Skip to content
AI News HubLIVE

This edition’s highlights

Models

GPT‑6 Cyber: Advancing the frontier of aligned cybersecurity agents

  • GPT‑6 Cyber scores 48.5% on ExploitGym, up from 32.1% for GPT‑5.6 Cyber.
  • Alignment evaluation shows 99.93% out-of-scope avoidance without prompts, 100% with prompts.
OpenAI NewsIn-site articleGPT‑6 Cyber: Advancing the frontier of aligned cybersecurity agents

Merge builds product and business tools with OpenAI technology

  • Merge CEO Shensi Ding uses Codex to review product code, including work on a new setup interface for Merge for Workforce.
  • Teams use GPT-6 Astra for product testing and marketing, and automate business tasks like sales updates and executive dinner planning.
OpenAI NewsIn-site articleMerge builds product and business tools with OpenAI technology

SRRM: A Retrieval-Reconstruction Benchmark for Evaluating Long-Context Memory in Small Language Models

  • SRRM is a lightweight, reproducible benchmark that jointly evaluates targeted retrieval and global information reconstruction.
  • The Memory Preservation Ratio (MPR) quantifies the relationship between retrieval and reconstruction performance.
arXiv Computational LinguisticsIn-site articleSRRM: A Retrieval-Reconstruction Benchmark for Evaluating Long-Context Memory in Small Language Models

Human-like cross-language generalisation in deep neural speaker embeddings and its acoustic foundations

  • The study compared speaker embeddings from 18 state-of-the-art models with human cross-language similarity judgments on Cantonese-English bilingual speech.
  • Embeddings exhibited cross-language generalisation that broadly paralleled human behaviour, including accurate identity discrimination and sensitivity to language switching.
arXiv Computational LinguisticsIn-site articleHuman-like cross-language generalisation in deep neural speaker embeddings and its acoustic foundations
Agents

‘Hogwarts Legacy’ Gets a Magical NVIDIA DLSS 5 Upgrade on GeForce NOW

  • Hogwarts Legacy receives DLSS 5 with 3D-Guided Neural Rendering, enhancing lighting and materials while preserving the game's art style.
  • World of Warcraft: Forever launches on GeForce NOW with CurseForge add-on support and a launch-week access reward for new users.
NVIDIA BlogIn-site article‘Hogwarts Legacy’ Gets a Magical NVIDIA DLSS 5 Upgrade on GeForce NOW

SlackBench: Benchmarking Agents on Collaborative Projects Grounded in Real Code Repositories

  • SlackBench features 140 queries across four task types: heterogeneous-source agentic search, abstention, goal-directed summarization, and circumstance inference.
  • Environments are privacy-preserving reenactments of real academic projects with multi-party dialogue, evolving decisions, and corresponding code changes in GitHub and Google Docs.
arXiv AIIn-site articleSlackBench: Benchmarking Agents on Collaborative Projects Grounded in Real Code Repositories
Research

Faster Retrieval for Better, More Efficient RAG

  • MIPS is reformulated as learning the support function of the key set, which is convex and can be approximated by input convex neural networks.
  • Two models are proposed: SupportNet learns per-cluster support functions for routing; KeyNet directly predicts the best-matching key and acts as a query rewriter for existing IVF indices.
Apple Machine Learning ResearchIn-site articleFaster Retrieval for Better, More Efficient RAG

Kimi K3.1 Pro on Fireworks: More Intelligence per Dollar, Delivered Fast

  • Ranked first of 19 on Sierra's τ³-Banking at 64.7%, 8.7 points ahead of Claude Opus 5.5
  • Post-trained from K3 with sparse attention replacing full-attention layers for more efficient long-context serving
Fireworks AI BlogIn-site articleKimi K3.1 Pro on Fireworks: More Intelligence per Dollar, Delivered Fast
Robotics

OdorSpaces: Visual-Olfactory Embodied Navigation in 3D Environments

  • OdorSpaces augments Habitat-Sim with Fire Dynamics Simulator for 3D odor transport and dispersion.
  • The benchmark includes two source-seeking scenarios: invisible gas-leak localization and early-stage fire localization.
arXiv RoboticsIn-site articleOdorSpaces: Visual-Olfactory Embodied Navigation in 3D Environments

You have reached the end of this edition.

