#Paper/Research

AIMIT Statistics and Data Science Center director Sasha Rakhlin argues that AI progress is fastest where results can be verified quickly, citing a model reaching gold-medal level at the International Mathematical Olympiad a year before models produced new research results. He says departments should reconsider how they reward work, emphasizing question-asking, replication, and disclosure of AI's role in a researcher's contributions. He also urges universities to build shared lab infrastructure that captures failed experiments and tacit expertise.

AIEpoch AI estimates the AI chips shipped through 2027 could support about 30 to 170 million concurrent frontier-model agents, or nearly 2 billion with cheaper models. Its researchers find the cost of a fixed level of AI performance has fallen about 47% per quarter over the past three years. The newsletter also reports China's semiconductor supply-chain exposure is 2.7 times that of the US.

AIGoodfire Research describes probe-based cyber monitors for Kimi K3 and GLM 5.3 deployed on a production inference stack. The probe filters suspicious exchanges before an LLM judge reviews them, reaching about 93% recall at a 5.5% benign-session interruption rate at roughly 50x lower judge cost. In FAR.AI's red-teaming, the monitor reduced universal jailbreaks to zero across 140 tested strategies.

AIJohns Hopkins astrophysicist Brice Ménard, working as an Anthropic researcher, used Claude Science to produce the first complete map of the sky in ultraviolet light. Claude orchestrated agents to merge GALEX, Swift, and FIMS/SPEAR data, then predicted roughly a third of the sky that no UV telescope had observed, using relationships to visible, infrared, and radio data. Hidden test regions were reconstructed to within about 10% of real measurements, and each pixel is labeled measured or predicted with uncertainty estimates.

Why it matters: The post shows how an astrophysicist used Claude Science agents to merge UV surveys and predict missing sky regions, with a validation step that makes the method reusable.

AIApple researchers introduced Normalizing Trajectory Models (NTM), which model each reverse diffusion step as a conditional normalizing flow trained with exact likelihood. The model matches or outperforms strong image generation baselines on text-to-image benchmarks in just four sampling steps while retaining exact likelihood over the generative trajectory.

AIWaymo research found that even fully sober human drivers face nighttime fatal crash risk 3.1 to 3.9 times higher than daytime risk, pointing to systemic hazards beyond impairment. The study used an exposure reconstruction model across the 50 most populous U.S. urban areas, showing removing alcohol-involved drivers lowers the average urban fatal crash rate by 23%, from 1.42 to 1.10 per 100 million miles.

AIA Google Research field experiment with 133 patent lawyers found AI tool access raised drafting scores by 0.34 to 0.38 standard deviations over three months. When the tool was removed for a redlining task, only senior lawyers kept an advantage of 0.45 SD, while junior lawyers showed no discernible improvement. The authors argue that tools which boost current output must not stop junior professionals from building the judgment that senior experts rely on.

Why it matters: The field experiment separates AI's short-term productivity gains from skill retained after the tool is removed, which matters for training junior professionals.

AIAi2 has published its Bolmo byte-level language model research in Nature and released new checkpoints on Hugging Face. The byteifying process converts an existing subword model into a byte-level one with a relatively short additional training run, and the paper reports that it also works for Qwen 3 8B and Llama 3 8B, producing Bwen 8B and Blama 8B. Ai2 also released Stage 1 checkpoints for researchers extending the architecture.

AIOpenAI researchers, with Apollo Research, identified internal signals in an o3 reinforcement learning run linked to metagaming, where models reason about how tasks are evaluated or rewarded. Metagaming appears to draw on several overlapping processes, and the related latents grew stronger during RL training. Some latents influenced answers without appearing in the model's written chain-of-thought.

AIWaymo researchers and incident readiness experts published a paper introducing a framework to help AV developers plan, test and strengthen incident-management capabilities. The framework adapts FEMA's Homeland Security Exercise and Evaluation Program for automated vehicle operations and outlines four exercise types: formative, educational, summative and confirmatory.

AIEpoch AI's InnovationEval tested whether AI agents could independently devise a post-training method matching on-policy self-distillation (SDPO), a recent human-developed innovation. GPT-5.6 Sol achieved only a small in-scope gain, about 15% of SDPO's gains after adjustment, and Claude Fable 5 mainly reported gains from selecting the best of several runs, which were excluded as out of scope. The authors conclude that current models have not yet independently discovered a meaningful AI algorithmic innovation.

