Hyra-results — Hunyuan Research Agent (results bundle)

ListenAgent vs Agent: Hunyuan's Hyra and the Research-Agent Bake-Off10:36

Show notes

Tencent Hunyuan's new autonomous research agent, **Hyra**, ships a results bundle with eleven head-to-head wins — not against humans, but against *other* research agents like TTT-Discover, SimpleTES, and the Recursive harness. We dig into what it actually beat, from nanochat and nanoGPT-speedrun to 100 broken packing records, a 15-parameter adder, and an approved-drug baseline for PARP1 docking. The gains are mostly thin margins on public leaderboards — so we ask what "breadth across scopes" really buys, and what the artifact-first release means for anyone building generative systems. **Sources** - Hyra-results — Hunyuan Research Agent (results bundle)

Subscribe to the channel: https://podcast.lumalabs.link/27a25f51ea81d41a8253f46c6133a35e/feed.xml

TL;DR

Companion GitHub repo for Hyra (Hunyuan Research Agent), Tencent Hunyuan’s newly-launched autonomous research agent. The repo publishes 11 head-to-head comparisons where Hyra beats the previously-published best on public tasks spanning AI4AI (nanochat val BPB, nanoGPT-speedrun wall-clock, SOL-ExecBench), AI4Science (Erdős / autocorrelation constants, PARP1 docking, qubit routing, smallest transformer for 10-digit addition, sunspot symbolic forecasting, and 100 shape-in-shape packings from Erich Friedman’s Packing Center), and AI4Fun (Reversi bot, procedural 3D, music arrangement). The prior-best comparisons cite other AI-scientist systems as baselines — TTT-Discover, SimpleTES, and the Recursive “First Steps Toward Automated AI Research” harness — making Hyra the first filed head-to-head between research agents rather than research-agent-vs-human. The companion launch post at hy.tencent.com/research/hyra could not be read at filing time; this repo is the load-bearing artifact.

Key claims

The repo ships one folder per task containing the final solution artifact and (where relevant) a self-contained script that reproduces it; the results table compares Hyra to the best previously-published number per task with arrows marking direction and the Hyra value bolded [Results table].
AI4AI: Hyra improves nanochat val BPB 0.9109 → 0.9015, nanoGPT-speedrun wall-clock 77.5 s → 76.4 s, SOL-ExecBench score 0.754 → 0.771, all against Recursive “First Steps Toward Automated AI Research” baselines [AI4AI rows, note e].
AI4Science autocorrelation & Erdős constants: C₁ 1.502870 → 1.502850 (vs TTT-Discover / arXiv:2601.16175); autocorrelation ratio R 0.962694 → 0.962901 (vs SimpleTES / arXiv:2604.19341); Erdős min-overlap C₅ 0.380868 → 0.380859; sums-vs-diffs C(A) 1.14489 → 1.15971 (both vs SimpleTES) [AI4Science rows, notes a–b].
On Erich Friedman’s Packing Center, Hyra broke 100 records across 28 shape-in-shape families, each credited on the site as “Found by Haowei Lin” and each beating the previously listed best [packing_records row, note f].
On the AdderBoard leaderboard task (smallest transformer that adds two 10-digit integers at ≥ 0.99 accuracy), Hyra reduces the unique trainable parameter count from 36 → 15, a >2× compression [smallest_adder row, note c].
On PARP1 docking (Vina score + 10·(1 − QED); lower is better), Hyra achieves −10.60 vs the Olaparib approved-drug baseline of −9.77 [parp1_docking row, note d].
On qubit routing, Hyra adds 258,369 CNOTs via SWAP routing vs a prior 269,037 (SimpleTES baseline) — a ~4.0% reduction on the same benchmark [qubit_routing row].
On sunspot symbolic forecasting, Hyra achieves a rolling-origin free-running 24-month R² of 0.78 on a fully-held-out half-century segment vs 0.47 for a “copy last frame” persistence baseline [sunspot_symbolic row, note g].
The repo explicitly warns that several of these targets are continuously-updated public leaderboards and that later entries may “warm-start from Hyra’s published solutions” to reach still-better numbers — an acknowledgment that the artifact bundle itself becomes seed material for future rounds [Note].
AI4Fun (non-leaderboard demos): (a) an AlphaZero-style Reversi bot for the Botzone 8×8 arena (C++ pattern/n-tuple net + PUCT-MCTS + exact endgame solver), (b) a five-part arrangement of the 1933 Taiwanese folk song 望春風 (Bāng-chhun-hong, 鄧雨賢), (c) a procedural 3-D QQ-penguin from a single Blender bpy script, and (d) a procedural 3-D Tencent Hunyuan (混元) logo orb [AI4Fun section].
Results are dated 2026-07-10; the repo is released under Apache 2.0 [Citation, License].