Your reading list
Other updates (57)
Agents

Push-based CDC into Unity Catalog with Zerobus Ingest and Debezium

Databricks and NVIDIA collaborated to add a Zerobus Ingest sink to Debezium Server, enabling CDC events to be pushed directly into Unity Catalog Delta tables without an intermediate Kafka cluster. This architecture simplifies data replication from operational databases, especially in private networks with only outbound connectivity, and ensures reliable at-least-once delivery with idempotency keys.

Databricks BlogIn-site articlePush-based CDC into Unity Catalog with Zerobus Ingest and Debezium

AI-Driven SaaS Platform Integration: AI Express for Unified Generative AI Services with Pay-As-You-Go Credit Systems and Scalable Microservice Architecture

An arXiv paper introduces AI Express, a SaaS platform that consolidates text generation, image synthesis, and code assistance into a single pay-as-you-go service built on a scalable microservice architecture, aiming to reduce costs and workflow disruption caused by fragmented generative AI subscriptions.

arXiv Computational LinguisticsIn-site articleAI-Driven SaaS Platform Integration: AI Express for Unified Generative AI Services with Pay-As-You-Go Credit Systems and Scalable Microservice Architecture

Towards Robust Long-Lived Agents: Benchmarking Experience Lifecycles in Evolving Environments

Researchers introduce ReScopeBench, a benchmark spanning 18 interactive environments to evaluate how agents acquire, reuse, and revise experience across task streams. Testing 20 models reveals that retained experience improves task performance, but agents struggle to keep it current as environments change, and correct experience recording does not guarantee task success, highlighting selective maintenance and effective use as complementary challenges.

arXiv AIIn-site articleTowards Robust Long-Lived Agents: Benchmarking Experience Lifecycles in Evolving Environments

Challenge-Aware Knowledge Recruitment with Meta-Agent Supervision for Medical Multi-Agent Reasoning

The paper proposes challenge-aware adaptive recruitment, where an external CHALLENGER agent inspects the reasoning trace to identify factual or expertise gaps and triggers targeted retrieval or mid-discussion specialist recruitment. On the MEDAGENTSBENCH hard subset, in-between retrieval outperforms upfront retrieval, and CHALLENGER-triggered recruitment raises GPT-4.1-mini MedQA accuracy from 46% to 54%. Naively combining both channels degrades performance, highlighting the need to coordinate adaptive interventions.

arXiv AIIn-site articleChallenge-Aware Knowledge Recruitment with Meta-Agent Supervision for Medical Multi-Agent Reasoning

Uncertainty-Aware Resource Orchestration for Agentic AI Systems

An arXiv paper formalizes the uncertainty-aware agentic orchestration problem for LLM-driven multi-step pipelines whose compute, bandwidth, and latency demands are unknown at dispatch time. It proposes CRO, an online algorithm that uses conformal prediction over sliding execution traces to produce dispatch-time reservations with finite-sample, distribution-free coverage guarantees. Simulations show CRO is the only evaluated method that meets target probabilistic coverage at all risk levels while achieving the lowest over-reservation among coverage-feasible methods, roughly halving the resource waste of a Chebyshev-bound baseline.

arXiv AIIn-site articleUncertainty-Aware Resource Orchestration for Agentic AI Systems

Agent Harness Awareness: Turning Verified Telemetry into Operational Control for Tool-Using Agents

An arXiv paper introduces Agent Harness Awareness, a trust-separated design that provides verified tool-health state snapshots to agents before consequential actions, closing the "operational blindness" gap. Experiments show a 96.9% reduction in policy regret and an increase in single-dispatch completion from 69.4% to 95.8%.

arXiv AIIn-site articleAgent Harness Awareness: Turning Verified Telemetry into Operational Control for Tool-Using Agents

Smart Approvals in Amazon Quick: Human oversight where it matters

Smart Approvals in Amazon Quick reviews eligible agent actions within existing controls, automatically clearing routine steps while returning consequential decisions to a person, with audit evidence available through AWS CloudTrail and Amazon CloudWatch Logs.

