10 selected stories for 2026-11-16, grouped by topic.
Edition details
Dates use UTC+8. The latest edition uses a day with at least 5 completed reports when available; until then an earlier edition may appear. There is no fixed delivery time.
Codex reduces coding time from task start to feature merge by 30–60%, freeing engineers to focus on design and higher-stakes decisions.
A FinCrime agent built on OpenAI models automates screening and initial investigations, letting Revolut monitor more transactions without scaling investigator headcount one for one.
FrontierML tests agents on 21 long-horizon ML tasks with $10, $50, or $200 budgets; top system Opus 5.5 gains 58.8 points at $200 versus 66.1 for a human researcher.
Petrichor plans a wet lab exposed as a machine-targeted API so models can submit experiments in plain language or structured form and receive results without human relay.
GDP_xlsx comprises 70 real workplace requests spanning 13 domains, with files totaling 2.18 million populated cells and 791 rubric criteria.
Agents work in a container with spreadsheet tools; an agentic judge that opens the file applies binary criteria emphasizing completeness and exclusivity.
Streaming support for output renderers plus a new Response.stream() method lets plugins stream every row of a table in one go; built-in TSV and CSV exports now use it.
A security fix, covered by a GitHub advisory, addressed exposure of private stored query titles and SQL, though the query itself could not be executed.
Presents a synthetic multi-turn HRI dialogue dataset annotated with action intents, object references, and task plans across scenarios and user personas.
Frames intent extraction as a tool-calling problem where each predicted action is a structured function call with dialogue-grounded arguments.
Benchmarks measure models answering alone, but deployment involves contested answers, and models can abandon correct answers under faulty counter-reasoning.
The authors define dialectical competence as knowing when to hold a position and when to yield, which does not follow from first-answer accuracy and can degrade under accuracy-only training.
Amazon Bedrock now supports prompt caching for open-weight models, offering implicit and explicit modes to reduce input cost and latency for repeated context. The post demonstrates both modes via the Chat Completions API, shows how to read cache usage, and applies implicit caching to an OpenCode coding agent session.
This paper examines pre-trained visual representations for contact-rich insertion tasks, finding that the best representation varies even between similar tasks, yet spatially-grounded and contact-centric information offers distinct value for policy learning, while reliable estimation of contact cues under realistic noise remains difficult.
Researchers propose fitting a small residual adapter on frozen visual world model embeddings to improve latent goal costs for image-goal planning. Using an expanding curriculum and trajectory order, the method boosts success rates from 54% to 98% on DotWall and 82% to 96% on PointMaze without task-specific progress labels.
This paper studies budget-conditioned object search, comparing two supervision methods for VLN models: a privileged TSP expert using full map and true target locations, and a room-level Bellman expert maximizing expected return under partial observability with semantic priors. Despite higher expert performance of TSP, Bellman-supervised VLN consistently outperforms TSP-supervised VLN on held-out HM3D scenes, with advantage growing at longer budgets. At 2000 steps (500m), Bellman-supervised VLN finds 71.4% of reachable targets vs. 65.5% for TSP-supervised and 44.8% for pretrained VLN.
Researchers introduce Agentic-Teacher, an autonomous data engine that improves vision-language-action models by analyzing failures, generating reusable demonstrations, and curating high-quality data. In simulation, it boosts π0 success rates on LIBERO-10, MetaWorld-10, and RoboCasa tasks.
This paper introduces Random Filter Loss (RaFL), which compares predictions and targets through filters with random weights, avoiding pretrained features and task-specific loss engineering. RaFL outperforms or matches pixel loss on 2D and 3D image reconstruction tasks with minimal computational overhead.
An arXiv paper accepted to the NeurIPS Creative AI Track 2026 presents Typhoon, an audiovisual installation that seeds a large language model with Joseph Conrad's novella and lets it feed endlessly on its own output. Using activation steering, including a new adaptive steering method, the authors control narrative rhythm and model digressions over long generations, showing that steering is both effective and unpredictable.
The paper formalizes Hierarchical Angular Separation (HAS) observed in CLIP embeddings and proves that, in an unconstrained features model with fixed image embeddings, text-side HAS violation is bounded by image-side violation, driven by hierarchical co-occurrence of image-text pairs.
Researchers propose a full-block estimator for post-training masked diffusion language models that evaluates each block exactly once, eliminating block-selection variance and achieving lower variance than random block sampling. Experiments on five benchmarks show highest accuracy on four, with an average gain of 12.79 points over the baseline.
