A new arXiv paper observes that while uniform discrete flow allows repeated updates at every generation position, that same continued revision can overwrite correct intermediate predictions. A Sudoku experiment found 9.4% of generated cells were correct mid-trajectory but wrong in the final output. The authors propose LEDFlow, a training-free sampler that introduces generation order via selective absorption, fixing chosen predictions while preserving uniform-flow velocity at still-active positions, and ordering absorption by local entropy. LEDFlow reaches 0.845 Nikoli Sudoku solve accuracy and improves text-to-image and multimodal understanding results at inference cost comparable to standard flow sampling.

A new arXiv paper evaluates federated learning with Sherpa.ai's Blind Vertical FL protocol to pair a hybrid quantum-classical active party with a classical passive party, reporting accuracy gains from 0.7227 to 0.8757 on a custom parity benchmark while using far fewer trainable parameters than classical baselines.

A paper accepted to COLM 2026 and KONVENS 2026 workshops finds that the chat template acts like a switch, boosting disclaimer-style self-reports and suppressing experiential ones across 8 open-source instruct models up to 9B parameters, and identifies an activation-space direction that can reproduce or suppress the effect.

Researchers introduce IntLawNER, a named entity recognition dataset and benchmark for codified international law, with 2,987 gold-annotated sentences and 8,094 entity spans from ICJ decisions, UN Security Council resolutions, and ECtHR judgments. The paper also exposes pitfalls in silver-to-gold evaluation and benchmarks models including GLiNER and several LLMs.

A new arXiv paper shows that LLM judge consensus is less reliable than it appears because judges make correlated errors. In a ten-judge panel, the average pairwise error correlation was 0.21, making the panel roughly equivalent to 3.5 independent judges; ignoring shared errors changed significance in up to 28% of comparisons.

A new arXiv paper adapts a three-stage human deliberation paradigm to large language models from three different families and tests it across four domains of increasing real-world stakes. Deliberation reduced collective error beyond passive aggregation of independent answers, and post-deliberation individual judgments retained that collective gain—but the advantage required model diversity, as groups composed of clones from a single model did not benefit.

The paper separates replication, measurement sensitivity, and persistence in behavioral evaluations of hosted language models. Using Regent Chess, it finds that a previously reported Gemini 3.1 Flash-Lite deficit replicates on fresh games under its historical configuration, but rebuilding the evaluation-and-inference configuration under the same public identifier lowers the endpoint and reverses a 4K comparison between Gemini 3.1 and Gemini 3.7. The authors argue hosted-model claims should be indexed by tested identifier, serving period, instrument, and inference configuration.

This paper proposes a goal-driven approach to categorizing process variants that reverses the usual workflow. Instead of clustering variants by structural similarity and then manually assigning business meaning, analysts first author an organization's goal model, which predefines the categorization axis. Each variant is converted into a textual narrative of its behavior, and an LLM interprets that narrative against the goal model to assign the variant to a category. The approach was implemented end-to-end and evaluated on three public logs of widely differing scale and behavioral diversity; goal-model guidance produced partitions that differ from unguided induction and that respond to controlled edits of the declared alternatives, at the cost of authoring a goal model.

A population-supervised framework infers latent biophysical quantities from single 2D red-blood-cell images and aggregates them into mean corpuscular volume, red-cell distribution width, and mean corpuscular haemoglobin, using shared local inference, a structured decoder, learned instance weighting, and device calibration. Evaluated on 390 specimens and 1,105 acquisitions across six devices, it reports Pearson correlations of 0.86–0.98 against a Sysmex analyser. The authors also formalise identifiability limits: population agreement alone does not identify single-cell properties or 3D geometry.

A new arXiv paper transplants the Music Lab social-influence design into a market for academic attention. In one experiment, 1,000 AI agents chose among all 114 regular research articles published in the American Economic Review in 2025; agents who could see earlier choices in their community picked 17.2 percent fewer papers each, concentrated their selections more heavily, and collectively covered only 73 papers versus 90 under independent choice. A second experiment found that randomly assigning papers five initial selections lifted their subsequent selection rate by 45.55 percentage points.

PAANI is an on-device perception-to-guidance architecture for river monitoring robots, combining YOLO11n detection, MobileNetV3 Small segmentation, and explainable corridor policies on an Arduino UNO Q. It achieves strong accuracy but distinguishes model performance from validated on-water collision avoidance.

A new arXiv paper presents an $88, fully offline AI-integrated smart cane built on a Raspberry Pi Zero 2W that fuses RGB vision with Time-of-Flight ranging and delivers distance-aware vibrotactile and audio alerts. Indoor tests show a macro-averaged F1 of 0.82, roughly 330 ms end-to-end latency and 2.8 W peak power, with a 12-participant usability study reporting a SUS score of 78.5.