AWS Machine Learning BlogIn-site articleSmart Approvals in Amazon Quick: Human oversight where it matters

Four layers of context engineering for production AI agents: How Deep Insight applies them in working code

In production AI agents, context management is the real bottleneck. This post presents a four-layer context engineering framework (architecture, prompts, file-based tools, and verification agents), implemented in working code from Deep Insight, an open source multi-agent data analysis solution on the Strands Agents SDK and Amazon Bedrock AgentCore.

AWS Machine Learning BlogIn-site articleFour layers of context engineering for production AI agents: How Deep Insight applies them in working code

How U Mobile built agentic telco operations with Amazon Bedrock AgentCore

U Mobile, Malaysia's newest 5G network provider, built AiDaH, an agentic AI operations platform on Amazon Bedrock AgentCore that moves beyond AI assistance to autonomous action across telco operations. Nine months after production launch, AiDaH handles over 150 operational scenarios across six divisions, cutting trouble tickets by up to 84.5%.

AWS Machine Learning BlogIn-site articleHow U Mobile built agentic telco operations with Amazon Bedrock AgentCore
Research

GRASP: Gradient-Guided Sampling for Contact-Rich Physics-Based Motion Retargeting

GRASP introduces a physics-based motion retargeting method that requires no differentiable simulator, recovering gradients from archived rollouts to guide sampling, and scaling to long demonstrations via MPC, achieving 92.8% success on bimanual OakInk2 demos and improving PPO tracking.

arXiv RoboticsIn-site articleGRASP: Gradient-Guided Sampling for Contact-Rich Physics-Based Motion Retargeting

OUCH: Objective, Uniform, Controlled Haptic Stimulation for Automated Neurological Assessment

OUCH is a wrist-worn pneumatic haptic device that uses stacked pouch actuators, a rigid indenter, and a buckling spacer to deliver controlled noxious stimuli with high local contact pressure at lower total force, offering a foundation for objective neurological assessment in disorders of consciousness.

arXiv RoboticsIn-site articleOUCH: Objective, Uniform, Controlled Haptic Stimulation for Automated Neurological Assessment

ELA: End-to-End Latent Action Policy Learning from Video Demonstrations

This paper introduces ELA, an end-to-end framework that jointly trains a latent action model, a latent action policy, and an action decoder so that robot policies can exploit abundant action-free video demonstrations with only a handful of action-labeled robot demonstrations. ELA consistently outperforms prior latent action baselines on simulated and real-world manipulation tasks and integrates with existing imitation learning methods through lightweight 4M-parameter action decoders.

arXiv RoboticsIn-site articleELA: End-to-End Latent Action Policy Learning from Video Demonstrations

MESA: An Evaluation Framework for Compositional, Semantic, and Spatial Generalization in Robotics

Researchers introduce MESA, a framework for systematically evaluating the generalization of language-conditioned robotic manipulation policies. It offers five test suites that isolate in-distribution performance and four axes of generalization: spatial configuration, object instance, object category, and subtask composition. The team also develops MESA-Gen, a pipeline for scalable task and demonstration generation, and releases both a single-arm MESA suite and BiMESA, the first large-scale benchmark for bimanual generalization. Across seven baselines, the best model achieves only 52% on MESA and 49% on BiMESA, indicating ample room for improvement. MESA rankings broadly align with RoboArena scores.

arXiv RoboticsIn-site articleMESA: An Evaluation Framework for Compositional, Semantic, and Spatial Generalization in Robotics

SO(3)-Equivariant Learning on CAD Boundary Representations

Researchers introduce SO3brep, an SO(3)-equivariant graph neural network on B-rep face-adjacency graphs. It computes node features in local frames, maps them to a shared frame with Wigner D-matrices, and uses Clebsch-Gordan tensor products to preserve relative orientation between adjacent faces, yielding SE(3)-invariant outputs. Across three face-segmentation and two classification benchmarks with randomly rotated test parts, SO3brep achieves the highest accuracy with the fewest parameters, with the largest advantage in low-label industrial CAD settings.