Researchers analyze global attention in multi-view 3D transformers like VGGT, finding each layer plays one of three roles, and propose RAWATTA, a training-free method using 8-bit attention with selective 16-bit precision to accelerate inference, fitting 200 images on a 12GB GPU with minimal accuracy loss.
A new arXiv paper evaluates small open-weight language models (8-36B parameters) for crowd decision-making on restaurant review classification. Repeating queries and output-level fusion stabilize results but do not eliminate sensitivity to participant sampling; systems output candidate sets rather than single winners. Three larger models match a hosted reference, suggesting local CDM is feasible.
Researchers find that widely used text-to-image reward models perform poorly on visual reasoning tasks, and preference fine-tuning erodes these capabilities. They propose a data-centric intervention: converting visual reasoning datasets into preference pairs that share the same image but differ in prompt, then training on a mixture with conventional preference data. This yields a Pareto improvement in visual understanding and preference prediction. The resulting reward model improves GenEval text-to-image generation by 4.0% and 5.5% on average for inference-time best-of-n selection and RL post-training, respectively, outperforming HPSv3 trained on 6x more preference pairs.
The paper proposes embedding a generative neural sensor within the control loop, treating illumination as a generative prior and decoupling high-latency semantic recovery from low-latency control. It achieves 87.1% mAP on ExDark and 100% lap completion in CARLA.
An arXiv paper by Vincent Cellucci and a co-author investigates whether AI can write an infinite poem through two experiments on couplets. The results show LLMs can produce interesting pairings but still lack human-level creativity and understanding of poetic relationships, while suggesting poetry may revive by incorporating AI workflows.
A TACL-accepted paper by Yang Wang and six co-authors proposes a task-centric framework that characterizes the model-internal processing elicited by benchmarks using attention-, activation-, and representation-based signals. Evaluating 16 LLMs across 24 benchmarks, it finds systematic processing differences between knowledge-centric and reasoning-centric tasks.
A new study places Gemma 3 12B as a defendant in London silver-fixing litigation and runs 260 market simulations to test how gain-loss framing sways LLM agents toward collusion, then injects internal states at layer 22 to probe causal mechanisms.
Researchers adapt diffusion-based crystal generative models at inference time using Diffusion Tree Sampling (DTS), a Monte Carlo tree search over the denoising trajectory, enabling guidance toward multiple conflicting properties without fine-tuning and extending to multi-objective sampling across a Pareto front.
The paper proposes NLS depth, a measure of the longest path through a model's computation that avoids natural-language-rooted interpretable bottlenecks. It can be adjudicated before training from architecture, training recipe, and deployment configuration, and is insensitive to ordinary post-training. Across open-weight models, NLS depth correlates strongly with layer count and has grown slowly; latent-reasoning architectures such as full-bandwidth transformers could increase it substantially. The authors argue developers should track and report NLS depth alongside direct CoT monitorability measurements.
The paper proposes replacing patient-level nearest-neighbor retrieval with a tabular population prior as the retrieval object for clinical RAG systems predicting medication sets from EHRs. On 1,004 held-out MIMIC-IV admissions, swapping only the retriever in PACERAG raises F1 from 0.303 to 0.338 and Jaccard from 0.218 to 0.237, with no extra LLM calls or inference-time patient index.
Researchers propose WETW, a learned distance that adapts embedding comparisons per pair of time series, combined with the WESEE encoder and ranking regularization, achieving 11.4x speedup over the fastest exact method with mean recall@10 of 0.93 across eight datasets.
This paper introduces LoRAWeaver, a training-free adapter routing method that treats adapter selection as a representation alignment problem in the adapter-output space, matching adapter-conditioned query representations against task prototypes via cosine similarity. Experiments with LLaMA-2-7B and LLaMA-2-13B on a 48-task FLAN-v2 benchmark and out-of-distribution tasks show it outperforms training-free baselines and approaches a trained retrieval encoder without learning a router.
A reproducible single-GPU audit finds that hallucination detector rankings invert across small 4-bit quantized on-device models (Spearman rho = -0.80), and a cheap lexical baseline is statistically indistinguishable from sampling-based methods. The study also shows the standard TruthfulQA-Generation labeler yields near-chance AUROC (0.46–0.59), a labeler failure rather than a detector failure.
A new study shows that hard-deletion filtering in LLM memory systems can reduce downstream accuracy when resolver errors delete query-critical facts. Across three models, hard-drop shifts from gains of +0.186 to +0.230 at perfect resolution to losses of -0.176 to -0.181 at the lowest reliability. The authors recommend treating deletion as cost-sensitive: validate false-drop risk on matched data or default to label-free retention.
