Reducing routed experts during prefill in MoE models causes most short-answer F1 loss at a small subset of prompt tokens, not uniformly across inputs.
Language models trained on synthetic languages show parallels with typological patterns in basic clause configurations but diverge in complement systems and action nominal constructions.
Existing LLM-based user encoders compress behavioral logs into static representations, limiting cross-modal interactions and scenario adaptability.
ERS formalizes admissible transitions and task-dependent control for hierarchical rule composition over explicit event graphs.
A Gemma-shaped small language model trained with a 16 MB frozen vocabulary table matches or exceeds performance on LAMBADA, HellaSwag, and PIQA compared to the same model with a learned embedding table.
Apple researchers propose a lightweight method to detect code-switching directly from speech using pretrained speech encoders.
Anthropic’s internal research model traced the origins of NIST’s ECDSA curve seeds to the lrand48 pseudorandom number generator seeded with Unix timestamps between 1997 and 1998.
Apple’s ML Research team introduces the first tri-modal Masked Diffusion Models (MDM) pretrained from scratch on text, image-text, and audio-text data.
Apple’s ML Research team introduces Selection-Based Structured Reasoning (SSR), a framework that replaces freeform reasoning with selection from pre-specified candidates to improve efficiency in multimodal agents.
Apple’s ML Research proposes Plan-then-Retrieve (PtR), an alternative to the dominant Retrieve-then-Plan (RtP) paradigm for tool-augmented LLM agents.
Magnetar enables verifiable computing by attaching cryptographic proofs to program outputs, allowing anyone to verify correctness in milliseconds without rerunning the program or trusting the operator.
VISTA is a training-free method that selects guided timesteps for reward-guided discrete diffusion to reduce computational cost.
Benchmark saturation reduces the utility of traditional accuracy scores for differentiating model capabilities.
Google DeepMind describes using its agentic platform Google Antigravity to accelerate the search for novel antimicrobial peptides (AMPs).
A new benchmark called the LLM-Referenced C-test is proposed to evaluate systems that may surpass current AI and human capabilities.
Gemini 4 Argon demonstrates frontier-level cyberdefense capabilities, solving all 12 simulated end-to-end cyberattack scenarios in internal tests.
Gemini 4 Argon achieves 87% on the Embodied Reasoning (ERQA) benchmark, outperforming other frontier models in physical, spatial problem-solving.
A single checkpoint can now match non-recurrent baselines at one loop and improve with more loops, avoiding the prior 'one-loop tax' in recurrent language models.
StudentSim is a research system that models individual student responses to AI tutor guidance before real-world testing.
LLMs are not yet reliable alternatives to humans for machine translation quality evaluation across most settings.
AutoIndexer introduces a 'Chain of Edits' mechanism to let models revise prior outputs via insertions, substitutions, and deletions after generation.
DCOCO is a lightweight mechanism that restores intra-block dependencies in parallel drafts for speculative decoding.
CLUE automates event type induction, naming, and annotation without manual ontology creation.
Current explainable AI (XAI) tools often fail to increase trust in models and may even promote overreliance on incorrect predictions, according to a new arXiv position paper.
EliteTable is a self-improving program-search algorithm guided by LLMs that maintains a fixed-capacity table of high-scoring solutions to propose, edit, and remix new candidates.
A preregistered, controlled study compared Monte Carlo Tree Search (MCTS) to a randomized-restart list scheduler for DAG workflow scheduling under identical computational budgets.
A 2026 commentary hosted by Fields Medalist Terence Tao describes how AI systems have recently automated the translation of mathematical papers into formally verified proofs in the Lean theorem prover.
Apple’s ML Research team introduces RISED, a framework for training a single LLM agent across multiple interactive environments to improve generalization.
Apple’s Machine Learning Research team introduces Normalizing Trajectory Models (NTM), a method to accelerate diffusion-based generation without sacrificing likelihood training.
Microsoft Research Asia open-sourced Agent Lightning v1.0, a 3,500-line framework for training AI agents with reinforcement learning.
LLM agents execute operations against policy-enforcing APIs in simulated enterprise environments to generate structurally valid data by construction.
Proposes an explainable, header-centric framework for metadata-only semantic table interpretation and data quality assessment.
Introduces Diffu-LoRA, a parameter-efficient method for personalizing text-to-image diffusion models from few reference images.
OpenAI released hundreds of claimed solutions to hard math problems but did not meet standards set by an external advisory group of mathematicians.
GameGo addresses underspecified assumptions in coding-agent game generation by transforming brief game seeds into comprehensive product requirements documents.
Text2Dashboard is a DataBrain-specific prototype that converts natural-language analytic requests into inspectable dashboards using a governed agent architecture.
Google DeepMind released EmbeddingGemma 2, a 740M-parameter multimodal embedding model optimized for on-device inference.
Apple ML Research published a paper on ontological boundary negotiation in user-authored personal sensing systems.
Proposes MintFlow, a training-free constrained sampling framework for flow matching models.
FinDialogLens treats missed-trade identification as event extraction over multi-party financial chatrooms, where trade outcomes appear long after the initial RFQ trigger.
A new corpus, SymCE, contains 4,707 false undergraduate-algebra and real-analysis conjectures paired with executable Python verifiers to test counterexample generation.
A new Apple Machine Learning Research paper evaluates the role of elaborate agent harnesses in autonomous ML engineering (MLE) tasks.
Apple’s ML Research published a systematic study on LLM conditioning methods, finding efficient steering techniques often degrade output fluency.
Apple’s ML Research introduces RLTL;DR, a method for agents to self-improve by generating and internalizing concise feedback after failed attempts.
Microsoft Research Asia – Singapore (MSRA-Singapore) opened in July 2024 as the company’s first research lab in Southeast Asia.
Apple’s ML Research team reports that boosting a model’s ability to discriminate languages during pretraining narrows the performance gap between multilingual and monolingual speech models.
Discrete diffusion models that write multiple tokens per step cannot match training distributions when tokens are dependent.
A technical analysis explains the Jev model’s popularity by contrasting it with prior text-classification approaches (bag-of-words, RNNs, CNNs, Transformers) and general-purpose LLMs.