arXiv Computer VisionIn-site articleSO(3)-Equivariant Learning on CAD Boundary Representations

SurgTransfer: Multi-Step Latent Action Modeling for Action Transfer in Surgical World Model

The paper introduces SurgTransfer, a multi-step latent action learning framework that learns context-invariant latent actions from unlabeled surgical videos for transfer within surgical world models. The authors validate the approach on the public SurgToolLoc-2022 and PitVis-2023 datasets using quantitative metrics and qualitative evaluation by surgical experts, and release their code.

arXiv Computer VisionIn-site articleSurgTransfer: Multi-Step Latent Action Modeling for Action Transfer in Surgical World Model

PointNDF: Cascaded Neural Displacement Fields for Point Cloud Denoising

PointNDF is a point cloud denoising framework that combines patch-level geometric context with local displacement prediction, using cascaded neural displacement fields to refine point positions and achieve more uniform point distributions.

arXiv Computer VisionIn-site articlePointNDF: Cascaded Neural Displacement Fields for Point Cloud Denoising

mmHand: Toward Pixel-Level-Accuracy Hand Localization Using a Single Commodity mmWave Device

The paper introduces mmHand, a device-free, privacy-preserving dynamic hand localization system that reaches pixel-level accuracy using a single commodity mmWave radar, validated on 12 subjects across three scenarios.

arXiv Computer VisionIn-site articlemmHand: Toward Pixel-Level-Accuracy Hand Localization Using a Single Commodity mmWave Device

VAR-Q: Tuning-free KV Cache Quantization for Visual Autoregressive Image and Video Generation

VAR-Q is a tuning-free KV-cache quantization framework for visual autoregressive image and video generation that exploits the distinctive block-wise cache statistics of these models, cutting KV memory by up to 78.5% across nine models and enabling 6x larger inference batches.

arXiv Computer VisionIn-site articleVAR-Q: Tuning-free KV Cache Quantization for Visual Autoregressive Image and Video Generation

Same Origin, Different Verdicts: Text Production Conditions Confound Machine-Generated Text Detection in Russian

A study on Russian machine-generated text detection finds that production conditions significantly shift detection scores even when origin is fixed, with disagreement among four operationalizations and a flip from machine to human verdict caused solely by removing emphasis markers.

arXiv Computational LinguisticsIn-site articleSame Origin, Different Verdicts: Text Production Conditions Confound Machine-Generated Text Detection in Russian

DS@GT ARC at CheckThat! 2026: A Verdict-Anchored Pipeline for Generating Fact-Checking Articles

The DS@GT ARC team developed a four-stage pipeline for CLEF 2026 CheckThat! Lab Task 3 that generates full fact-checking articles from a claim, its original verdict, and scraped source documents. The system ranked 3rd out of 9 teams with a mean score of 0.428, outperforming the baseline of 0.272.

arXiv Computational LinguisticsIn-site articleDS@GT ARC at CheckThat! 2026: A Verdict-Anchored Pipeline for Generating Fact-Checking Articles

Discourse Analysis of Legislative Bill Discussions

A new arXiv paper introduces an end-to-end pipeline for parsing rhetorical structures from legislative committee hearings and proposes a rhetorical structure grammar; 55.21% of California state bill discussions in its dataset parse with that grammar, and a logistic regression classifier with grammar-based output masking and tag smoothing reaches a 52.8% F1 score over 100 hearings and more than 2,000 annotated utterances, outperforming more complex models.

arXiv Computational LinguisticsIn-site articleDiscourse Analysis of Legislative Bill Discussions

Support Vector Machines Directly Minimise KL Divergence

A new paper proves that the dual objectives of support vector classification and regression are exact constrained KL divergence minimisation tasks between Gaussian measures, providing an information-theoretic interpretation for SVMs and unifying kernel methods.

arXiv Machine LearningIn-site articleSupport Vector Machines Directly Minimise KL Divergence

