Translate
Live intelligence

Latest AI updates

1,721 updates
AllModelsProductsIndustryResearchTutorialOpinion
Research

Only Accepted Tokens Count: Energy Accounting for Speculative Decoding Under Latency Targets

The paper measures the energy cost of speculative decoding under production serving conditions on H100 and H200 GPUs using SGLang with CUDA graphs, at batch sizes up to 64 across five corpora. It finds that on dense models verification, not drafting, dominates energy, and that the saving shrinks as batch size grows because batching has already shared the weight read that speculation would divide among accepted tokens. At large batch sizes block drafters push verification past the arithmetic-memory balance point into a compute-bound regime where the saving becomes a loss, and low acceptance rates raise energy per token. Meeting a tighter latency target with speculation therefore costs additional energy, while larger batches lower energy per token without speculation.

arxivOct 30, 2026Score 76
Speculative DecodingLLM Inference ServingEnergy Efficiency
Research Editor selected

Loop Scaling, Specialization, and Parallel Inference for Transformers

This arXiv study examines looping in dense decoder-only transformers, where weights are reused to make models deeper without increasing parameter count. It finds that looping the full transformer gives the strongest quality at a fixed parameter budget, while looping a middle block is better when training FLOPs are constrained; gains can saturate, reverse, or become unstable as loops increase. Loop-specific low-rank adapters improve quality, especially with more loops, and a modified Parallel Loop Transformer formulation recovers quality gains while allowing loop stages to run concurrently.

Why it matters: It provides concrete design trade-offs for weight-reuse looping and a fix for the decoding slowdown that limits parallel loop inference.

arxivOct 30, 2026Score 82
transformer architecturesmodel scalinginference efficiency
Research Editor selected

GDGraph Compiles Radiology Guidelines into Executable Decision Graphs, Lifting LLM Accuracy to 90.8%

Researchers present Guideline-derived Decision Graph (GDGraph), a framework that decouples guideline compilation from case evaluation for radiological risk stratification and management planning. During compilation, a large language model partitions guidelines into section-based chunks, extracts local decision subgraphs with linked definitions, and aggregates related branches into an executable decision graph. At inference, a fast System One model answers localized clinical questions while deterministic traversal resolves Boolean logic, yielding a traceable decision path without task-specific training. Across 347 cases spanning renal cyst, adnexal mass, and thyroid nodule guidelines, GDGraph raises System One accuracy from 57.3% to 90.8%, outperforming retrieval-augmented GPT-4o (80.1%) and graph-guided GPT-4o with chain-of-thought (82.7%) at substantially lower latency and cost.

Why it matters: It shows how scattered guideline prose, tables and cross-referenced definitions can be compiled into an auditable decision path that beats GPT-4o baselines while costing less.

arxivOct 30, 2026Score 84
GDGraphClinical Decision SupportRadiology AI
Research

A Unified Hessian-Guided Framework for Post-Training Quantization

The paper proposes a unified post-training quantization framework for large language models that jointly optimizes quantization ranges, linear transformations, and weight rounding. It derives a Hessian-guided approximation of output changes from joint weight and activation quantization, using a tractable Fisher approximation to form a common output-sensitivity objective. The authors claim the method remains practical for modern LLMs and outperforms state-of-the-art transformation and rounding methods in channel-wise and block-wise settings.

arxivOct 30, 2026Score 74
Post-training quantizationLarge language modelsModel compression
Research

EpiCoach: Auditable LLM Agent for N-of-1 Wearable Physiology

EpiCoach is an LLM-agent system that treats a single cross-country mountain biker as the subject of auditable N-of-1 science, keeping claims in an evidence-gated hypothesis registry that requires a probabilistic prediction about a future session before it occurs. Forecasts come from a mechanistic digital twin built on heart rate, GPS and barometric altitude, combining a lagged heart-rate response model, a fitness-fatigue model with learned time constants and a latent technical-descending skill state measured on discovered trail segments. The agent prescribes each training week as a constrained experiment that maximizes information gain subject to load and safety constraints, and is evaluated over eighteen months of data with a rolling-origin replay benchmark and a live prospective phase. In the first three live weeks the registry used 20 predictions and scored 14, the twin beat standard forecasters on heart rate but not segment time, and the safety officer vetoed the agent twice; the registry, benchmark and protocol are released.

arxivOct 30, 2026Score 72
LLM agents for healthdigital twinswearable sensing
Research Editor selected

OpenConjecture Tests Frontier Models on Randomly Sampled Math Conjectures

Researchers introduce OpenConjecture, a dataset of conjectures drawn from arXiv mathematics papers published since February 2026, and use a random sample of 200 as open problems. Frontier GPT models were run over three passes with a reproducible protocol and no problem-specific mathematical guidance. Models produced candidate solutions for roughly 60% of conjectures, with counterexamples about 1.3 times more common than proofs. The analysis finds some resolutions deviate from author intent, such as proposing the trivial group as a counterexample, and that many proofs follow closely from existing results, though some show at least moderate novelty.