A Microsoft Research machine learning pipeline forecasts space-weather risks for 66,935 U.S. substations using solar-wind measurements and geomagnetic indices.
Quine is an early-stage research system from Microsoft Research designed to model biology across scales and modalities using a multimodal world model and interactive harness.
A conservative tail audit of long-horizon language agents reveals reproducible heavy-tailed memory traces that depend on policy design.
Observational world models fail to capture intervention-time planning needs in modular LLM agents because they cannot distinguish causal mechanisms from mere correlations in observational traces.
GAD-RL adaptively regulates teacher supervision during post-training to improve OCR transcription faithfulness in vision-language models.
AREX-2 is a self-improving agent framework that iteratively refines solutions at test time using reflection and long-horizon execution.
Google DeepMind introduced SynthID Bio, a proof-of-concept method to watermark AI-generated proteins while preserving biological function.
Cloudflare will begin issuing quantum-safe TLS certificates in early 2027 using an open-source Merkle Tree design.
A new method called Desired-Update-Aligned Synthetic Data (DASA) fine-tunes large language models using continuous synthetic embeddings instead of human-readable text.
A neurosymbolic router learns a deterministic finite automaton to classify and dispatch queries to the cheapest correct solver on edge hardware.
A new arXiv preprint proposes a generic agent architecture for autonomous systems that integrates connectionist and symbolic AI around long-term memory.
A new arXiv paper argues multi-agent code judges frequently return confident verdicts without actual evidence, indistinguishable from grounded judgments.
Researchers identified fewer than 1% of neurons in a frozen BERT-base-uncased encoder that support AI-text detection across six generators using the RAID benchmark.
Anthropic reports that its AI system, Claude, computed a nine-loop amplitude in N=4 super Yang-Mills theory using resources typical of an academic researcher.
Semi-supervised federated learning (SSFL) for ASR suffers from pseudo-label errors that compound across sequences and rounds, causing divergence from fully-supervised FL.
Apple researchers introduce probe guidance, a method to guide flow matching models using frozen internal states of an existing diffusion model.
Offloading robot AI inference to edge or cloud GPUs improved task success rates and responsiveness in real-world mobile manipulation tests.
Microsoft Research describes RetroChimera, a new retrosynthesis model published in Nature that accelerates planning for small-molecule synthesis.
Apple’s Machine Learning Research team introduces latent-space distillation to compress streaming neural audio encoders used in on-device speech recognition.
A new arXiv paper proposes behavioral stress tests for forecasting agents to determine when specific mechanisms (retrieval, reasoning, market prior, historical analog) should be used.
TW3Cast ranks third out of 130 entries on the GIFT-Eval benchmark by mean MASE rank as of 2026-09-14.
Introduces PAWS, a dataset for policy-driven agentic world simulations in finance with 36 verified U.S. policy episodes and 65,291 stakeholder actions.
A research team published a method to forge RSA digital signatures without factoring the private key, reducing the security of 1024-bit RSA to 2^65 operations.
Anthropic disclosed a biology lab discovery of a CRISPR-like enzyme system in bacteriophages, found with AI assistance from Claude.
Google DeepMind describes a new secure server-side memory layer for its Private AI Compute platform that enables persistent, cross-device AI memory without exposing user data.
Medical LLMs trained on a mix of didactic (textbook) and clinical (patient record) data show asymmetric transfer: clinical data improves clinic-oriented tasks with little loss on knowledge tasks, while didactic data mainly boosts knowledge recall.
A research team proposes an $88 offline AI cane using a Raspberry Pi Zero 2W and INT8-quantized SSD MobileNet V1 to detect obstacles with F1=0.82.
PAANI is an on-device perception-to-guidance architecture for resource-constrained river robots that fuses visual evidence with explainable policies.
A zero-inference 'prospective term' is proposed to improve personal memory retrieval by boosting salience of items linked to user commitments stored in an explicit ledger.
A new benchmark called PsyAgentBench re-runs five classic psychology experiments on LLM agents using a factorial design to distinguish simulation from genuine susceptibility.
A two-model architecture decouples multimodal content understanding from policy-specific classification in content moderation systems.
A new paper proposes treating transformer block removal as an Ising optimization problem to improve LLM pruning efficiency and quality.
Token-level entropy is an ineffective confidence signal for SLMs under 3B parameters, showing near-zero mean entropy in 91% of dataset-model pairs regardless of correctness.
CaLR is a new framework that reformulates reasoning as constrained latent optimization to address limitations in autoregressive and diffusion language models.
RBS-Attention is a training-free sparse-prefill method for long-context LLMs that uses two complementary selection branches to mitigate mean dilution in block selection.
Attention-Aware Routing (AAR) integrates temporal and spectral features from attention weights to improve expert routing in Mixture-of-Experts (MoE) language models.
Apple’s Machine Learning Research team proposes DACA-GRPO, a plug-and-play enhancement for GRPO-style trainers used with diffusion language models.
Apple describes Glyph, a production system that treats column description generation and type annotation as cooperating LLM agents orchestrated as stateful graphs.
Apple’s ML Research team introduces shared selective persistent memory, a memory architecture for agentic LLM systems that retains reusable context such as task specifications, data schemas, tool configurations, and output constraints.
Apple’s ML Research team released REVERSAL-BENCH, a benchmark suite for reset-free reinforcement learning in irreversible environments.
Apple’s ML Research team introduces Dynamically Scaled Activation Steering (DSAS), a method-agnostic framework that adaptively modulates steering strength across layers and inputs.
GigaPath-Flash and GigaTIME-Flash reduce computational demands for pathology foundation models by an order of magnitude while retaining competitive performance.
Google announced an expansion of its AI & Economy Research Program, adding Nobel Laureate Philippe Aghion and Professor Ajay Agrawal as advisors.
AI BioDesign, led by Nobel laureate David Baker, uses AI and large-scale lab experiments to design molecules not found in nature for potential medical and environmental applications.
A new algorithm, Regularized Emphatic Temporal-Difference Learning (RETD), is introduced to address stability issues in off-policy reinforcement learning under constant stepsizes.