FedSharp: Byzantine Robust Decentralized Federated Learning Under Noisy Channel

The paper proposes FedSharp, a robust decentralized federated learning framework that addresses data heterogeneity, Byzantine attacks, and channel noise. By screening suspicious models, averaging the lowest-loss candidates, and applying noise-adaptive regularization, FedSharp achieves 93.5-94.9% accuracy on MNIST with 33 Byzantine nodes, far surpassing benchmarks.

arXiv Machine LearningIn-site articleFedSharp: Byzantine Robust Decentralized Federated Learning Under Noisy Channel

The Confidence Score: a New Measure on Kernels

Researchers propose the "Confidence Score", a new kernel measure that replaces traditional alignment-to-ideal-kernel metrics. It measures alignment between an encoding of ground truth and the kernel-projected ground truth, allowing kernels that preserve intra-class structure to score highly, and shows higher correlation with classification accuracy on real and synthetic datasets.

arXiv Machine LearningIn-site articleThe Confidence Score: a New Measure on Kernels

Objective-Conditioned Active Preference Learning for Multi-Objective Alignment

This paper introduces OGC-DPO, an active preference learning framework that treats each (response pair, objective) combination as a separate query, measures informativeness via the policy's implicit gradient, and selects queries that most expand each objective's gradient coverage, reducing annotation cost for multi-objective alignment. Experiments show OGC-DPO outperforms existing methods under the same annotation cost, and informative conflicts provide useful training signals.

arXiv Machine LearningIn-site articleObjective-Conditioned Active Preference Learning for Multi-Objective Alignment

Matched but Not Better: An Equivalence-Tested Benchmark of Quantum Machine Learning on Six Clinical Datasets and the QML-Health-READY Reporting Framework

An arXiv preprint benchmarked quantum classifiers against tuned classical models on six clinical datasets. Quantum models did not significantly outperform classical ones; finite-shot estimation was the main practical cost, not gate noise; configuration errors could reverse conclusions. The authors propose the QML-Health-READY reporting framework.

arXiv Machine LearningIn-site articleMatched but Not Better: An Equivalence-Tested Benchmark of Quantum Machine Learning on Six Clinical Datasets and the QML-Health-READY Reporting Framework

Constraint-Preserving Flow Matching for Spectral Graph Generation

The paper introduces Laplacian-Constrained Stiefel Flow Matching (LaSFM) to preserve the Laplacian nullspace constraint alongside eigenvector orthonormality and positive eigenvalues during spectral graph generation, and extends it to property-guided generation with improved spectral and graph-level metrics.

arXiv Machine LearningIn-site articleConstraint-Preserving Flow Matching for Spectral Graph Generation

Towards Efficient Whole Page Ranking with One-Pass Inference

The paper proposes Multi-Item Scoring (MIS), a prefix-aware non-autoregressive framework that captures order-dependent interactions in whole-page ranking while maintaining efficient inference. Experiments show MIS approaches autoregressive generation quality with 4.07x and 3.54x faster CPU decoding on two datasets, and in a large-scale social-video platform evaluation, it improves the realization metric by 0.104% and watch time by 0.058% while reducing ranking-service CPU consumption by 9.73%.

arXiv AIIn-site articleTowards Efficient Whole Page Ranking with One-Pass Inference

Can Large Language Models Review the Scientific Literature? A Test Against a 102-Article Benchmark

A study evaluates six AI tools across four stages of literature review, finding they can assist with discovery and interpretation—especially with institutional access—but cannot replace expert judgment in evidence evaluation and synthesis.

arXiv AIIn-site articleCan Large Language Models Review the Scientific Literature? A Test Against a 102-Article Benchmark

Franklin Templeton uses OpenAI to support investment research

Franklin Templeton is developing AI workflows to help portfolio managers compare research views and examine investment risks, starting with equity research.

OpenAI NewsIn-site articleFranklin Templeton uses OpenAI to support investment research

Use decision models with Amazon Bedrock AgentCore Evaluations

This post explains how to connect two decision-model backends, the hosted Jev model and the open-source Strands Decider 2B, to Amazon Bedrock AgentCore Evaluations through code-based evaluators, how their probability outputs map to AgentCore results, and how to validate a decision-model evaluator on your own labeled traffic before gating a release on it.