A study accepted for oral presentation at NLPCC 2026 uses token-matched experiments to examine how didactic data (textbooks) and clinical data (patient records) differently shape medical large language models. It finds asymmetric transfer: clinical data improves clinic-oriented tasks while staying competitive on knowledge-intensive ones, whereas didactic data mainly helps knowledge-intensive tasks. Error analysis points to a knowing-doing gap, where better knowledge recall does not reliably translate into clinical reasoning.

SpeakON has released a 25 g MagSafe button that carries its own microphone and battery and writes finished text straight into any iPhone text field via a system-wide iOS keyboard extension. It works while locked or offline, ships with Smart Polish, Smart List, Style, Translation and other text-shaping features, and costs $129 one time in the US with Pro Lifetime included.

Simon Willison and Jesse Vincent are hosting an evening birds-of-a-feather gathering in San Francisco on Wednesday, October 14th for people building strange and interesting things with and on top of coding agents. Framed as an "agentic show-and-tell," the event favors early-stage experiments, unfinished projects and privately held work over product pitches.

NVIDIA AI Day Singapore takes place Sept. 22-23 at the Raffles City Convention Centre, offering hands-on training, expert-led sessions and advanced tools. At the event, NVIDIA and its partners announced progress across Southeast Asia in public sector AI, region-specialized Nemotron open models, and the first ASEAN adoption of the NVIDIA Vera Rubin platform by Sea Limited.

OpenAI is rolling out ChatGPT Ads across Indonesia, Malaysia, the Philippines, Singapore, Thailand, Vietnam, and Taiwan, bringing the product to more than 60 countries. Ads appear only for Free and Go users, while Plus, Pro, and Enterprise stay ad-free.

Kaiku is a task tracker pitched as already usable by AI agents, with a Product Hunt listing that describes it as “The task tracker your AI agents already know how to use.”

OpenAI announced on September 23, 2026 that, under a new agreement, Airbnb is giving its engineering and product development teams broader access to OpenAI frontier models, including GPT-6 Astra. The expansion builds on Airbnb's use of Codex and makes models available through OpenAI APIs and Amazon Bedrock; early Astra use has also shown gains in debugging, system design, and non-coding work.

A Guardian opinion piece uses the resignation of AI researcher Jacob Coxon from Anthropic to examine growing insider warnings that AI could wipe out humanity, and lays out the five most-discussed doomsday scenarios, from a dangerous new bioweapon to total societal breakdown.

After a string of mathematical results turned into a reputational crisis, OpenAI has formed an independent nine-member panel of elite mathematicians, hosted by the Institute for Advanced Study, to advise on how AI companies present and release mathematical research. Researchers welcome the outreach but question its transparency, representativeness, and real influence, especially since the group's first priority is coordinating OpenAI's reported flood of model-generated results.

Apple researchers introduce probe guidance, a method that uses the frozen internal states of an existing diffusion model to construct a guidance signal without an extra forward pass at inference. It sets a new state of the art for unconditional generation with continuous diffusion language models, improves multiple-choice benchmarks on a 1.7B model, and offers insight into how autoguidance actually works.

Together AI has published a walkthrough showing how to fine-tune Qwen3.5 4B into a Jev-like classifier that takes a state plus predefined questions and returns a score, boolean, or multiple-choice answer. The run costs about $17 and takes roughly 25 minutes, after which the model is deployed behind a dedicated HTTP endpoint. A hosted version, together/Tev1-4B-experimental, is also available on Together's serverless platform.

Learn how we optimized performance and efficiency serving the workload that is changing software engineering forever — and you can too.

OpenAI and Grab launched GO Forward with AI on September 23, 2026, a two-year regional programme that will help 30,000 Grab driver-, delivery-, and merchant-partners build practical AI skills, starting in Singapore and expanding across Southeast Asia through 2027.

On 22 September 2026 Anthropic shipped Claude Opus 5.5 and, roughly an hour later, OpenAI released GPT-6 Sol and GPT-6 Luna. Both labs cut prices sharply: GPT-6 Luna lands at $0.10/$0.50 per million tokens, while Opus 5.5 drops 20% to $4/$20 with cache reads down 60%. Simon Willison's early testing also found that Opus 5.5 at its "max" thinking level exhausts the 128,000 output token limit while reasoning and returns nothing at all.

Bracket is positioned as "the memory layer for your business," a concise pitch for giving companies a unified way to store and recall business information. Public detail beyond the tagline is limited.

Google researcher John Platt joins Latent Space to discuss ERA, an LLM-driven system that automates scoreable scientific tasks, its use in climate change work such as contrail reduction and FireSat, hard-won lessons about reward hacking and overfitting, and why deep domain expertise and hands-on work still matter in the age of superintelligent AI.

Kelam is an AI product listed on Product Hunt that pitches letting an agent make the phone calls you would rather not handle yourself.

OpenAI released an improved prompt caching system for the GPT-6 family, delivering higher cache hit rates by default, discounts on eligible shared prefixes reused within a 30-minute window, and new tools for monitoring cache performance, diagnosing misses, and controlling what gets cached.