branch-and-bound coverage certificate re-executed by the independent checker. 42 have branch-and-bound search records that only the same prover can re-run.

Metrics are as defined by each benchmark; comparisons are against the cited published results and conjecture statements.

Metric notes.

  • C₁: first autoconvolution/autocorrelation constant, max(f∗f)/(∫f)² (minimize).
  • R: second autocorrelation ratio, ‖f∗f‖₂² / (‖f∗f‖₁·‖f∗f‖∞) (maximize).
  • C₅: Erdős minimum-overlap constant (minimize).
  • C(A): sum-vs-difference exponent, log(|A+A|/|A|) / log(|A−A|/|A|) (maximize).
  • params: unique trainable parameters of a transformer that adds two 10-digit integers at ≥ 0.99 accuracy (minimize).
  • objective: PARP1 docking objective, Vina score + 10·(1 − QED) (minimize).
  • CNOTs added: extra CNOTs inserted by SWAP routing (minimize; 1 SWAP = 3 CNOTs).
  • forecast R²: rolling-origin, free-running (24-month) forecast R² for monthly sunspot numbers (maximize).
  • Vol/w³ bound: largest proven universal lower bound on Vol(K)/w³ over all convex bodies K ⊂ ℝ³ of constant width w (maximize; the conjectured optimum is ≈ 0.419860).
  • C_BA: smallest proven uniform coefficient in ‖B‖_{Lᵖ→Lᵖ} ≤ C_BA·(p*−1) for the Beurling–Ahlfors transform, over all 1 < p < ∞ (minimize; the conjectured optimum is 1).
  • regime exponent: infimum of the power-law exponents u for which m/nᵘ → ∞ is known to force the asymptotic formula for N(n,m) (minimize; the endpoint value itself is not claimed).
  • log exponent: smallest proven k in m(ε) = O(logᵏ(1/ε)) for commutators within ε of the identity (minimize; the matching lower bound is k = 1).
  • dimension settled: highest dimension n for which every real n-dimensional normed space is known to contain n + 1 equidistant points, as conjectured by Petty (maximize).
  • conjecture status: status of the unrestricted claim that every closed manifold admitting strictly positive sectional curvature admits an effective smooth circle action. The counterexample has π₁ ≅ (ℤ/6)²; the simply connected version remains open.
  • board non-CSS codes with exact d: number of the board's 184 general stabilizer codes whose Pauli-weight distance is proven exactly, not only bounded (maximize).

Note. The results above were current as of 2026-07-10, except 3d_blaschke_lebesgue, added 2026-08-17, the three entries added 2026-08-19, petty_dimension_five and yau_circle_action_conjecture, added 2026-11-02, and hyqec, added 2026-11-03. Several of these problems live on public, continuously-updated leaderboards; later entries there may warm-start from Hyra's published solutions to reach still-better numbers.

Model releases

Trained models, each with its weights, inference code and technical report:

  • AI4Science/hyinv/: HyInv, a 1.96-million-parameter message-passing graph network for protein inverse folding (backbone → sequence), trained with structure-reward reinforcement learning. The 7.8 MB weights ship in the folder; see its README for the CLI and Python API.
  • AI4Science/hyplm_flash/: HyPLM-Flash, a 650-million-parameter protein language model with an INT8 GPU inference engine (about 130,000 residues per second on one H20). The 2.6 GB weights are a GitHub Release on this repository; weights/download.sh fetches, merges and verifies them.
  • AI4Science/hycell/: HyCell, a 16.1-million-parameter embedding model for single-cell transcriptomes, pretrained without cell-type labels on 79.5 million human cells of the CZ CELLxGENE Census. It maps raw counts to 512-dimensional embeddings and decodes them back to expression, and can be fine-tuned into a cell-type annotator. The 64 MB weights ship in the folder as part of the hycell package.
  • AI4Science/hyff/: HyFF, five machine-learning force fields for molecular simulation (Presto, Vivace, Moderato, Andante and Largo; 2.3 million to 1.2 billion parameters, 83 elements), with the hyff package for single-structure inference, NVE/NVT/NPT molecular dynamics and batched prediction. The weights (9 MB to 4.8 GB per model) are a GitHub Release on this repository; the package downloads and verifies them on first use.
  • AI4Science/hypde/: HyPDE, a 323-million-parameter 2D foundation model for partial-differential-equation simulations, trained from scratch on thirteen datasets of The Well and FlowBench. From the last six frames of a field it forecasts the next four, and rolls out autoregressively; zero-shot against Walrus it has the lower median VRMSE on 9, 10 and 8 of 13 datasets in one-step, T[1:20] and T[21:60]. The 1.3 GB weights are a GitHub Release on this repository; weights/download.sh fetches and verifies them.

AI4Fun

Creative and game-playing demos (no leaderboard comparison):

  • AI4Fun/reversi/: a AlphaZero-style Reversi (Othello) bot for the Botzone 8×8 arena (C++ pattern/n-tuple net + PUCT-MCTS + exact endgame solver).
  • AI4Fun/music/: a five-part arrangement of 望春風 (Bāng-chhun-hong, 1933, 鄧雨賢).
  • AI4Fun/3d_penguin/: a procedural 3-D QQ-penguin built by a single Blender bpy script.
  • AI4Fun/3d_hunyuan/: a procedural 3-D Tencent Hunyuan (混元) logo orb.

Citation

If you use these results, please cite:

@misc{hyra2026,
  title        = {Hyra: Hunyuan Research Agent},
  author       = {{Hyra Team}},
  year         = {2026},
  howpublished = {\url{https://hy.tencent.com/research/hyra}},
}

License

This repository is licensed under the Apache License, Version 2.0; see LICENSE.