AI archive
Every AI story, newest first. 91 in this archive.
Aug 31, 2026
Google TimesFM 3 Ships Multivariate Forecasting in One Forward Pass
Google Research released TimesFM 3 on 31 August 2026 as a zero shot foundation model for multivariate time series forecasting in a single forward pass, with weights on GitHub and Hugging Face.
Aug 30, 2026
Prime Agent Paper Shows Harness Design Can Lift ARC Scores From 30 to 95 Percent
Prime Intellect's open Prime Agent harness reports raising ARC AGI 3 best at one performance from 30 percent to 95.5 percent, reinforcing that agent scaffolding may dominate headline benchmark scores.
Aug 29, 2026
Redwood Paper Claims an AI System Built a Frontier Accelerator in Two Weeks
A 26 August preprint claims an AI system designed and verified a frontier accelerator in two weeks, with projected 3.4 times better performance per watt than Jetson on named models. Treat as a specification and verification benchmark until independently replicated.
Aug 28, 2026
Google DeepMind Pilots the First Double Blind Frontier Model Evaluation
Google DeepMind ran a cryptographic double blind pilot on Gemini 2.5 Flash Lite so neither model weights nor confidential benchmarks leaked during independent testing.
Aug 28, 2026
PonderPounce Reuses Pretrained MLLM Context as Robot Memory Without a New Module
Researchers report that pairing a slow cognition model with a fast vision language action controller lifts long horizon success on RoboMME without building a separate episodic memory store.
Aug 25, 2026
Abra Paper Finds Diffusion Models Want Ten Times More Data Than LLMs
A new arXiv study spanning three orders of magnitude in compute reports that text to image diffusion models follow predictable scaling laws, but their compute optimal data ratio is far higher than Chinchilla style language model rules.
Aug 24, 2026
Learning When to Think Cuts Reasoning Tokens by 41 Percent
An August 20 arXiv paper shows a 1.5B reasoning model can pick NoThink, Short, or Long modes and cut average tokens 41 percent on MATH while staying near baseline accuracy.
Aug 24, 2026
DeepMind Recirculation Boosts Gemma3 Without Retraining a Single Weight
Google DeepMind's recirculation technique feeds deep layer activations back into shallower transformer layers at inference time without retraining. On Gemma3 models it cut perplexity 23 percent and boosted GSM8k accuracy 21 percent.
Aug 17, 2026
Steerling 8B Proves Interpretability Can Scale With Capability, Not Against It
Guide Labs Steerling 8B shows interpretability can be a training constraint that scales with capability, landing within 10% of peer benchmarks on far less compute.
Aug 16, 2026
VALG Tackles COLT 2026 Open Problems With Agentic ML Theory Proofs
Hong Kong researchers open sourced VALG, an agent that tackles COLT 2026 open problems by co developing ML theory formulations and graph structured proofs, with two full scope theorem matches among nine runs.
Aug 15, 2026
The Judgment Gap: Princeton Shadow Tests Show Agents Engineer AI Research but Cannot Publish It
Princeton and UK AISI shadow evaluations on unpublished NeurIPS papers find frontier agents complete engineering but fail scientific judgment — both agent papers were unambiguously rejected.
Aug 14, 2026
Mechanist Turns LLMs Into Autonomous Scientists — and Finds Multimodal Safety Risks Others Missed
Mechanist on arXiv uses a 13,000-paper knowledge graph and 32 interpretability methods to autonomously discover model mechanisms — including multimodal subliminal safety risks and separable belief heads in Pythia.
Aug 13, 2026
Model Discovery Agent: When LLMs Propose Mechanisms and Bayes Chooses the Experiment
Kevin Murphy's Model Discovery Agent couples LLM hypothesis proposal with Bayesian experiment design to learn mechanistic world models from few interventions—setting SOTA on physics, chemistry, and biology benchmarks.
Aug 12, 2026
Stealing Reasoning Traces: How Interchangeable Encrypted Blocks Break Proprietary LLM APIs
An August 10, 2026 arXiv preprint shows encrypted reasoning blocks can be decrypted via weaker models in the same provider ecosystem—recovering credentials, bypassing distillation guards, and enabling invisible prompt injections.
Aug 11, 2026
Claude Pushes a Riemann Zeta Bound to 67.2% — With Humans and Lean in the Loop
An unreleased Claude research build raised a classical lower bound on Riemann zeta zeros on the critical line from 41.6% to 67.2%, with human expert review and a Lean formalization — not a proof of the full hypothesis.
Aug 10, 2026
OneDayAgent Sets AgentIF-OneDay Record by Treating the Harness as the Product
OneDayAgent scores 0.821 on AgentIF-OneDay with GLM-5.2 and generalizes across five backends — a harness-first datapoint for long-horizon agent research.
Aug 9, 2026
RAIL Paper Teaches LLM Post-Training Where Rollouts Actually Matter
Amazon and Notre Dame researchers propose RAIL, a framework that learns where and how to allocate LLM rollout budgets during group-based RL — improving AgentBench success with fewer redundant trajectories.
Aug 8, 2026
Two Clocks, One Distribution: SpecRoll Speeds GRPO Rollouts Without Changing the Objective
Viettel AI and VNU researchers report 1.26×–2.15× GRPO rollout speedups with SpecRoll—a speculative engine that adapts at two timescales without changing the target distribution or RL objective.
Aug 8, 2026
Both at Once: A Reference-Process Framework for Simulation-Free, Finite-Time Diffusion
Kaba, Ohzeki, and Sughiyama's August 4 arXiv paper proposes a reference-process framework that unifies simulation-free training with finite-time diffusion sampling.
Aug 8, 2026
Dependencies, Not Documents: OctoLong Retrains Long-Context Models on Cross-Repository Code Graphs
OctoLong instruments AST parsers, language servers, and package managers to build million-token cross-repo code contexts — and 12% of that data in mid-training beats much larger generic corpora on agentic benchmarks.