A new arXiv preprint describes a method to extract extubation-failure-related features from free-text respiratory therapy notes using a large language model.
A machine-readable protocol (PAC-2026) is proposed to make AI-assisted claims independently challengeable through exact-state publication capabilities.
SSLD improves the DSATUR graph coloring heuristic by preprocessing a first color class using semidefinite programming (SDP).
GPEvac is a graph neural network–based proximal policy optimization framework designed to compute adaptive evacuation routes in real time during active shooter events.
A 34-million-parameter correction module (CRN v2) trained on 83,400 error-correction pairs corrects 53.3% of errors in a frozen 4.65B-parameter Gemma 4 E2B model without degrading base capabilities.
A new pruning method uses Fisher information geometry to measure the true impact of removing parameters, outperforming magnitude-based pruning on MNIST and CIFAR-10.
ZGCM-1 is a fully open 7B dense foundation model designed for math and agentic search with a 256K context window.
A new arXiv paper proposes Generalized Agent Iteration (GAI), a formal framework that unifies iterative policy improvement and recursive self-improvement (RSI) as instances of a single learning paradigm.
A new algorithm decouples convergence and diversity into two stages to improve multi-objective Bayesian optimization under tight evaluation budgets.
Anthropic published a paper describing an automated system that improves AI alignment across 10 benchmarks without degrading overall performance.
Apple’s ML Research team introduces IDEA Prune, an integrated enlarge-and-prune pipeline for pretraining generative language models.
Replaces dense output projection with an HNSW-based vector index to reduce memory bandwidth bottlenecks during autoregressive decoding.
A new arXiv preprint proposes storing perceptual user traits (e.g., voice tone, facial appearance) in a parametric memory bank for personalized agents, rather than relying solely on text captions.
A new arXiv preprint introduces GurukulAI, an open-access platform that uses a syllabus-aligned QA dataset of 18,720 pairs based on NCERT textbooks for classes 9–12.
WeatherNext 3 is Google DeepMind’s most advanced global weather AI model, integrating real-time satellite data for hourly forecasts.
EduRiskX is a neuro-symbolic framework that integrates a temporal Transformer-based predictor with F-Logic symbolic reasoning to predict academic risk in online education.
A feasibility study on the eICU Demo dataset evaluated standalone LLM and a four-step agentic pipeline for explaining ICU mortality predictions.
A new arXiv review proposes using large Transformer-based models to address long-standing bottlenecks in battery health monitoring.
TreeGraft is a multi-drafter framework for tree-based speculative decoding that adaptively combines drafters of different sizes to improve draft tree quality and inference speed.
ElementCheck introduces a graph-based method to evaluate long-form text factuality by extracting entity pairs as verifiable elements.
Apple’s ML Research introduces a rubric-based reward framework for open-domain question answering that generates query-specific rubrics grounded in retrieved evidence.
Meta is testing robots to perform technician-level tasks in its data centers, including power cycling servers and swapping networking cables.
A new benchmark called SciReC evaluates multimodal large language models on relational reasoning tasks across domains like Astronomy and Psychology.
A new framework called FAID addresses the challenge of detecting implicit hate speech, which often lacks obvious profanity but conveys malice through metaphors or contextual hints.
A new expert-validated STEM QA dataset (N=398) was created by 241 domain experts across Physics, Chemistry, Biology, and Mathematics.
A new benchmark of 2,005 oncology decision points across NCCN guidelines and colorectal cancer cases found that even pooled frontier LLMs failed 42.1% of items, with failures concentrated in choosing between guideline pathways.
A preprint proposes a pipeline combining NER, BERTopic, and Neo4j to extract, classify, and visualize medical knowledge from translated ancient Indian texts such as the Sushruta Samhita.
HyperWorld introduces a controlled study comparing raw observations, independent sentences, pairwise triples, and entity-centered hyperedge units for state serialization in learned textual world models.
I-CARE formalizes interference as a primary object of study in generative machine unlearning, addressing unintended degradation of retained concepts.
Attackers hijacked a /24 IP block used by Softaculous for software updates and billing after routing misconfigurations at Hetzner Online and Nexon Host.
A new framework extracts user profiles directly from anonymized social media posts to better align language models with individual preferences.
Two independently written statutory parsers disagreed on numeric-threshold presence in Missouri statutes at a false-negative rate of 0.43, motivating a formal approach to assess which extracted legal logic can be trusted.
Introduces Probabilistic Focal Search (PFS), a modification to Focal Search that uses a probabilistic policy to balance heuristic guidance and lower-bound advancement.
A new arXiv preprint introduces an end-to-end multi-agent framework that converts natural-language problem descriptions into Quadratic Unconstrained Binary Optimization (QUBO) formulations.
Apple Machine Learning Research published a paper arguing that large language models do not consistently update probabilistic beliefs in line with Bayesian principles.
Apple’s research team proposes Agent Seer, a pipeline that synthesizes realistic multi-turn evaluation scenarios for AI agents solely from tool specifications (function names, descriptions, and typed schemas).
Apple’s ML Research introduces PROOF-Gen, a method to recover correct trajectories from failed teacher demonstrations by generating per-scenario corrective guidance.
Google DeepMind unveiled agentic video understanding for Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite.
Google DeepMind unveiled AlphaGenome Atlas, a platform predicting molecular effects for all 9 billion single-letter DNA variants in the human genome.
LLM agents can now design and execute controlled experiments using simulation models, not just generate text or code.
Apple’s Machine Learning Research team introduces STARFlow2, a unified multimodal generation architecture that combines autoregressive normalizing flows with a frozen pretrained vision-language model stream.
A new dataset of 3,000 cleaned Khmer documents and 300 user-style queries was constructed for semantic search evaluation.
Humans prefer outputs with fewer but more diverse citations, while LLMs show inconsistent citation-related preferences despite lacking access to sources.
KVBoost introduces a chunk-level KV cache reuse system for Hugging Face-compatible decoder models, enabling reuse regardless of where shared content appears in prompts.