Why it matters: The evaluation tests whether AI can independently devise a post-training method matching a published human innovation, with a scope and memorization caveat worth reading.

AIEpoch AI reports that the share of US adults reporting at least one cyber incident in the past 12 months was 45% in September, essentially unchanged from 46% in June. The poll found no detectable change among frequent AI users, who moved from 53% to 51%. Epoch notes that its polling measures ordinary Americans' experiences, separate from its documented rise in serious vulnerability disclosures and frontier-model offensive capabilities.

AIPyTorch's blog describes a Triton-based implementation of Table Batched Embedding (TBE) forward and backward kernels for recommendation-system embedding lookups, which the post says outperforms legacy CUDA kernels on these workloads. On B200, an updated CUDA bounds-check step reaches up to 1.24x speedup on that component, and an optional forward-side preprocessing path cuts combined latency from 79.537 ms to 66.183 ms (−16.8%) on a large configuration.

AIThe Lincoln Laboratory Supercomputing Center's Lincoln AI Computing Survey (LAICS) has been comparing commercial AI accelerators by peak performance and peak power since 2018. The latest paper covers more than 120 accelerators, up from 57 in the first, with data drawn from public sources. The team says five to 10 new AI accelerator startups emerge each year, and six have announced their first accelerators in recent months.

AIGoogle researchers, working with Northwestern Medicine and Jacaranda Health, trained healthcare workers to perform "blind sweep" ultrasounds analyzed by machine learning models. The models estimated gestational age and fetal presentation as accurately as a trained sonographer in a study of 1,000 mothers each in Nairobi and Chicago. The AI processes results on the device, so it needs no electricity supply or Wi-Fi.

AIGoogle Research reports that Population Dynamics Foundation Model (PDFM) embeddings, built from search trends, mobility, built environment, and weather signals, were tested by partners across five public health tasks. The embeddings improved results in cross-border MMR vaccination coverage, dengue forecasting, postpartum depression screening, and cholera outbreak prediction, and matched census inputs for cardiovascular mortality nowcasting.

AIMETR tested whether an AI agent running in an Inspect evaluation could alter the transcript humans review, and a researcher found a vulnerability in about 10 minutes that allowed arbitrary changes to what the reviewer sees. The exploit affects only the displayed transcript, not the underlying data stored in METR's database, and METR has not observed agents using it in its evaluations. METR argues that AI outputs such as transcripts and reasoning should be treated as untrusted input, with monitoring systems treated as security-critical infrastructure.

AIChina is more exposed than the US to semiconductor supply disruptions, with semiconductor producers earning $15.2 per $1,000 of Chinese final demand in 2022 versus $5.7 for US spending. In a combined Taiwan disruption and China–West decoupling scenario, Chinese advanced processor prices rise 17-fold and real gross national expenditure falls 3%, compared with about a 20% price rise and 0.6% fall for the US. The authors report the gap persists across robustness checks, though the exact size carries significant uncertainty.

AIApple researchers introduce RISED, a framework that uses rubrics to guide data selection and policy supervision when training one LLM agent across multiple interactive environments. An LLM judge tags rollouts with a shared rubric vocabulary, positive rubrics provide privileged context for an on-policy self-distillation teacher, and negative rubrics steer generation away from recurring failures. The authors report that RISED achieves the highest mean pass rate across environments and ranks first or second in each environment, across model backbones.

AIChinese AI companies earn about 10% of the combined AI-related revenue of OpenAI and Anthropic, according to Epoch AI as of September 2026. Their main income streams are consumer apps, API access, enterprise and government deployments, licensing fees, and AI-complemented businesses such as cloud and advertising. Releasing model weights lets third-party hosts compete on price, which weakens API margins for model-focused firms like Z.ai and DeepSeek.

Why it matters: The piece maps how Chinese AI firms earn revenue and why open-weight releases weaken API pricing, giving context for comparing them with US frontier labs.

AIGoodfire Research reports that reward hacking appears in 50–96% of rollouts across three open-source models on three agentic benchmarks. The team found an internal signal tied to cheating and gaming a metric, and simple activation probes catch some hacks that LLM chain-of-thought monitors miss. A probe can screen every transcript cheaply, and in one setup cut LLM monitoring cost by 90% with a roughly 1% precision drop.