Why it matters: It offers a reproducible, sampling-based benchmark that moves LLM math evaluation past cherry-picked breakthroughs toward fidelity, attribution and novelty of solutions.

arxivOct 30, 2026Score 84
LLM mathematical reasoningbenchmark and evaluation designOpenConjecture dataset
Research

TOME-Math Converts 2,118 Textbooks into 380K Structured Proof-Supervision Records

Researchers introduce TOME-Math, a large-scale corpus that turns 2,118 university-level mathematics textbooks into roughly 380K structured proof-supervision records while preserving chapter order, proposition-proof links, dependency references and source provenance. Fine-tuning on the full corpus lifts the unweighted mean score across held-out evaluations and established public benchmarks by 16.7 percentage points over the base model, and exceeds all three public instruction-data baselines in aggregate. The work also characterizes scaling patterns with supervision volume and source breadth, and explores curriculum learning across mathematical subjects. The authors state the corpus will be released to support further research on structured mathematical supervision.

arxivOct 30, 2026Score 74
TOME-Mathmathematical reasoningtraining data organization
Research

Understanding and Mitigating Role Spoofing in LLMs

This arXiv paper examines role spoofing, an attack in which malicious text—such as content from a tool output or webpage—exploits LLM role confusion to impersonate the user. The authors show that models track speakers using both content cues, including semantics and writing style, and structural cues, including control tokens and formatting, and that attacks combining fake structure and content are highly effective. In a toy setting, adversarial training reduced susceptibility to content-based injections but not structure-based attacks. The paper also proposes detecting role spoofing with linear probes trained to predict conversational roles and reports effectiveness across settings.

arxivOct 30, 2026Score 74
LLM securityprompt injectionrole confusion
Research

SCARL: Sequence Confidence-Aligned Reinforcement Learning for RLVR

The paper identifies a sequence-confidence misalignment in reinforcement learning with verifiable rewards: because the advantage is blind to a policy's confidence, a response the policy barely found and one it produces with near certainty receive the same update, and a confident error is corrected no harder than an exploratory one. SCARL reshapes each advantage, without changing its sign, by the alignment gap between the response's group-relative confidence and the outcome that its advantage certifies, normalized into per-class shares of the advantage budget; it provably reduces to the base estimator when a class's confidences agree and adds no extra passes and only one hyperparameter. On Qwen3-4B-Base it lifts average Pass@32 by 2.4 points over DAPO and is the strongest of nine compute-matched methods on coverage and per-draw accuracy. Applied unchanged across four estimators and four backbones, it improves 14 of 16 settings while preserving policy entropy and widening the confidence-correctness separation; code is released.

arxivOct 30, 2026Score 78
Reinforcement learningLLM reasoningConfidence calibration
Research Editor selected

On-Policy Training Helps Multi-Hop RAG Retrievers Recover From Mistakes

A new paper proposes training chain embedding retrievers on their own retrieval histories, including mistakes, rather than only on gold-standard histories. The method mixes these on-policy states with teacher-forced examples and masks alternative correct passages to avoid penalizing valid continuations. The authors report more complete evidence retrieval and better recovery from errors, with transfer without additional training to harder long-chain variants of HotpotQA and 2WikiMultiHopQA constructed for the study.

Why it matters: The study addresses a concrete reliability problem in multi-hop RAG: retrieval errors that compound when subsequent searches depend on previously retrieved passages.

arxivOct 30, 2026Score 85
Retrieval-augmented generationEmbedding modelsOn-policy training
Research

Neurosymbolic Framework Links Audio Classifier Neurons to Acoustic Concepts

Researchers present a model-agnostic neurosymbolic framework for interpreting the final hidden layer of audio classifiers through symbolic concept induction. The method is evaluated on music genre classification with GTZAN and instrument classification with IRMAS, analyzing a 32-neuron layer. Several hidden neurons were associated with meaningful induced concepts, and most confirmed neuron-concept pairs were statistically supported using the Mann-Whitney U test. The work suggests concept-level induction can explain latent audio representations beyond input-level attribution.

arxivOct 30, 2026Score 70
audio classificationneurosymbolic AIinterpretability
Research

IMPACT: Inverter-Informed Multimodal Learning for Photovoltaic Generation Forecasting

The paper presents IMPACT, a physics-informed multimodal deep learning framework for 10- to 110-minute-ahead probabilistic photovoltaic power forecasting. Its four-branch architecture combines plant-level temporal history, inverter-level spatial cloud dynamics via graph attention, solar geometry, and current meteorological state through gated cross-attention. The decoder anchors predictions to a smart-persistence baseline and applies online kernel-weighted adaptive conformal calibration for quantile forecasts. In reported experiments, the model outperforms all 14 baselines with an FSS of 0.2950, and ablations show every branch contributes positively, including under overcast and variable-sky conditions.

arxivOct 30, 2026Score 72
Photovoltaic power forecastingMultimodal deep learningConformal prediction
Research

NECA Uses Similar Patients' Event Timing to Improve Multi-Horizon ICU Deterioration Prediction