Apple researchers introduce split-and-scale, a method that treats 3D scenes as continuous functions of 3D coordinates so a single pretrained tokenizer can be applied to arbitrarily small sub-regions at inference time, expanding latent capacity and detail without raising training costs.
Physicist Nathaniel Craig and GPT-6 Astra propose a candidate nonperturbative lattice definition of the full Standard Model, with Astra helping complete the detailed proof.
OpenAI's Economic Research team analyzed more than 580,000 scientific messages from 1,819 scientists in the ChatGPT for Academic Researchers summer pilot and surveyed 251 participants, finding scientists estimated the same work would take 1.9 times as long without AI and 98% expected AI to accelerate discovery.
LVMH Group’s Media Division and OpenAI have signed a strategic agreement to deploy OpenAI tools and ChatGPT Enterprise across its publications, support newsroom uses of AI, and make its journalism more accessible through ChatGPT with summaries, excerpts, and links to original articles.
An arXiv paper presents an AI-guided robotic in situ printing platform that fabricates wireless wound electrotherapy devices directly on transparent wound dressings. The system integrates 3D vision, a six-degree-of-freedom robotic arm, direct-ink-writing, and geometry-aware RF optimization. AI-based wound segmentation determines wound dimensions and local curvature, while reconstructed 3D geometry and tissue dielectric loading inform the electromagnetic design of the receiving coil and interdigitated electrodes to compensate for geometry-induced resonance shifts. Robotic fabrication was demonstrated on an anthropomorphic tissue-equivalent forearm phantom, achieving resonant operation near 25 MHz, establishing a perception-to-design-to-fabrication framework for personalized on-body electro…
This paper presents a control method for swinging up a direct-drive planar two-link robot whose motors deliver only about 6% of the shoulder gravity torque. The approach guides the robot along its nonlinear normal modes—two-dimensional invariant manifolds of periodic orbits—by decomposing the control torque into a power-neutral steering component and a scalar pumping component. The allocation is formulated as a linear program with two variables, maximizing pumping subject to torque limits and a control-Lyapunov constraint. Hardware experiments demonstrate swing-ups along both modes and compare the method with summing-and-clipping and prioritized stabilization.
QCAP-VQA supervises a subset of cross-attention heads with a question-conditioned prior from ranked segmentation proposals, using a capped prior-mass objective that penalizes insufficient attention only below a threshold. It matches baseline accuracy while increasing attention on annotated structures, but prior-referenced and expert-referenced evaluations rank models in largely opposite order.
This paper introduces a training-free open-world scene deocclusion framework that recovers the full amodal appearance of occluded objects from a single RGB image by decomposing the scene into depth-ordered RGBA layers. Accepted at WACV 2027, code will be released.
An EACL 2027 paper probes sentence-level semantics in massively multilingual self-supervised speech models across 76 languages. It finds that sentence meaning is linearly accessible from frozen speech representations but highly uneven across languages and resource tiers; cross-lingual probes can match language-specific probes on unseen languages, yet per-model monolingual decodability does not predict transfer. Scaling OmniASR wav2vec from 300M to 7B disproportionately benefits high-resource languages, and ASR tuning improves semantic decodability while making representations more language-specific. Semantic content is also entangled with acoustics, inflating performance on single-speaker corpora.
This paper studies how to choose prediction target regions for I-JEPA self-supervised learning on retinal OCT images. It compares anatomy-guided placement based on image intensity and predicted retinal segmentation against I-JEPA's default uniform placement, finding that intensity-guided rectangular targets achieved the highest linear-probe AUC for glaucoma classification on FairVision, while greater anatomical specificity did not consistently improve performance.
The paper introduces CamSpokenNLI, a pilot spoken NLI dataset of 114 scenario-matched items in each of Cameroon Pidgin English and Camfranglais, recorded by one native speaker from Buea. XLM-RoBERTa baselines reach 44.7% and 46.2% on reference text but are statistically indistinguishable from hypothesis-only input; disjoint-item cross-variety transfer scores 40.1% and 45.6%, while a counterpart-leak audit inflates them to 60-64%. Zero-shot Whisper transcription is poor, with best word error rates of 66.6% and 45.5%, and Whisper Large-v3 mislabels 93 of 114 Pidgin recordings as Yoruba or Swahili.