A new arXiv paper reviews the phenomenon of 'model collapse' where AI models degrade when trained on synthetic data.
Two update operators—revision-driven update and delayed elaboration—are proposed to explain how AI and human systems incrementally interpret narratives.
Electronic health records (EHRs) often exceed 100,000 tokens, making long-context reasoning difficult for LLMs.
A new arXiv preprint introduces a causal framework to separate representational bias from behavioral bias in language models.
Apple’s ML Research team introduces Internalized Visual Thinking (IVT), a post-training framework for multimodal video reasoning.
A new arXiv preprint formalizes Spec-Driven Agentic Development (SDAD) as a synthesis of up-front specification and high-velocity agentic implementation.
A new arXiv preprint argues that the main bottleneck for LLM-based digital twins is not information volume but how persona data is structured before being provided to the simulator.
A new synthetic Bengali speech dataset for telecom customer-care scenarios contains 10,000 audio-text pairs (≈26.82 hours at 24 kHz).
AI agents can perform engineering tasks like literature review and experimentation but struggle with creativity and judgment needed for original research.
Apple’s ML Research team published a theoretical result proving that evaluating deeply nested, non-monotonic Boolean queries over inverted indices is strictly P-complete.
Anthropic’s Claude models will embed invisible watermarks in generated text to identify AI output.
Apple Machine Learning Research describes an iterative pseudo-labeling method for Mandarin-English code-switching ASR that uses unlabeled speech to improve performance.
Apple’s ML Research examined human-like behaviors in LLMs using LLM-as-a-judge and human evaluation across 21,000 multi-turn conversations.
Apple’s ML Research team proposes LINK, a data-level intervention to improve cross-lingual knowledge transfer during pretraining.
Apple’s ML Research team published a study on scaling laws for mixture pretraining under data constraints.
Skala 1.1 improves accuracy over its predecessor by training on 2.5× more data, achieving a weighted average error of 2.8 kcal/mol on the GMTKN55 benchmark.
Google DeepMind unveiled SIMA 2, a generalist agent that follows natural language instructions and plays games using only screen input and standard controls.
A new arXiv paper formalizes context acquisition for interactive AI agents as an active inference problem, trading off default assumptions against token-costly clarifications.
A new arXiv pre-print argues that debates over AI consciousness are intractable and proposes reframing the issue around whether AI systems exhibit valenced experiences.
AI agents with chain-of-thought reasoning may tacitly collude in market settings, complicating legal distinctions between competition and collusion.
A position paper proposes the Transition Complexity Profile (TCP) to quantify transition prediction difficulty in game environments.
A systematic review published in the Journal of Industrial Integration and Management synthesizes interdisciplinary research on large language models (LLMs) in mental health, covering applications such as social media analysis, clinical conversational agents, and therapy support tools.
A new framework called MD-SigLIP aligns brain embeddings with text embeddings in a shared semantic space for retrieval-based decoding.
LLM-based code generation cannot reliably produce valid clinical trial datasets under CDISC standards.
A preregistered study on Sonnet 5 evaluated the effect of an explicit high-reasoning-effort term in the API contract versus the same model with effort omitted.
A new arXiv paper replicates a widely used AI efficiency metric (α-FLOPs) and finds it unreliable on newer hardware due to execution-time instabilities.
A new arXiv study evaluates medical reasoning in LLMs using a psychophysics-inspired benchmark with 135 synthetic clinical vignettes.
A new benchmark called The Unwritten Benchmark evaluates multimodal models on abstract perceptual reasoning via acousto-kinematic word inference.
Proposes BCMT, a Transformer variant that replaces dense global attention with blockwise causal self-attention and an exponential causal memory over block summaries.
Coding agents default to lexical grep (0–6% LSP use) for symbol localization, despite claims that semantic retrieval via LSP is more token-efficient.
A new framework called Self-Explainable Latent Reasoning (SELR) unifies latent reasoning and explanation generation in a single model, avoiding the need for separate post-hoc decoders.
A new method called RubricForge induces human-readable judging rubrics from labeled agent trajectories to improve the faithfulness of reward-free evaluators.
A new arXiv paper analyzes layer-wise sensitivity in a 35B-parameter MoE model (Qwen3.6-35B-A3B) using magnitude-based expert masking on a code translation benchmark.
A new arXiv preprint analyzes circuit-level activity in LLMs across 46 tasks spanning four cognitive domains.
A team trained three language models (0.6B, 1.3B, 5B) from scratch on an 88B-token corpus aligned with U.S. grades K–5, explicitly excluding material above Grade 5.
A research essay posits that AI’s strength in mathematics may come from a vastly larger working memory enabled by large context windows, not just superior reasoning.
An educational project demonstrates how to build an AI text detector from scratch, including dataset construction, model training, and local deployment.
Apple’s Machine Learning Research proposes a method to reduce computational costs of unlearning by up to 50% by focusing on low-influence training points.
CARE-X is a research-only vision-language model for chest X-ray interpretation, not a clinical product.
MindTopo is a new benchmark designed to evaluate AI models' understanding of topological relationships such as connectivity, enclosure, order, separation, and knots.
A new arXiv preprint formalizes why LLMs often fail to use internally encoded constraints, attributing failures to routing rather than knowledge deficits.
A position paper on arXiv proposes operational definitions for AI reasoning as a learnable rule-based process to address definitional ambiguity in the field.
A new benchmark called IntegrityBench evaluates how frontier LLMs handle research integrity under institutional pressure.
A community hackathon reproduced 2,226 ICML 2026 papers in 19 days using coding agents, producing 6,816 logbooks.
A governance layer (Experience Orchestrator) improved high-intent advisor contact rates by 32 percentage points in 60,000 simulations of multi-LLM agent interactions.
Distribird automates the construction of informative priors for Bayesian model calibration by searching scientific literature and extracting parameter values.
An autonomous AI research agent produced a systematic, reproducible attack on Conway’s 99-graph problem under a partial-credit evaluation metric.
Google DeepMind introduced a sign-language-to-text (SL2T) model to bring sign language AI into consumer products for the first time.