AWS Machine Learning BlogIn-site articleUse decision models with Amazon Bedrock AgentCore Evaluations
Robotics

Correct Robots Before They Make Mistakes: Proactive Human-in-the-Loop Intervention via Preference Learning

The paper introduces PHIL, a framework that lets operators preview a policy's predicted future motion through an XR interface and intervene before undesirable actions occur, pairing rejected predictions with corrective action chunks for preference-based post-training and substantially raising real-world manipulation success rates.

arXiv RoboticsIn-site articleCorrect Robots Before They Make Mistakes: Proactive Human-in-the-Loop Intervention via Preference Learning

$\mathcal{C}^\infty$ Smooth Trajectory Planning with Application to Soft Robots

This paper introduces a $\mathcal{C}^\infty$-smooth trajectory generator that produces infinitely differentiable trajectories while respecting kinematic limits, suitable for soft continuum robots with high smoothness requirements.

arXiv RoboticsIn-site article$\mathcal{C}^\infty$ Smooth Trajectory Planning with Application to Soft Robots

Contact-Implicit Whole-Body Trajectory Optimization for Dynamic Humanoid Locomotion on Parameterizable Uneven Terrain

This paper presents a contact-implicit whole-body trajectory optimization framework that enables humanoid robots to plan dynamic motions on complex, parameterizable terrains without predefined contact sequences. The method approximates ground geometry with smooth surfaces and introduces novel contact constraints to extend applicability from flat ground to arbitrary smooth surfaces, with preliminary tests in simulation and on the ErgoCub robot.

arXiv RoboticsIn-site articleContact-Implicit Whole-Body Trajectory Optimization for Dynamic Humanoid Locomotion on Parameterizable Uneven Terrain

Ctrl-F: Force as an Interface for Humanoid Loco-Manipulation

Ctrl-F introduces a hierarchical motion-force framework that augments humanoid motion tracking with explicit 3-D force commands, enabling loco-manipulation tasks requiring large, sustained interaction forces and successfully manipulating 20 kg boxes on hardware.

arXiv RoboticsIn-site articleCtrl-F: Force as an Interface for Humanoid Loco-Manipulation

Jev with Foresight: Safe Language-Guided Autonomous Driving via a Language World Model

Jev-AD combines a typed decision model with a language world model to foresee consequences of driving actions, reducing collision rates from 26.0% to 2.4% in simulation and from 48.2% to 2.8% in real-world replays.

arXiv RoboticsIn-site articleJev with Foresight: Safe Language-Guided Autonomous Driving via a Language World Model
Models

From Tokens to Trajectories: Parameter-Efficient Continuous Adaptation of Unified Multimodal Models

FlowShift converts pre-trained discrete unified multimodal models into continuous rectified-flow generators using LoRA and dual-objective velocity vector distillation, significantly improving image generation quality while preserving over 99.5% of multimodal understanding.

arXiv Computer VisionIn-site articleFrom Tokens to Trajectories: Parameter-Efficient Continuous Adaptation of Unified Multimodal Models

Augmentation with Reinforcement Transformation for Low-textured Scientific Sketch Image Retrieval

This paper proposes Augmentation with Reinforcement Transformation (ART), which formulates augmentation selection as a multi-armed bandit problem to adaptively construct view banks for fine-tuning pretrained models on low-textured sketch images. On Florida manatee sketches, ART achieves 71.40% Top-1 accuracy for transformed-scar retrieval and 55.84% for location-aware retrieval, outperforming random augmentation and frozen CLIP.

arXiv Computer VisionIn-site articleAugmentation with Reinforcement Transformation for Low-textured Scientific Sketch Image Retrieval

PEEK: One-Step-Look-Ahead Exposure Bias Correction for Diffusion Sampling

PEEK is a training-free deterministic sampler plugin that mitigates exposure bias in diffusion models by performing a one-step look-ahead evaluation and interpolating predictions, achieving up to 49.3% FID improvement on multiple models.

arXiv Computer VisionIn-site articlePEEK: One-Step-Look-Ahead Exposure Bias Correction for Diffusion Sampling