This paper introduces ActGCG, a method that optimises an input suffix to match target activations produced by activation steering, enabling black-box reproduction of steering vector behaviour. Experiments show that for steering a model to answer in German, a discrete suffix almost perfectly replicates the steering vector's effect and transfers across prompts and models, while for anti-refusal steering the equivalence is weaker, succeeding under a phrase-avoidance metric but only partially under the stricter HarmBench classifier.
ParseBench-2 introduces a method for developing document-parsing metrics from severity policies, using test scenarios to optimize metric implementations. The benchmark covers 4,317 pages across six parsing-quality dimensions and combines a content-recovery score with a severity cap to prevent high recovery from hiding serious errors. Experiments show the metrics align more closely with expert judgments and penalize at most 3.4% of benign variations, and are used to evaluate frontier vision-language models, specialized parsers, and open-weight systems on rankings, capability performance, and cost tradeoffs.
SURFACE introduces an adaptive scalarization weight update method that estimates the mapping between weight grid regions and Pareto front areas to allocate a finite number of solutions according to a desired preference distribution. Theoretical analysis shows linear convergence to a finite-resolution error floor, and experiments demonstrate improved target-mass balance and Pareto front coverage over baselines.
A new paper investigates whether ensembling multiple routers improves mixture-of-experts models. Testing router ensembles on two image-classification and two tabular-regression tasks, the authors find that more routers lower validation loss and report four concrete findings, though the gains are offset by extra training cost.
An arXiv paper explores how astrocytes, which make up about 50% of brain cells but do not emit electrical signals, may act as the brain's conductors. It proposes two algorithms based on astrocyte characteristics: one that modulates thousands of synapses simultaneously for faster, noise-robust supervised learning applicable to any neural network, and a rhythmic sharing algorithm for rapid anomaly detection adaptation and prevention of catastrophic forgetting. The authors report that astrocyte-derived algorithms significantly outperform neurocentric ones and address key weaknesses of state-of-the-art ML models.
This paper studies episodic adversarial MDPs with unknown dynamics and full-information rewards, achieving regret Õ(√(H³SAK)). It introduces VOPO, a policy optimization algorithm with a Stability-Matching Bonus, and provides a reduction from unknown to known dynamics for downstream-dependent algorithms, narrowing the remaining gap to finding an Õ(H√K) known-dynamics algorithm.
This study develops an availability-aware historical-feedback learning framework for predicting EAF endpoint temperature and final concentrations of twelve chemical elements. Eight industrial data sources are integrated into 19 static, 39 temporal, and 47 history-informed features, with historical outcomes incorporated only under an assumed 60-minute measurement-release delay. Using 19,738 temperature heats and 3,341 chemistry heats, twelve tuned regression algorithms are evaluated via chronological forward validation. The historical-feedback representation reduces RMSE in 147 of 156 matched comparisons; ExtraTrees achieves 12.290°C RMSE, a 5.54% reduction over its temporal-only counterpart.
This paper introduces model annealing, a training method that uses the remaining budget for review and consolidation by relearning from the network's own knowledge via reheated episodes, a prediction memory, and adversarial dreams, improving accuracy, calibration, and robustness across datasets and architectures without extra inference cost.
An arXiv paper evaluates MXFP8 and 6-bit MXFP6 microscaling formats for industrial-scale recommendation and ranking (R&R) inference on AMD MI350X, reporting significant GEMM kernel speedups over FP16 and end-to-end throughput gains from selective MXFP6 under a strict NE parity threshold.
This paper extends a unified geometric perspective to exponential-tail losses in separable linear classification, identifies a curvature-defined reference subspace replacing the minimizer manifold, derives explicit iterate bounds revealing a three-phase damping structure, and uncovers a transient implicit bias induced by large step sizes.
Simon Willison released datasette-maplibre 0.1a0, a new Datasette geospatial plugin combining MapLibre GL and OpenFreeMap, built mostly from a single prompt to Claude Fable 5.2 with features like GeoJSON SVG previews and lazy-loaded WebGL.
OpenAI announced computer-use agents, improved trusted clinical search, and more than 20 new healthcare plugins for ChatGPT for Healthcare, aiming to help teams work across EHRs, scheduling systems, and payer portals. A blind evaluation of over 1,300 physician reviews found the updated clinical experience rated more highly across all eight dimensions.
TouchTwin uses an 880-taxel glove and MuJoCo simulation to rebuild real human touch recordings, closing the force gap in vision-based real-to-sim-to-real pipelines. A touch-grounded policy grips at 0.99x recorded force and succeeds in 20/20 held-out trials and 35/36 unseen perturbations.