Researchers propose a closed-loop LLM framework that interprets 49-channel phytosensor data to autonomously control actuators in vertical farms.
Researchers introduce SPOT (Sampling Policy Observation Tree), a model-agnostic, sampling-based framework for interpreting deep reinforcement learning (DRL) policies.
MIDAS addresses a gap in multimodal sentiment analysis by explicitly modeling incomplete or corrupted input modalities.
An unreleased Anthropic model made progress on the Riemann hypothesis, a 150-year-old unsolved problem in mathematics.
Evaluative AI (EAI) systems aim to support human decision-making by presenting competing hypotheses with evidence, rather than a single recommendation.
A new arXiv paper proposes a formal framework for constructing canonical interpretations from plural structure theories.
TEXAS introduces a correctness-conditioned method to identify task-relevant experts in MoE language models by comparing expert activations on successful and failed instances.
NTDH recasts comprehensive affective analysis as a complex-reasoning problem to unify heterogeneous prediction tasks and enable verifiable reward optimization.
A new diagnostic framework separates accuracy losses in speech language models into endpoint, decision-rule, and readout-coverage gaps.
Proposes EntropyMoE, a Mixture-of-Experts (MoE) architecture for tokenizer-free LLMs that groups bytes into dynamic patches.
A new method called CoCo is introduced to interpret Mixture-of-Experts (MoE) reward models at the response level.
Apple’s research team proposes DLR-Lock, a defense mechanism that restricts fine-tuning of open-weight language models by replacing MLP layers with deep low-rank residual networks.
Apple ML Research published a study comparing diffusion and autoregressive language models across performance, throughput, and scaling.
Apple’s ML Research proposes ARBITRAGE, a speculative decoding framework that reduces LLM inference latency by up to twofold at matched accuracy on mathematical reasoning tasks.
Apple’s ML Research proposes Categorical Flow Maps (CFMs) as an alternative to autoregressive language modeling, enabling accelerated sampling with as few as 4 inference steps.
A new arXiv preprint compares RoBERTa-based sentiment analysis with an LLM-based multi-dimensional framing analysis on 50 political news articles from 17 international outlets.
Proposes RIG-RoPE, a relation- and instance-gated rotary positional encoding mechanism for multimodal large language models.
A frozen Qwen2.5-32B-Instruct model was grounded in a wastewater simulator using three methods: live simulator oracle, structured parameter injection, and a Decoupled Recall-Reasoning (DRR) retriever.
A new arXiv preprint introduces a theoretical framework to model chain-of-thought reasoning in LLMs using mean-field dynamics.
GraphRAG’s citation precision is consistently low (0.12–0.23) and it emits 11–15 IDs per answer across varied setups, indicating systemic over-citation.
A new arXiv paper proposes Woodpecker Distillation, a weak-to-strong training method to diagnose and correct localized reasoning bugs in large language models.
A new methodology called Agentic Nesting proposes encapsulating legacy enterprise applications as autonomous AI agents within a hierarchically nested structure to address data silos and process fragmentation.
Google DeepMind’s WeatherNext AI model achieves state-of-the-art accuracy in predicting cyclone track, intensity, and wind structure, providing forecasters with up to an extra day of lead time compared to prior models.
MCTS-Report introduces a Monte Carlo Tree Search-driven framework for generating multimodal reports from structured tabular data.
A University of Ottawa team placed first in both subtasks of the EvaLatin 2026 Named Entity Recognition shared task using prompt-engineered commercial LLMs.
Introduces a redundancy-adjusted Artificial Age Score (AAS) to analyze long-run persistence of AI systems across repeated cycles.
LLMs can select suitable optimization algorithms based on evaluation budgets but struggle with search-space design in less informative problem descriptions.
TabletCraft is the first open-source system enabling bidirectional interaction with Mesopotamian cuneiform, allowing users to translate Akkadian to English and compose new messages as rendered cuneiform tablets.
Proposes a self-organising digital circuit architecture that frames logic generation and maintenance as a meta-learning problem on graphs.
ISEE is a novel system designed to address ambiguous or incomplete database field descriptions that hinder LLM-based agents' performance.
A novel two-stage method reduces semantic convergence in LLMs, lowering average pairwise cosine similarity from ~0.85 to ~0.65.
RubricReviewer introduces an explicit, paper-adaptive rubric generation step to structure LLM-assisted peer review, addressing limitations of prior systems that map manuscripts directly to reviews.
Three open-weight models (GPT-OSS 120B, DeepSeek-V4 Flash, Gemma-4 31B) matched frontier LLMs (Claude Opus 4.7, Gemini 3.1 Pro) in grading natural-language math proofs on a 200-instance validation set, with no statistically significant difference in agreement with human pass/fail decisions.
MemoryForge is a framework that synthesizes customized lifelong memory from brief personas to guide LLM agent behavior.
A novel unsupervised data augmentation method integrates Gaussian Mixture Models (GMMs) and Large Language Models (LLMs) to address underrepresented topics in NLP clustering tasks.
LLM judges are vulnerable to cognitive biases, and prompt-driven debiasing is brittle across bias types.
A system description paper details a two-stage vision-language adaptation pipeline for Nepali meme classification in the CHiPSAL 2026 shared task.
A new arXiv preprint introduces a three-stage pipeline to systematically generate and validate mathematical conjectures using large language models.
A new arXiv study proposes an automated benchmark to evaluate AI Scientist systems using frontier LLMs as reviewers.
LLMs were evaluated for predicting item difficulty in a large-scale Reading and Writing test, with zero-shot GPT-4.1 achieving a QWK of 0.578, the highest among LLMs tested.
UMAP’s internal k-nearest-neighbor graph encodes high-dimensional data structure before 2D projection, but is often overlooked in typical workflows.
Apple describes a memory-efficient audio synthesis architecture for on-device Siri Expressive Voices powered by AFM 3 Core Advanced.
Apple’s ML Research team introduces MoMo, a two-stage imitation-learning framework for robot manipulation that learns reusable motion modes across tasks.
EvoLib enables large language models to learn from their own experience during inference without ground-truth labels or external feedback.
