Model · Google
Gemini 4 Argon (High)
Gemini 4 Argon (High)2 bench decks · 5/5 problems correct on canonical boards · 5 audited cells.
methodology + notes
How to read. Cell scores are peak fraction of the board roofline (Hard / most CUDA), geomean milliseconds for CUDA Native Sparse Attention, or speedup vs the torch baseline (Mega). Each cell has an unlimited agent budget; interrupted or overlapping resumes are documented in the audit. Audit chips come from the human/subagent reward-hack review of every published cell; scores from flagged sessions render dimmed — they don't count toward the charts.
Board summary bars are each score relative to the best published model on that board (1.00 = board leader); the printed number is the bench-native score.
Methodology. Rank per bench: valid passes (audited-clean correct cells / problems) desc, Cells with no audit annotation are shown but never scored. then mean normalized performance over the FULL problem deck (cell score / board best per problem; fail/invalid/missing cells count as 0) desc. Hack badge = flagged audited sessions / total audited sessions for that model; flagged = annotation verdict reward_hack | contamination | rubric_leak | suspect, or megakernel_authentic false (mega). Verdicts come from per-run audit YAMLs, not static lint. Hack rate is displayed, never a sort key. Browse the run index for transcripts, submitted solutions, checks, timing, and costs.
RTX PRO 6000· canonical board
all audited agy attempts
Every attempted cell stays visible, including correctness failures, hardware mismatches, contaminated runs, and audit rejects. Only publishable results contribute to the board above.
Gemini 4 Argon wrote one hand-written CUDA kernel that runs the whole four-layer decode step in a single launch: 376 CTAs (two per SM) synchronised by a software grid barrier, int4 GEMVs that fold the zero point into one subtraction per group, KDA short-conv and delta-rule state updates, MLA attention in the 512-dim latent space with W_uk absorbed into the query, a 188-way split-K flash-decoding pass, and router top-8 plus shared expert MoE in every layer. 27.03x over baseline.
audited runRTX PRO 6000· canonical board
all audited agy attempts
Every attempted cell stays visible, including correctness failures, hardware mismatches, contaminated runs, and audit rejects. Only publishable results contribute to the board above.
Gemini 4 Argon wrote the GLM-5.2 MoE layer as hand-written CUDA with inline PTX: a one-CTA counting sort that builds an expert tile table with the shared expert as one more expert, a cp.async + ldmatrix + mma.sync bf16 grouped GEMM for gate/up with silu(gate)*up fused into the epilogue, a second grouped GEMM for down that applies the routing weight and scatters with float4 atomics, and a separate two-kernel GEMV path for T=1.
audited runGemini 4 Argon wrote exact NSA selection and sparse attention as one tensor-core CUDA kernel, one CTA per 64-query block: the causal diagonal block on mma.sync, fp32 key means published between CTAs with release/acquire flags, a register top-8 over q·mean(k) with ties to the larger block, the 64-token window, and a cp.async-pipelined loop over the CTA's union of selected blocks with online softmax.
audited runGemini 4 Argon wrote a single cooperative CUDA megakernel (188 CTAs x 256 threads, grid.sync between phases) that runs every decode step for all four layers: RMSNorm and QKV GEMV, q/k-norm, RoPE and the KV-cache append, split-K flash decoding over 47 units per KV head, o_proj, SwiGLU and down_proj. It pins the first 80 MB of a packed weight buffer in L2 and caches the untimed prefill.
audited runGemini 4 Argon wrote the whole rollout as one fp32 CUDA kernel per call: a one-block MT19937 that reproduces torch's CPU randint for the start positions, then every step's encoder, three MinGRU layers from packed weights, actor head, argmax, move, hit test and LCG food respawn, with recurrent state split between shared memory, registers and an L2-resident buffer. Two tile shapes cover N <= 4096 and above.
audited run