The paper proposes Neighbor Event-Curve Augmentation (NECA), a model-agnostic method that retrieves training-set patients with similar early-window ICU profiles and converts their subsequent organ dysfunction onset times into cumulative event-curve features at 6h, 12h, 24h and 42h horizons. These features are concatenated with the original patient representation and passed to standard downstream predictors, addressing the common practice of treating each patient as an isolated sample. On PhysioNet 2012 and a 4,000-stay MIMIC-IV subset, NECA improves average PR-AUC and ROC-AUC across four organ dysfunction tasks and seven supervised learners, with the largest gain on neurological dysfunction (PR-AUC 0.461 to 0.557 on PhysioNet 2012). Ablations attribute most of the improvement to clinically meaningful retrieval and horizon-specific neighbor risk, while the full cumulative curve adds task-dependent temporal context.

arxivOct 30, 2026Score 71
ICU organ dysfunction predictionretrieval-based augmentationclinical prediction models
Research

Prefix Executability: Evaluating Tool-Using Agents Beyond Final Success

This arXiv paper proposes prefix executability, a trajectory-level evaluation protocol that checks whether every prefix of a predicted tool-call sequence stays executable under tool schemas, state preconditions, and data dependencies. It summarizes reliability with Prefix Executability Curves, First-Error Depth, Trajectory Executability Rate, horizon-normalized AUC-PEC, and Recovery Rate. The protocol is instantiated on stratified samples from BFCL-v3 multi-turn and ToolSandbox, plus the complete Airline/Retail base splits of the tau^3-bench. The authors report that outcome-only and prefix-level rankings diverge, with some plan-conditioned agents scoring higher final success while having much lower trajectory executability, and that explicit pre-execution plan checking can improve executable-prefix survival while keeping terminal success competitive.

arxivOct 30, 2026Score 76
tool-using agentsagent evaluationbenchmark reliability
Research Editor selected

Simulation-based reinforcement learning improves AI psychiatric intake in role-played encounters

An arXiv preprint reports that multi-turn reinforcement learning in simulation improved a language model’s structured psychiatric intake performance compared with its identically prompted base model. In 120 encounters with psychiatrists role-playing standardized patients, independent psychiatrist evaluators preferred the trained agent in 76.7% of matched scenario pairs, with improvements in communication and differential diagnosis. In adversarial simulated presentations, under-triage of acute safety risks fell from 30.5% to 12.9%; the findings do not establish effectiveness with real patients, and residual gaps require safeguards and clinician oversight.

Why it matters: The study provides concrete evidence that simulation-based training can improve psychiatric conversational agents while highlighting safety limitations and the gap between role-played evaluation and real-world care.

arxivOct 30, 2026Score 87
Reinforcement learningPsychiatric intakeConversational AI
Research

FIRME: Registry-Anchored Memory Graphs for Evolving, Heterogeneous Investor Knowledge

FIRME is a registry-anchored memory graph designed to reconcile recurring investor observations into persistent canonical relationships while retaining source, time, and context. It uses rule-based within-period fusion to link investor-call discussions with portfolio cash-flow signals, and cross-period alias memory to direct recurring observations to the same relationship history. Evaluated on proprietary investor-call and cash-flow records and S&P 500 earnings-call transcripts, FIRME improved final-period Correct Match Rate by 7.8–12.1 percentage points over a registry without alias memory across four settings. A blinded human audit also reported higher Correct Match Rate and reattachment recall, and a separate audit found valid evidence paths for 293 of 300 outputs in the investor-call setting.

arxivOct 30, 2026Score 72
AI memoryknowledge graphsfinancial AI
Research Editor selected

MetaRefine Improves Image Quality Assessment with Geometry-Aware Score Refinement

Researchers introduce MetaRefine, a lightweight architecture that refines existing image assessment scores using original-image geometry, multi-level visual features, and the frozen model's score. Across 100 configurations on 21 datasets, it improved PLCC in 88 configurations and SRCC in 85, while reducing MAE in 90 and RMSE in 89. The reported median parameter overhead was 1.11%.

Why it matters: The work offers a practical method for improving image quality assessment models while preserving existing scoring pipelines and adding limited parameter overhead.

arxivOct 30, 2026Score 88
MetaRefineimage quality assessmentcomputer vision
Research

Neural triage decoder speeds high-rate quantum LDPC error correction

Researchers introduce neural triage decoding, a framework that uses syndrome consistency to adaptively allocate neural computation for high-rate quantum low-density parity-check codes. A local neural message-passing network rapidly predicts corrections, additional iterations reduce residual syndrome violations, and a rarely invoked diversification stage searches for consistent corrections. The Triage decoder is reported to handle ultra-high-rate codes encoding a few hundred logical qubits, including lifted-product and Kasai codes, matching near-optimal algorithmic decoders while reducing decoding latency by roughly one to two orders of magnitude. The work also analyzes a syndrome-inconsistent error floor and shows robustness to mismatched error priors.

arxivOct 30, 2026Score 74
quantum error correctionneural network decodersqLDPC codes