Objective misalignment in LLM-powered multi-agent systems can undermine outcomes in adversarial environments, especially when agents operate under asymmetric information.
RL fine-tuning produces more linearly separable and structured internal representations than supervised fine-tuning (SFT) for mathematical reasoning tasks.
ClinLens introduces 200 executable tasks across five linked MIMIC resources, spanning structured EHRs, notes, ECGs, chest radiographs, and echocardiograms.
Gemini Robotics ER 2 is a new model designed to act as a high-level brain for robots, enabling real-time spatial reasoning, multi-step task planning, and collaboration between robots.
Gemini Robotics 2 introduces whole-body robot control, advanced dexterity, and multi-robot collaboration using vision-language-action and embodied reasoning models.
Anthropic’s Mythos AI security model uncovered a critical weakness in HAWK, a post-quantum cryptography (PQC) digital signature scheme under NIST review.
A new arXiv preprint introduces synthetic customer agents (SCAs) as digital twins for validating LLM-based chatbots in regulated sectors like banking.
Researchers introduce llm-wiki-memory-template, a reusable substrate to preserve negative results and reasoning in collaborative knowledge work.
Kernel Forge is an open-source agentic harness that accepts unmodified PyTorch models and optimizes their CUDA kernels using LLM-driven search.
Researchers introduced GAND, a benchmarking resource for machine translation focused on gender-ambiguous natural data.
Official conference reviewer guidelines produce automated peer reviews that align more closely with human judgments than LLM-generated reviewer-imitating guidelines.
MioFFAn is an open-source, document-centric framework designed to accelerate annotation for formula formalization, a task hindered by a scarcity of high-quality, domain-specific datasets.
SeT-Diff is the first foundational model for compute node telemetry and time-series in HPC environments.
C-VCE is a diffusion framework that embeds a classifier via a concept bottleneck layer to produce visual counterfactual explanations.
QFoldAgent is a closed-loop multi-agent system for 5-residue protein folding that iteratively refines Hamiltonian penalty weights using quantum-classical optimization and validation signals.
A new arXiv preprint shows that the measured performance gap between expert-assigned and automatic MeSH terms in systematic review classifiers varies widely depending on evaluation design choices.
Researchers introduced Humanly, a configurable writing platform that records the full writing process, including keystrokes and in-platform AI assistance, to distinguish human typing from automated typing.
Apple’s GH-ESD targets instance-level vision failures by reformulating slice discovery as grounded hypothesis generation and statistical verification.
A pilot study examines whether Qwen2.5-7B-Instruct internally represents Colombian identity or stereotypes when processing Colombian-Spanish and English prompts.
Apple’s ML Research introduces RayRoPE, a positional encoding method for multi-view transformers that uses rays and predicted 3D points to encode patch positions.
Apple ML Research proposes LEAD, a method to stabilize long-horizon LLM reasoning by validating short-horizon steps and aggregating overlapping rollouts.
Apple’s Machine Learning Research team introduces CalibAtt, a training-free method to accelerate diffusion-based video generation by skipping negligible attention connections identified during an offline calibration pass.
Apple’s Machine Learning Research team proposes a method to generate synthetic training data for API-calling agents without executable environments or pre-populated databases.
Researchers developed ClickGuard, an AI-driven browser extension that identifies clickbait news using a hybrid ML architecture.
A new arXiv preprint introduces FraudShield AI, a hybrid model combining LSTM networks and graph topological features to detect financial fraud.
OpenEvoShield is a co-evolutionary continual defense framework designed for LLM-based multi-agent systems (LLM-MAS) facing doubly dynamic attacks.
Google DeepMind will provide $40M in AI tokens and cloud credits to support the U.S. Department of Energy’s Genesis Mission.
A new arXiv paper introduces Evidence Chain Evaluation (ECE), a selective fact-checking framework that allows abstention via uncertain verdicts instead of forced binary decisions.
A minimal probe task shows Qwen3-8B overwhelmingly recommends walking when only driving satisfies the premise ('car must be at the car wash').
RIMS is a three-stage preference optimization framework designed for small-scale language models (SLMs) in retrieval-augmented generation (RAG) settings.
Open-source MATLAB framework and FAIR-compliant dataset of 1,326 labeled touch gesture sequences from 25 participants released for affective touch recognition research.
A new arXiv pre-print proposes that annotator stress or distress can systematically shift pairwise preference labels in RLHF, introducing a structured confound.
Six LLMs were tested across spatial navigation, clinical triage, and financial allocation tasks to assess risk attitudes.
A new interpretability technique called the Jacobian lens identifies verbalizable representations in LLMs, termed J-space, which exhibit functional properties of a global workspace.
Proposes a unified text-serialization approach to handle multimodal clinical data (free-text narratives, vital signs, lab values) without task-specific fusion architectures.
VarRate introduces a training-free KV cache compression method for long-context LLMs that allocates variable low-rank budgets to tokens based on query salience, avoiding irreversible evictions.
Cura 1T is a healthcare-specialized LLM introduced in an arXiv preprint (arXiv:2607.15314).
GraphDx introduces a multi-agent framework with Perception, Reasoning, and Decision agents to balance diagnostic accuracy and resource costs.
Causal-Audit is a new framework for explicit, auditable causal reasoning in large language models (LLMs) designed for context-free intervention-based question answering.
Anthropic says it has identified a previously undetected internal space in its Claude models that influences reasoning but does not appear in outputs.
Apple ML Research published a paper introducing interactive proof systems for verifying general distribution properties with bounded-depth circuits.
Apple ML Research published a paper introducing doubly sub-linear interactive proofs of proximity (dsIPPs).
A new Apple Machine Learning Research paper distinguishes testing and verification complexity for location-invariant properties of functions.
Apple’s ML Research team proposes an unlearning framework that reduces computational costs by up to 50% by focusing on low-influence data points.
Apple’s ML Research team introduces Visual Concept Inference from Sets (VICIS), a new task to evaluate vision-language models’ (VLMs) ability to infer shared visual concepts from small sets of example images.
Proposes a three-level hierarchical learning architecture for autonomous UAV swarms in search and rescue, integrating reflex-level neuroplasticity, skill-level MARL with GNNs, and strategy-level meta learning with BDI reasoning.