When Does Sharing Work? Scaling a Shared SigLIP2 Across Image, Video, Audio, and Audio-Visual Representations

Researchers introduce Omni-Encoder, a single SigLIP-2 Vision Transformer with fully shared parameters that handles image, video, audio, and audio-visual inputs, reaching competitive results on 16 benchmarks while joint training improves every modality over its separately trained counterpart.

arXiv Computer VisionIn-site articleWhen Does Sharing Work? Scaling a Shared SigLIP2 Across Image, Video, Audio, and Audio-Visual Representations

Visual Forensic Detection of Fabricated Bangla News Photocards using Deep Learning

This study recasts photocard fake-news detection in Bangladesh as visual image forgery detection. It builds a dataset of 969 authentic and 1,077 fabricated Bangla photocards, augmented to 32,736 samples. A fully connected ANN achieves only ~52% accuracy, while a compact four-block CNN reaches 87.45% and a fine-tuned YOLOv8n-cls backbone reaches 97.36%, highlighting underfitting in the compact model.

arXiv Computer VisionIn-site articleVisual Forensic Detection of Fabricated Bangla News Photocards using Deep Learning

RAGCompass: Generalizable Retriever Routing via Reranking

RAGCompass routes among retrievers by reading the retrieved passage sets themselves, using a contrastively fine-tuned reranker to judge which evidence set best supports the answer. It improves answer F1 by 28% over the best fixed retriever and generalizes to unseen corpora, retrievers, and readers.

arXiv Computational LinguisticsIn-site articleRAGCompass: Generalizable Retriever Routing via Reranking

Rewrite, Then Veto: Rewriter-Agnostic Edit Gating for LLM-Based Post-OCR Correction

The paper proposes EditVeto, a veto gate placed after a full-context LLM rewrite: the rewrite is character-aligned to the OCR input, differences are split into blocks, and a 0.6B-parameter gate keeps or reverts each block. On the eight HIPE-OCRepair-2026 test sets, the gate improves all five rewriters tested, including two unseen in training.

arXiv Computational LinguisticsIn-site articleRewrite, Then Veto: Rewriter-Agnostic Edit Gating for LLM-Based Post-OCR Correction

When to Ask, When to Answer: Strategic Clarification for Ambiguous Reference Resolution in Dialogue

This paper studies strategic clarification in dialogue, where LLMs must decide when to ask questions to narrow candidate referents instead of prematurely answering. Using a controlled environment built from SIMMC 2.1, the authors evaluate baseline LLMs and post-training methods, and propose BACE, a lightweight action chooser that more consistently selects the optimal number of clarifying questions than GRPO and ArCHer.

arXiv Computational LinguisticsIn-site articleWhen to Ask, When to Answer: Strategic Clarification for Ambiguous Reference Resolution in Dialogue

CCDWIKI: Decoupling Case Formulation from Response Generation in Longitudinal CBT Dialogue

CCDWIKI is a plug-and-play advisor that models longitudinal CBT case formulation as evidence-grounded hypothesis adjudication, improving conceptualization and intervention quality without retraining the backbone language model.

arXiv Computational LinguisticsIn-site articleCCDWIKI: Decoupling Case Formulation from Response Generation in Longitudinal CBT Dialogue

Recasting Visual Language Models as EEG Foundation Models for Clinical Neurodiagnostics

A new study shows that generalist vision language models, without large-scale EEG pretraining, perform competitively with specialist EEG foundation models on clinical neurodiagnostic tasks, and adapt better in few-shot settings and with model scaling.

arXiv Machine LearningIn-site articleRecasting Visual Language Models as EEG Foundation Models for Clinical Neurodiagnostics

TSFM Meets LLM: Context as Covariates

This paper proposes Context as Covariates (CoCo), a forecasting framework where an LLM distills textual reasoning into numerically grounded forecast and confidence covariates to guide a TSFM backbone. The approach first fine-tunes the LLM via GRPO with a reward tied to forecasting improvement, then fine-tunes the TSFM on the aligned LLM's covariates. It achieves up to 12% MAE improvement over state-of-the-art TSFMs and LLM forecasters like GPT-4o on real benchmarks spanning finance, economics, weather, traffic, and security.