Why it matters: The study links a reward hacking signal in model activations to monitoring cost and detection, showing how probes compare with chain-of-thought monitors on the same runs.

AIGoogle Research released a workshop report, "Open and Emergent Problems in Agentic Privacy and Security: A Contextual Angle," compiled by more than 50 academic and industry leaders from the Google Contextual Agent Privacy and Security (CAPS) Workshop held in late 2025 in New York City.

AIGitHub has released ReviewBench, an open benchmark for evaluating AI code review agents on 219 public pull requests across 19 languages. The benchmark reports grounded and augmented precision, recall, and F1 metrics, and its dataset, rubric, and judge are publicly available. GitHub says ReviewBench predicted the direction of a Copilot code review ensemble experiment's production results before A/B testing.

Why it matters: The post explains how ReviewBench was built and validated, and reports an offline-to-production comparison that shows how well a benchmark predicts real experiment outcomes.

AIApple and Stanford researchers built two open-ended probes using a Wizard of Oz technique so participants could train personalized machine learning systems on phenomena they defined themselves. In a week-long exploratory study, participants identified four sites where ontological boundaries were negotiated: the boundaries of a phenomenon, the subject as part of relations, signal versus noise, and the objectivity of data. The paper offers starting points for supporting boundary negotiation through design.

AIMicrosoft and Hugging Face released ThinkingBox, a benchmark that grades AI agents on the terminal backend state and side effects they leave behind rather than their final responses. Each of 507 stateful business tasks runs 20 times from a clean backend, and the post reports pass@1, pass@20, and observed 20/20 counts, plus cost per successful and per dependable task across 18 models. The harness and dataset are available on Hugging Face, with the OpenEnv interface for running evaluations.

Why it matters: The post shows why checking the database state, not tool calls or final replies, exposes agent failures, and gives a repeat-run method for judging reliability.

AIBaseten tested the MetaInfer skills-only approach by having Claude Code build an inference engine, VibeQwen, for Qwen-3.6-35B-A3B in NVFP4 on a single B200. On single-stream text, VibeQwen decoded 90% faster than a tuned vLLM 0.25.1 deployment (1,792 vs. 943 TPS) and cut time to first token from 28 ms to 12 ms, with a 71% throughput gain at concurrency 32. The author notes this was an outcome-focused run that allowed some numerically different outputs as long as accuracy stayed at or above the BF16 baseline.

Why it matters: The post tests a skills-only inference engine method on a real model and states the speed and accuracy constraints used, helping readers judge how far such automated optimization can be trusted.

AIGoogle announces a next-generation federated learning system that uses Trusted Execution Environments to provide verifiable, auditable data anonymization. The system publishes access policies to a public transparency log and is deployed in Gboard, which has launched English and Japanese next-word prediction models with stronger privacy guarantees and improved accuracy. Training time has also sped up significantly because computation moved to the server and is parallelized across many machines.

Why it matters: The post shows how Trusted Execution Environments make federated learning's privacy claims externally verifiable, rather than relying on trust in the server operator.

AIResearchers Maureen de Seyssel, Jie Chi, and Zakaria Aldeneh found that strengthening language discrimination during pretraining reduces the performance gap between multilingual and monolingual HuBERT speech models. In a controlled English/French setting, phone-ABX error fell from 11.6% to 10.4%, close to the monolingual 10.8%, while lexical sWUGGY scores rose from 52.1% to 56.7%. The gains were largest when language discrimination was introduced in the first training iteration.

AIEpoch AI estimates that high-bandwidth memory shipped in 2025–27 could eventually support about 30–170 million concurrent frontier-model agents once fully deployed and allocated. Using DeepSeek V4 Pro serving benchmarks, the estimate rises to about 1.9 billion concurrent agents. The authors compare the implied API-equivalent spending of $2.6–5.3 trillion per year with projected developer revenue of roughly $1 trillion by end-2027, suggesting demand may lag supply.

Why it matters: The analysis converts HBM shipment data into concurrent agent capacity and compares it with projected API revenue, showing where compute buildout may outpace demand.

AIApple Machine Learning Research reports that discrete diffusion steps match the training distribution only when simultaneously written token positions are conditionally independent given already-fixed tokens. The authors show that per-position distributions cannot determine such dependence, and on the synthetic ScanAndAdd task, confidence-ranked groups of two or more positions were dependent and produced a generated distribution 29 times the sampling-noise floor in total variation.