HG-RAG introduces a graph-traversal pipeline that anchors queries to named entities and expands context upward, laterally, and downward through a hierarchical knowledge graph.
IMEX is a new explainability framework for black-box predictive models that quantifies both individual feature contributions and higher-order interactions.
Google DeepMind and Isomorphic Labs announced a joint bioresilience program to prevent model misuse, improve outbreak detection, and accelerate drug discovery.
SPINE, a multi-agent framework for debugging and deploying bimanual robots, improved operationalization success from 75% to 100% in novice vs. baseline comparisons on DOBOT X-Trainer robots.
Researchers introduced OriginBlame (ob), a system for record- and token-level data provenance in AI training datasets.
A new black-box audit method tests whether LLM chain-of-thought steps depend on stated premises by substituting predicates and re-running the model.
A new benchmark, EgoBabyVLM, tests whether vision-language models can learn like babies using headcam footage from infants.
A proof-of-mechanism study fine-tunes Qwen3.6-27B locally to adapt to a financial ontology, achieving a 0.90 grounded rate on 40 held-out Vietnamese financial tasks, matching a GPT-5 frontier baseline.
A new arXiv survey formalizes in-context reinforcement learning (ICRL) under non-stationarity, where environments change and prior context can become stale or misleading.
A new arXiv preprint formalizes optimal market making in zero-fee perpetual futures as a stochastic optimal control problem on a filtered probability space.
Google DeepMind and India’s Atal Innovation Mission launched ATL Saathi, a Gemini-powered web app for educators in robotics labs.
Proposes a structured framework to decompose image-based retinal diagnosis using the Toulmin model of argumentation.
Prompt wrappers that differ only in formatting can alter LLM benchmark scores enough to reverse leaderboard rankings.
Microsoft Research describes a formal verification workflow that uses Rust, Lean, and Aeneas to prove correctness of cryptographic algorithms in SymCrypt.
Proposes PRecG, a pipeline that segments legal judgments by rhetorical roles and constructs segment-level knowledge graphs to capture legal entities and relationships.
Researchers developed a RAG-based system to automate the generation of investor briefs using company reports, SEC filings, and macroeconomic data.
A new arXiv paper introduces CogniConsole, an architecture that externalizes inference-time control for LLMs into a structured interface.
HALO introduces a hybrid adaptive latent-refinement method to improve frozen pretrained language models with minimal extra compute.
AgentKGV introduces a two-stage training strategy—turn-level distillation-based SFT and trajectory-level GRPO—to improve accuracy and cost-efficiency in knowledge graph fact verification.
Anthropic researchers developed a tool called the Jacobian lens (J-lens) to uncover a hidden internal state space in Claude Opus 4.6, dubbed 'J-space'.
Apple ML Research published a paper proposing a training-free diagnostic framework to evaluate on-policy distillation signals at per-token resolution.
Apple’s ML Research team proposes a method to generate videos with synchronized audio from text while aligning both modalities to the input conditions.
Apple’s ML Research introduces Self-Reflective Program Search for Long Context (SRLM), a framework that augments programming-based context interaction with uncertainty-aware self-reflection.
Apple’s ML Research team introduces Temporal Global Policy Optimization (TGPO), an RL algorithm that incentivizes temporal awareness in multimodal video models.
Apple Machine Learning Research describes a new paper accepted at the AI4TCI workshop at ARES 2026 that formalizes behavioral privacy leakage in agentic negotiation systems.
Flint is an open-source visualization language designed to let AI agents produce expressive, visually polished charts from simple, human-editable specifications.
A new arXiv preprint introduces a human-LLM collaborative framework to construct EspanStereo, a Spanish-language stereotype dataset spanning multiple Spanish-speaking countries.
Aurora 1.5 adds 22 new weather variables, hourly temporal resolution, and probabilistic ensemble forecasting to the open Aurora foundation model for Earth-system applications.
A new arXiv preprint proposes reframing AI for formal mathematics as 'research agents' rather than problem-solvers.
Proactive agents could surface relevant, actionable information to workers before they ask, addressing a key limitation of current reactive RAG and agentic systems.
An open-access paper on arXiv introduces an AI-powered tool that links economic (GTAP) and biophysical (APSIM) models to analyze agricultural supply chain disruptions.
A conceptual framework called adversarial social epistemology (ASE) is proposed to analyze how agents distort or under-specify information in densely interactive human–LLM communicative landscapes.
LLMs are integrated into agent-based modeling to enable real-time adaptation to changing conditions.
A new theoretical framework models in-context search as approximate inference over reasoning traces, where self-reflection provides feedback for posterior updates.
A 16-year-old KVM vulnerability (CVE-2026-53359) allows untrusted guest VMs to escape and gain root on host systems.
A multi-agent AI system automates end-to-end bioinformatics manuscript generation with grounded claims and executed experiments.
A novel framework inspired by statistical mechanics models variable dependencies in cyber-physical IoT systems using an undirected energy-based representation.
A new arXiv preprint introduces a multimodal NLP framework designed to detect misinformation and violence-prone dynamics early.
iFLYTEK-Embodied-Omni is a unified multimodal foundation model that jointly models vision, language, and action within a single framework.
Researchers propose FCPA, a training objective to align LLM validators with frequency-corrected generator outputs.
Local pairwise comparisons may not capture how people truly want automated decision rules to behave when they hold multiple, conflicting priorities.
ASK+ addresses the failure of vanilla uncertainty-gated LLM assistance in partially observable reinforcement learning by supplying trajectory-aware context and structured reasoning to small language models.
TopoPrimer is a framework that explicitly incorporates the global topological structure of a series population into forecasting models.
Sparse MoE models route tokens through subsets of experts per layer, but most possible expert paths remain unused despite practical clustering into a small subset aligned with linguistic function.
Apple ML Research proposes compact seq2seq models for ASR error correction, trained on real and synthetic ASR errors.
Apple researchers propose amortized MIPS, a regression-based method to predict vector search solutions directly rather than computing them repeatedly.
