From overseas AI news on September 12, 2026. Mathematics first: a proof-generation system trained from Nemotron 3 Ultra scored 30 points at the 2026 International Mathematical Olympiad (IMO) — the gold medal threshold — and has released its entire recipe. On the applications side, the U.S. Princeton Plasma Physics Laboratory announced an AI framework that entrusts safety-critical control of fusion systems to millisecond-scale decisions. Rounding out the lineup: measurement notes from training an image generation model entirely on a single GPU, research on compute allocation in reinforcement learning and memory management for agents, and, from the Japanese-language community, articles quantifying agent self-correction and failure.
Math Proof System Built on Nemotron Scores 30 Points at IMO 2026, Reaching the Gold Medal Threshold — Full Recipe Released
Drawing attention on arXiv is "An Open Recipe for IMO Gold: Training Nemotron for Olympiad Mathematics." Starting from NVIDIA's open model Nemotron 3 Ultra, the work trains two math-specialized checkpoints through supervised fine-tuning (SFT) and reinforcement learning (RL), and systematically evaluates the design of checkpoint selection, verification, and refinement. The finished system runs a test-time compute pipeline in which three checkpoints — including publicly available models — run a search that iterates proof generation, verification, and refinement, while a separate high-compute stage picks the final submission. It writes every proof in natural language, using no formal verifier, no external tools, and no internet access.
The system reportedly scored 30 out of a possible 42 points at IMO 2026, reaching the gold medal threshold. What's notable is not just the score but how "open" it all is. In addition to the two post-training checkpoints, the team released the training data, training and inference code, the solutions actually submitted, and even "Nemotron-IMO-Bench," a new 200-problem benchmark that does not overlap with existing problems. In contrast to reports — covered here recently — that GPT-6 Astra deliberately deprioritized strengthening its math capabilities, open models continue to push the frontier of proof generation.
Millisecond-Scale Control Decisions for Fusion Systems — PPPL's "Pacman" AI Framework
The U.S. Princeton Plasma Physics Laboratory (PPPL) has announced [an AI framework that controls fusion systems safely](https://www.pppl.gov/news/2026/pacman-ai-framework-controlling-fusion-systems-safely-makes-key-decisions-milliseconds), which was featured on Hacker News. According to the announcement's title, the framework, named "Pacman," makes key decisions in fusion system operation on millisecond timescales.
In fusion plasma control, responding to signs of instability is demanded on timescales beyond human reaction speed, so deciding fast — and on the safe side — has been a longstanding challenge. Even when handing judgment over to AI, in a world where a control mistake leads directly to hardware damage, it matters that "safely" leads the title. As of this collection we haven't been able to confirm the technical details, so we'll wait for follow-up reports on the substance of the method.
Training a 210M-Parameter Image Generation Model on a Single GPU — Three Fresh Measurements
On Reddit's MachineLearning community, attention is on [measurement notes](https://huggingface.co/blog/ivanmikhnenkov/tinydit-text-to-image-from-scratch-one-gpu) from training a 210M-parameter diffusion transformer (DiT) for text-to-image generation from scratch over 3.5 days on a single RTX PRO 6000. It's a personal experiment — 4.2 million 256px images, with the goal of understanding the whole recipe — and the author shares three measurements that, in their words, are "not clearly stated elsewhere."
The first is where attention "sink" mass ends up. After adding register tokens and learned spare key/value slots, as training progressed these slots in intermediate blocks gathered roughly 90% of the cross-attention mass, while the EOS token — traditionally prone to becoming the sink — dropped to around 4%. The second is that the flow-matching loss is a "health signal," not a "quality signal": the loss barely moved over the course of training (0.805 → 0.754), while validation FID improved markedly from 33.7 to 27.0 and object-detection-based accuracy rose from 65% to 90%. The third is an inference-time timestep shift: keeping 20 sampling steps and applying a shift of 2.8 gave better quality than raising the count to 50 steps. [Code, weights, and a demo](https://github.com/ivanmikhnenkov/tinydit) are public as well.
Pouring RL Compute in Right Before "Belief Changes" — belief-shift branching
In reinforcement learning with verifiable rewards (RLVR), a known technique forks a chain of thought mid-stream to create sibling rollouts and estimates each step's value from how their outcomes diverge. Place the fork badly, though, and the siblings arrive at nearly the same conclusion, yielding almost no credit signal. "Belief-shift branching," posted to arXiv, proposes a placement that reads the model's own "belief" in its answer at candidate boundaries and puts the fork immediately before the step where consecutive beliefs diverge the most.
Three implementations are provided, none requiring step-level supervision: a black-box probe, a depth profile via the logit lens, and an activation direction learned offline. The probe's added cost reportedly stays at about 1% of per-step compute on math and under 5% on code. Compared against Monte Carlo value curves, it beat existing entropy- and structure-based metrics across all eight model × benchmark settings, and in actual RL it showed consistent improvements across three model families — reportedly +2.6 on aggregate math scores for OLMo-3-7B and reaching +6.5 on LiveCodeBench-medium.
Curating an Agent's Persistent Memory Against Its Environment — environment-probing curation
Persistent memory that lets long-running AI agents accumulate experience is entering practical use, but managing it is hard. A curator agent that organizes memory by looking only at completed task trajectories can save past errors as correct knowledge, over-generalize from partial evidence, or keep holding on to stale knowledge. Research posted to arXiv proposes "environment-probing curation": letting such curators inspect the real environment with read-only, least-privilege tools. No retraining is required, and the task agent's configuration stays unchanged.
Evaluated in a production-like harness built with the GitHub Copilot SDK, the approach reportedly improved accuracy and efficiency at the same time on the database exploration benchmark CLBench: pass rate from 39% to 73%, queries per task from 8.8 to 4.7, and task agent cost from $3.38 to $1.68. On APEX, a management consulting task suite, it beat the baseline in all six worlds while cutting tool calls by 16–75%. Taking inventory of memory against the environment rather than from "recollection" is a practical touch, and this one is worth watching as a reliability measure for agent operations.
From the Japanese-Language Community — Quantifying Agent "Self-Correction" and "Failure"
On Zenn, an article analyzes 1,239 agent Edit calls extracted from 389 session logs on hand. Only 12 (1.0%) failed because the replacement target could not be found — but 50.0% of successful replacements were rewrites of strings the agent itself had written earlier in the same session. What's more, the median gap between writing and fixing was 16 steps, with only 3.4% of fixes coming immediately after. A high success rate and writing it right the first time are different things — a good lesson for the design of metrics that measure agent work quality.
A Qiita article by the same author examines tool calls across 9,906 session logs and classifies the 283 (2.9%) that carried a failure flag. Genuine failures came to just 34.3%, with terminations due to permission-related governance at 33.6% and cases that produced output but exited non-zero at 21.2%. Two-thirds of what "looks like failure" is a different kind of event altogether — and how you read the logs could considerably change the outlook on running agents.