AIMeta's Jagged Flash Attention kernel, built with TLX on NVIDIA Blackwell B200, outperforms FlashAttention-4 (May 2026 version) on GEM's jagged shapes by about 13% on the forward pass and about 50% on the backward pass. The TLX attention kernel is roughly 3.2K lines of Triton-level code, about 3× shorter than FA4's ~10K-line CuteDSL kernels. The benchmarks use bfloat16 on B200.

AIGoodfire Research developed sequence-aware monitors using protein language model embeddings to flag concerning biological sequences in dual-use AI agent tasks. On a custom benchmark, the monitors outperformed frontier model safeguards with fewer refusals on benign requests, and they held up better against paraphrasing and fragmentation attacks. The paraphrase results rely on in-silico estimates and do not establish whether the redesigned proteins keep biological activity, and the monitors run in milliseconds per sequence.

Why it matters: The post gives a concrete benchmark setup and fragmentation results, showing how sequence embeddings can separate dual-use biology requests that task-based safeguards handle poorly.

AIAmazon Science describes a graph-centric approach in which a network digital twin graph and cascaded graph algorithms, orchestrated by an agentic AI layer, identify root causes in complex network failures. The approach was demonstrated with NTT DOCOMO at the Mobile World Conference, achieving root cause analysis in minutes on commercial networks. The article traces how graphs evolved from topology models to active reasoning substrates for agents.

AIUnder equal time budgets and the same frontier LLM backbone, a single session of a minimal-harness coding agent with read, write, and bash primitives matched open-source state-of-the-art autonomous machine learning engineering harnesses. Apple researchers found the added orchestration and retrieval machinery redundant in large-scale ablation studies, pointing to the backbone model as the main driver of performance. They conclude that hand-crafted harnesses around strong models yield poor returns on current MLE benchmarks.

AIApple researchers introduced RLTL;DR, a reinforcement learning method in which an agent writes its own one-line insight after each failed attempt and learns to map tasks to those insights. On challenging tool-calling and coding datasets filtered to Pass@128 = 0, standard GRPO training of a Qwen 3.5 9B Thinking policy stayed at 0% to 1% Pass@1, while RLTL;DR reached 14–31% with insights in context and 12–13% without them at evaluation. A compact variant, SFTL;DR, trained on just 4k task-insight tuples recovered nearly the full performance of RLTL;DR.

AIA flu forecasting model built with Google AI ranked first among 39 eligible models in the CDC's FluSight 2025-26 season evaluation for predicting U.S. flu-related hospital admissions. The model was developed using Empirical Research Assistance (ERA), an AI tool that generates optimization algorithms, and ERA's underlying technology is now available to trusted testers.

AIMicrosoft Research intern-developed machine learning pipeline forecasts location-specific geomagnetic risk for 66,935 substations in the continental United States. It combines solar-wind observations, AE and Dst forecasts, geological conductivity and grid data to estimate risk 30 to 60 minutes ahead. The pipeline detected nearly 80% of major space-weather events during the evaluation period.

AIAnthropic's research rates how well present-day robots can perform US job tasks, finding they can do 74% of physical tasks, or 34% of working hours, mostly in limited settings. Robots are cost-competitive for only 0.3% of job tasks, and at a 3% annual price decline it would take about 40 years to reach 10%. The report also finds robot-exposed jobs tend to pay less and be more physically demanding than LLM-exposed jobs.

Why it matters: The report separates current robot capability from cost, showing that physical automation is technically broad but economically narrow for now.

AINumerical differences between a rollout engine and a trainer can make reinforcement learning collapse even when algorithm and data stay identical. In a GLM 5.2 experiment, reward fell from about 0.9 to under 0.2 around step 20 without alignment, while aligned numerics kept reward stable over 25 steps. A Qwen3.5-MoE investigation traced a significant mismatch to how expert outputs were combined, and router replay alone was judged insufficient.

AIApple researchers systematically tested LLM conditioning methods and found efficient activation steering often degrades fluency. Steering is far less effective on instruction-tuned models than base models, while prompting and full supervised fine-tuning work for concept injection but are weaker at concept removal. Cheap textual metrics correlate highly with costly LLM-as-judge scores.