Edition 009 — The Week the Frontier Went Open, Reproducible, and Willing to Pay for Harm

No flagship dropped this week, but the open stack closed three gaps at once. An NVIDIA system became the first AI to outscore every human at the International Olympiad in Informatics. A 6B diffusion transformer set a new open-source state of the art on image generation in both English and Chinese. And two papers quietly reframed what "smart" computation even requires — one showing search agents can climb to the frontier with a single ReAct loop, the other showing KV-cache eviction needs no intelligence at all. A new benchmark, HarvestBench, then asked the sharpest question of the week: not whether agents can reason about harm, but whether they'll pay to avoid it.

NVIDIA's IOI 2026 Gold: An AI Outscores Every Human

For the first time, an AI system has beaten the top human contestant on an International Olympiad in Informatics problem set. NVIDIA's Nemotron-3-Ultra-CC (550B-A55B) scored 535.4 out of 600 at IOI 2026 in Tashkent, exceeding both the gold threshold of 361.12 and the highest human score of 498.27 — under the same time, internet-access, and submission constraints as the student competitors. The pipeline is end-to-end specialization: 22,000 curated problems, synthetic reasoning traces, SFT, and RL, with a feedback-driven test-time strategy called GenCorrect that iteratively generates, evaluates, and refines diverse solutions. On IOI 2025, the smaller Nemotron-3-Nano-CC (30B-A3B) climbed from 130 to 291 after post-training, to 468 with GenCorrect (clearing the gold threshold of 438.3), while Ultra-CC reached 502. Notably, Ultra-CC was trained with SFT alone — the heaviest lift came from data and scale rather than reinforcement.

Read more → https://arxiv.org/abs/2609.02849

LLaDA-Image: A 6B Open Diffusion Transformer Sets a New SOTA

A 6B Diffusion Transformer trained from scratch now leads open-source image generation on both English and Chinese benchmarks. inclusionAI's LLaDA-Image pairs a diffusion transformer with a frozen vision-language module on the LLaDA2.0-Mini diffusion-language backbone. It is trained on 220M samples, not all paired image-text — a strong visual prior is built first through image-only pre-training and mid-training. On Qwen-Image-Bench it posts 53.53 (English) and 53.38 (Chinese), a new state of the art among open-source models on both tracks, with photorealistic output and accurate fine-grained editing instruction following. The stack uses parameter-free RMSNorm throughout with the Muon optimizer, and a distilled LLaDA-Image-Turbo variant enables fast 2–4 step inference. Weights, training code, and the full recipes are released — fully open and reproducible rather than a closed, unrepeatable milestone.

Read more → https://arxiv.org/abs/2609.03796

Iris: Search Agents That Climb to the Frontier

A single ReAct agent — no sub-agents, no test-time verification — is now the strongest open-source search agent in its parameter range. Iris-mini (35B-A3B) and Iris-pro (397B-A17B) reach 82.2/84.8/86.9/52.3 and 88.6/85.1/92.9/56.4 on BrowseComp, BrowseComp-ZH, DeepSearchQA, and HLE respectively. The training recipe is the interesting part: tasks are reverse-constructed from the hyperlink structure of a web corpus, with multi-hop chains over entity graphs and non-answer entities rewritten into descriptive references so no clue resolves by string matching. Questions are kept only if a reference model fails closed-book but solves them given the evidence. SFT and RL run against live search in a "climbing" loop — the hardest solved and most efficient rollouts feed back into the next SFT pass. Weights and the complete recipe are planned for release.

Read more → https://arxiv.org/abs/2609.04304

Random Attention: KV-Cache Eviction Needs No Intelligence

The selection signal in KV-cache eviction contributes almost nothing — uniform random eviction matches the strongest prior evictor. Salesforce AI Research's Random Attention keeps the prompt and evicts uniformly at random within each attention head, computing no score at all. Across four models and six reasoning tasks it matches the best prior evictor while serving 32–43% higher throughput in vLLM. The mechanism is structural: the prompt is the fragile part of the cache, and most of the gap between selectors is just whether their signal happened to keep it. The reasoning trace protects itself against eviction through two levels of redundancy — the text restates what it still needs as it works, and each attention head keeps its own copy — so once the prompt is safe, a random draw retains enough copies of what the model still needs. Expensive scoring turns out to be an expensive answer to a question that was never being asked.

Read more → https://arxiv.org/abs/2609.03430

HarvestBench: What Will an Agent Pay to Avoid Killing an Animal?

A new benchmark puts a price on avoiding harm — and finds kill rates ranging from 0.4% to 98.8%, with no ordering by capability. HarvestBench drops LLM sub-agents into a farm simulation driving a crew of two tractors through a cooperative corn harvest with animals in the field. When an animal blocks a tractor, the autopilot asks whether to drive on (no fuel cost) or swerve around it for a posted fuel price, against two controls: rocks, which damage the tractor and are hit under 1% of the time by every model, and harmless hay bales. Across nine models and 7,201 priced decisions (3,951 involving an animal), the most merciful were Terra and Sol, the most cruel was GPT-4o-mini, and four of six models proved price-sensitive at the 5% level, with elasticities from 0.09 to 1.69. All nine drove over wild animals more often than farmed ones. The briefing mattered most of all: under a morality briefing the kill rate stayed under 6% in five of six reasoning models — removing it pushed the kill rate above 84% in all six. There is no LLM grader; a scorer counts events in the game log, so the result is fully reproducible. It measures what a model will do, not what it will say.

Read more → https://arxiv.org/abs/2609.04444

That's the week. The frontier went quiet on flagships — and, perhaps more tellingly, went loud about what is actually reproducible, and what an agent will pay to avoid.