arXiv Machine LearningIn-site articleTSFM Meets LLM: Context as Covariates

Learning Implicit Solves for Dynamic Simulations

Researchers introduce the Implicit Neural Operator Solver (INOS), which pretrains a neural operator from static equilibrium solutions and embeds it within Newmark-β time integration. This enables accurate dynamic simulations without dynamic trajectory supervision, achieving over 4,600× speedup over sequential finite-element solvers in nonlinear elastodynamics and stable million-step rollouts.

arXiv Machine LearningIn-site articleLearning Implicit Solves for Dynamic Simulations

A Feasibility Condition for Confidence-Gated Tiered LLM Inference

The paper presents a closed-form feasibility condition for confidence-gated tiered LLM inference, showing when no confidence threshold can make a cascade worthwhile. It introduces an architecture cost ratio κ, reveals a condition inversion between billed and sunk local compute, and derives three design requirements plus a future co-designed architecture and protocol.

arXiv AIIn-site articleA Feasibility Condition for Confidence-Gated Tiered LLM Inference

Who May Overrule the Agent? Evidence-Gated Authority in LLM Review Institutions

This paper frames reviewer authority in multi-agent LLM systems as a question of institutional jurisdiction, comparing broad override authority with an evidence gate. Experiments show the gate sharply reduces corruption of correct proposals, but its effectiveness depends on model behavior and is not a model-independent safeguard.

arXiv AIIn-site articleWho May Overrule the Agent? Evidence-Gated Authority in LLM Review Institutions

The Enterprise Case for Open-Weight Models on Databricks

Databricks field engineers observe enterprises reassessing their AI strategies after two years of standardizing on single frontier providers, driven by seven pressures: insufficient returns on AI spend, frequent model ranking changes, frontier guardrail limitations, evolving data retention policies, governance visibility needs, harness impact on benchmark results, and open-weight models reaching frontier-class performance at lower cost. Open-weight models offer cost efficiency, data sovereignty, and operational autonomy, while Unity AI Gateway provides unified access, smart routing, budgets, and centralized governance. Usage data shows open-weight adoption rising sharply, and Databricks announced a $3 billion raise partly to secure GPU capacity.

Databricks BlogIn-site articleThe Enterprise Case for Open-Weight Models on Databricks

Agile Deliberation: How AI helps humans define the gray areas

Google Research introduces Agile Deliberation, a framework that turns building subjective AI image classifiers into an iterative conversation between humans and AI, using an LLM-driven optimization loop to resolve edge cases and produce accurate, aligned classifiers.

Google Research BlogIn-site articleAgile Deliberation: How AI helps humans define the gray areas

How big is the open model economy today?

This is the first open model economy research note for paid subscribers, examining the size of the open model inference market, whether it is growing, and whether it is a rounding error relative to OpenAI and Anthropic. The author uses open model inference startups Baseten, Fireworks, and Together AI, along with OpenRouter's open model token share, as bellwethers and compares them with Google's reported token volumes.

Interconnects (Nathan Lambert)In-site articleHow big is the open model economy today?

How Indegene is modernizing and scaling enterprise data quality in life sciences

Indegene built a production-grade, LLM-powered Data Quality Agent within its NEXT Unified Data Platform on AWS. The agent generates context-aware data quality rules, validates them, and produces ready-to-run SQL and Python under human-in-the-loop governance, helping life sciences organizations scale enterprise data quality.

AWS Machine Learning BlogIn-site articleHow Indegene is modernizing and scaling enterprise data quality in life sciences
Chips
Policy

Setting a new standard for responsible AI infrastructure

OpenAI introduces its National Community Compact and Compute Campus Policy Blueprint to protect local resources, create jobs, and guide responsible AI infrastructure development.

OpenAI NewsIn-site articleSetting a new standard for responsible AI infrastructure
Tools

Flipsnack grows product signups with conversion optimization

Flipsnack focused its ChatGPT Ads campaigns on getting people to try its product, growing registrations while lowering the cost per signup.

OpenAI NewsIn-site articleFlipsnack grows product signups with conversion optimization