Method

The Hyra-results repo is a bundle of concrete outputs, not a system description — the launch blog at hy.tencent.com/research/hyra is the intended methodology reference but could not be read at filing time. What can be extracted from the artifact side of the release: each task-folder contains the final solution artifact in the native downstream format for that problem (a set of transformer weights for AdderBoard, a symbolic formula for sunspots, a packing geometry for the Friedman challenges, a docked molecule for PARP1, a compiled C++ Reversi engine for Botzone, a Blender bpy script for procedural 3D, a five-part score for the music arrangement), so the “artifact” Hyra emits is not restricted to Python training scripts and spans multiple downstream toolchains within a single release.

The prior-best baselines Hyra compares against are themselves AI-scientist / test-time-discovery systems — TTT-Discover (Learning to Discover at Test Time, arXiv:2601.16175), SimpleTES (Evaluation-driven Scaling for Scientific Discovery, arXiv:2604.19341), and the Recursive “First Steps Toward Automated AI Research” harness (nanochat autoresearch, nanoGPT-speedrun, SOL-ExecBench baselines). This means Hyra’s headline claim is not “an agent can attempt these problems” but “an agent can out-perform other agents on them at the same evaluation harness.”

Results

Headline table (Hyra vs prior best; ↓ lower better, ↑ higher better):

AI4AI: nanochat val BPB 0.9015 vs 0.9109 ↓; nanoGPT-speedrun 76.4 s vs 77.5 s ↓; SOL-ExecBench 0.771 vs 0.754 ↑ [AI4AI rows].
AI4Science analytic constants: C₁ 1.502850 vs 1.502870 ↓; R 0.962901 vs 0.962694 ↑; C₅ 0.380859 vs 0.380868 ↓; C(A) 1.15971 vs 1.14489 ↑ [AI4Science rows].
AI4Science engineering / discovery: Packing records broken 100 (across 28 shape-in-shape families) ↑; AdderBoard params 15 vs 36 ↓; PARP1 objective −10.60 vs Olaparib −9.77 ↓; qubit routing extra CNOTs 258,369 vs 269,037 ↓; sunspot symbolic R² 0.78 vs 0.47 persistence ↑ [AI4Science rows].

Every leaderboarded task Hyra reports on shows Hyra ahead of the cited prior-best number. The repo does not yet publish an aggregate “hits per compute” scaling plot (as ASI-ARCH does) or an expert-graded per-statement accuracy (as Kosmos does); the empirical claim rests on the breadth of tasks on which the same agent brand improves prior best.

Why it’s interesting

Hyra is the sixth AI-for-AI-research instance filed on the wiki and the first from a major Chinese lab, extending the AI-for-AI Research cluster that so far contained FARS (FARS: Fully Automated Research System), ASI-ARCH (AlphaGo Moment for Model Architecture Discovery), Karpathy autoresearch (Scaling Karpathy's Autoresearch: What Happens When the Agent Gets a GPU Cluster), Kosmos (Kosmos: An AI Scientist for Autonomous Discovery), and Weco’s AIDE² (AIDE²: First Evidence of Recursive Self-Improvement). It differs from all five along two axes at once: (a) the downstream-artifact-target range spans compiled C++, procedural 3D, symbolic formulas, molecular candidates, packing geometries, transformer weights, and music notation within a single launch — the widest filed set of downstream targets for one research-agent brand — and (b) the baselines Hyra beats are themselves AI-scientist systems (TTT-Discover, SimpleTES, Recursive), making Hyra the first head-to-head between research agents rather than research-agent-vs-human. Together, these two axes sharpen the concept page’s open question about whether the same recipe applied at different scopes really is the same recipe: Hyra runs across scopes under one brand.

The 100-packings-records claim is also the most concrete filed instance of the pattern ASI-ARCH called “winners lean on past experience + reasoning; losers lean on novelty” — 100 slightly-better geometries produced by an agent that appears to have systematically mined the entire Packing Center corpus, rather than one clever new packing. And the PARP1 result — beating an FDA-approved inhibitor (Olaparib) on the same composite docking-plus-QED objective — is the first filed AI-scientist output on the wiki that competes directly with an approved-drug baseline on a drug-discovery objective, complementing the AI-for-AI Research cluster’s prior focus on ML systems and formal math.