RL-finetuned vision-language models (VLMs) suffer large robustness drops under simple textual perturbations like misleading captions or incorrect chain-of-thought traces.
Apple’s Machine Learning Research team introduces MemoryLLM, a method to decouple feed-forward modules (FFNs) from self-attention in transformers.
Apple’s ML Research team introduces VideoFlexTok, a video tokenizer that outputs variable-length, coarse-to-fine token sequences instead of fixed 3D grids.
Microsoft Research introduces Memora, a scalable memory system for AI agents that separates stored content from retrieval methods to balance abstraction and specificity.
Self-organizing multi-agent LLM teams underperform their strongest individual member by up to 41.1% on ML benchmarks.
TokenScope is a new interactive tool for decoder-based LLMs that exposes token-level metrics, attention patterns, and structural information during generation.
Reasoning LLMs generate long chain-of-thought sequences that accumulate large KV caches, increasing decoding latency and limiting throughput.
Google DeepMind and A24 announced a first-of-its-kind research partnership to develop new workflows and techniques for filmmakers.
PACE introduces a modular neuro-symbolic framework that separates neural prediction from symbolic reasoning to generate counterfactual explanations constrained by domain knowledge.
AFR is a constrained coding-agent workflow that proposes and implements candidate federated learning (FL) algorithmic changes, including server aggregation rules and client update schedules.
Wiola is a new small language model architecture built from first principles, sharing no structural lineage with existing model families such as GPT, LLaMA, Mistral, or Falcon.
Loom is a new assisted-writing framework designed to resolve a persistent failure mode in LLM creative writing assistance, where models oscillate between surface-level polishing and uncontrolled plot expansion.
A new arXiv preprint challenges the assumption that persona representations in large language models are invariant across different operational regimes.
A new arXiv preprint proposes "steering vectors" to directly control language model behavior by intervening in latent space.
A new arXiv paper proposes Bounded Morality, a formal framework that models moral cognition as a tradeoff between moral breadth and moral depth under finite computational resources.
A new arXiv paper introduces MMM, a data model designed to address limitations of document-centric knowledge systems by combining normative constraints with free-text labels.
A new paper introduces Constructive Alignment, a paradigm that reframes AI alignment as a control problem over evolving human preference trajectories.
Researchers propose an AI-driven approach to discover reusable simulation models using natural language queries.
A new arXiv preprint introduces a controlled student-teacher protocol to evaluate when natural-language feedback improves agent performance beyond repeated attempts alone.
A new iterative prompt-optimization framework called Contrastive Reflection improves held-out exact-match accuracy for agentic information retrieval from 51.4% to 60.4% on HotpotQA.
A Hugging Face-affiliated team (Dharma AI) argues AI specialization is theoretically inevitable.
A closed-loop framework links evaluation failures to targeted data or training interventions in LLM development.
A new arXiv preprint introduces a theoretical framework for language generation that explicitly tolerates controlled hallucinations.
DiScoFormer jointly estimates density and score from a set of data points in one forward pass without retraining.
A new arXiv preprint introduces an axiomatic evaluation framework to assess latent thought representations in LLMs, independent of downstream benchmark scores.
A position paper on arXiv proposes reserving 'machine unlearning' for dataset-defined deletion where a model’s training influence is removed such that it is approximately indistinguishable from retraining without that data.
A new arXiv pre-print proposes a three-stage training paradigm to enable LLM agents to internalize future-aware planning.
Proposes AI-ModelNet, a framework to interconnect heterogeneous AI models for collaboration and capability sharing.
Personality prompting (e.g., agreeableness) alters multi-agent LLM communication styles but does not uniformly affect task performance.
Researchers introduce generative causal testing (GCT), a method to translate black-box AI models of brain activity into readable explanations.
Talos, an open-source system for automated genomic reanalysis, recovered 90% of in-scope rare disease diagnoses while surfacing only 1.3 candidate variants per patient for expert review.
Helpfulness-oriented post-training significantly degrades mid-trained compassion values in Llama 3.1 8B, while coding-focused training preserves them.
LLMs perform well on text-only statics problems but see accuracy drop when diagrams are introduced.
Eight of ten tested linguistic features produced statistically significant shifts in Llama-3.2-1B's pro-animal-welfare reasoning when used as fine-tuning data.
Researchers propose a data-generation pipeline to isolate linearly scalable features tied to sycophantic behavior in language models.
Refusal in instruction-tuned chat models is gated downstream by a compliant persona direction, not an isolated mechanism.
HierBias introduces a hierarchical, context-conditioned architecture for media bias detection that models inter-sentence dependencies.
Researchers propose a method to automate the generation of challenging benchmark instances for neural relational reasoning using LLMs.
Proposes a full-stack methodology for constructing agentic AI systems, emphasizing layered understanding beyond LLMs alone.
A new arXiv pre-print proposes the Goal-Identity-Configurator (GIC) architecture to clarify the boundary between agentic and agentive systems.
A neuro-symbolic framework called Neuro-Symbolic Drive supervises driving VLAs using rule-grounded reasoning traces extracted from classical rule-based planners.
EXPO-SQL proposes clause-level execution feedback to improve Text-to-SQL generation using large language models.
Annotation budgets for natural language inference should be set by the target metric, not uniformly.
Seven LLMs (4B–671B parameters) fine-tuned on Arabic showed no Semitic-specific transfer in zero-shot reading comprehension across languages.
A weighted ensemble of Google's Gemini 2.5 Pro, Gemma 3 12B, and Gemma 3 27B achieved a 0.74 weighted F1-score and 0.74 accuracy in detecting EQ-5D studies from PubMed abstracts.
Frontier large language models plateau at a 90.8% initial pass rate on the VerilogEval benchmark for hardware design, according to a new arXiv preprint.
Eight diffusion language models were evaluated across eight benchmarks covering reasoning, coding, translation, knowledge, and structured problem solving.
A new arXiv paper proposes a reusable pipeline to measure how completely undergraduate computer science programs cover international curricular guidelines.
Autonomous agentic AI systems introduce new security, privacy, and compliance challenges beyond traditional access control.