Researchers recast powder X-ray diffraction (pXRD) structure determination as a verifiable agentic search problem, introduce the 200-problem pXRD-Bench, raise structural recovery from 27% to 69% via agent adaptation, and use a reflexive evolution framework to improve profile agreement on 15 experimental patterns and transfer an analogue-substitution procedure for a phase missing from the reference library.
Researchers introduce Unfinished-Bench, a 97-scenario benchmark evaluating whether persistent AI agents violate rules embedded in realistic work environments during autonomous, user-free runs. Persistence raised violations of an internal-document rule from 32.2% to 43.6%, five times the 8.5% rate for the same rule stated by the user.
Over two days at COEX, more than 2,000 attendees converged for AI Day Seoul, bringing together Korea’s industry leaders, startups and developer communities advancing physical AI and AI agents. At the plenary sessions, Deepu Talla, vice president of robotics and edge AI at NVIDIA, said Korea is “punching above its weight,” with Doosan Robotics, Hyundai Motor Group, LG Electronics and SK Group — as well as startups like Real World and Luda Robotics — using NVIDIA platforms to bring general-purpose robotics into the real world. Jonathan Cohen, vice president of applied research at NVIDIA, showed how enterprises can build agents they own and control. Breakout sessions featured speakers from Coupang Intelligent Cloud, Doosan Robotics, LG AI Research, LG Electronics, NAVER Cloud, SK Telecom and…
Coding agents send large inputs and receive small outputs, making prefix caching critical for long sessions. The article examines cache-preserving practices on both the harness and inference engine sides, including tool loading, system prompt stability, reasoning effort, output truncation, subagents, and routing.
Together AI built agent evals to measure how well coding agents could use its platform, then fixed docs, skills, and API errors to raise task completion from 75% to 97%.
Together AI built agent evals to measure how effectively coding agents can use its platform. Small fixes to docs, skills, and API errors lifted task completion from 75% to 97% across tested models.
OpenAI launches a science blog to give scientists and academic researchers a place to share AI-driven discoveries, practical methods, and firsthand perspectives, highlighting recent results in mathematics and biology.
AGIBOT researchers present X-Nav, described as the first system to take a bipedal humanoid through a language-instructed long-horizon mission from indoors to outdoors. It combines a data engine, the RGB-only MELIOR navigation model, and the X-Navarch agentic workflow, and completes a roughly 100-meter real-world mission on the Lingxi X2 while also posting competitive results on R2R-CE and RxR-CE.
This paper evaluates image-specific and image-agnostic adversarial attacks on YOLOv3/YOLOv4 detectors paired with SORT, Deep SORT, Strong SORT, and OC-SORT trackers in the CARLA simulator. Image-specific attacks drop HOTA for YOLOv3-based trackers from 53.45-56.19% to 14.85-17.03%, with MOTA reaching -772.60% in the worst case. YOLOv4-based trackers are more robust but HOTA still falls to 25.06-26.86%. Image-agnostic attacks using a single Universal Adversarial Perturbation (UAP) can also degrade tracking accuracy without per-frame perturbations.
The paper introduces a vision-language-model-based framework that infers task- and scene-dependent safety constraints for robot manipulation, first rejecting tasks with non-recoverable hazards, then identifying preventable object-pair hazards and grounding them into continuous 3D risk fields for obstacle avoidance.
The paper introduces Reflect-Reasoning, a training method that teaches models how to reason about alignment to principles rather than which principles to follow; across five models it outperforms system prompting and explicit in-context reflective prompting, generalizes to out-of-distribution medical and financial principles, reduces rare but serious violations, and preserves safety and jailbreak robustness.
The King’s Trust and OpenAI have announced a £1 million partnership to help up to 25,000 young people across the UK build practical AI skills for work, prioritizing those not in education, employment or training (NEET). The 18-month program will combine AI skills learning with nationally recognized qualifications to help young people navigate an AI-enabled economy.
ERGo is a hybrid training framework that automatically switches from a full-parameter phase to a parameter-efficient phase during training, reducing off-chip memory access on processing-in-memory devices. It achieves up to 3.15x speedup and 10.5x energy efficiency on PIM platforms, and extends the lifetime of heterogeneous NVM-SRAM architectures by up to 33x.
Holiday Extras leverages ChatGPT-powered dots to proactively generate launch reports, flag potential issues, and discuss meetings on commutes, boosting efficiency.