Automated AI newsroom

AI News

High-signal AI developments, independently scored, summarized and organized for fast reading.

Live intelligence

Editor-selected AI developments

330 updates

Research Editor selected

Logprob Reduction Fails to Verify Causal PII Suppression in LLMs

This paper shows that SAE-based feature interventions for suppressing PII leakage in LLMs, when evaluated only by log-probability reduction, often produce false positives. Across 36 cases, 47% appeared successful under logprob-only evaluation, but only 2.8% passed a multimetric framework with random direction controls. The authors identify two failure modes and propose causally grounded PII suppression via circuit-level localization and causal mediation analysis.

Why it matters: Provides concrete evidence that a common evaluation metric is unreliable)Skip and proposes a more rigorous methodology, directly impacting AI safety research.

arxiv1d agoScore 85

sparse autoencodersinterpretabilityprivacy

Research Editor selected

Kierston & KierstonBench: Simulated City Benchmark for AI in Crime Investigation

The paper introduces Kierston, a city-scale simulation of Allentown, PA, and KierstonBench, a benchmark for evaluating language models in crime investigation and surveillance tasks. It generates realistic records from a common behavioral timeline and tests models on 32 task families. Evaluation of 16 models shows most fail, while a few frontier models solve most tasks but still fall for deceptions.

Why it matters: Provides a novel, concrete benchmark for a critical AI application area, with clear evaluation results and practical implications.

arxiv1d agoScore 85

AI safetycriminal investigationlanguage model evaluation

Research Editor selected

Dependence-Aware Belief Propagation for Unsupervised Truth Discovery Among Correlated AI Agents

This paper introduces BP-TD, a method for truth discovery among LLM agents that accounts for correlated errors from shared base models. It uses per-agent decorrelation weights in mean-field variational inference, outperforming majority vote on MMLU without labels and ranking agents by true accuracy. The method's benefit depends on the pool composition, and with a few labels, selecting the best agent beats all aggregators tested.

Why it matters: Provides a concrete, novel algorithm with empirical results addressing a practical problem of correlated AI agents, valuable for AI system design and evaluation.

arxiv1d agoScore 85

AI agent reliabilityEnsemble methodsVariational inference

Research Editor selected

Asymptotic Optimality of cμ-Rule for LLM Inference Scheduling under Memory Constraint

This paper studies LLM inference scheduling as an online queueing problem with finite KV-cache memory. The authors develop a fluid model that reduces the system to a multiclass single-server queue and derive a cμ priority rule where effective service rates account for computation, memory access, and memory capacity. They prove the rule minimizes cumulative holding cost in the fluid model and, for finite workloads, also minimizes completion time and maximizes output-token throughput. Experiments on an NVIDIA H800 show practical implementations reduce average latency while maintaining high throughput.

Why it matters: Provides a rigorous theoretical foundation and practical validation for a scheduling policy that improves LLM inference efficiency under memory constraints, directly relevant to serving systems.

arxiv1d agoScore 85

LLM inference optimizationqueueing theorymemory management

Research Editor selected

AgenticFinBench: Benchmarking Financial Agents on Diverse Real-World Tasks

AgenticFinBench introduces a benchmark of 230 real-world financial tasks across 8 categories, 2 languages, and 4 complexity levels, with a reference-free evaluation framework that audits hallucinations. Experiments across 18 agent configurations show hallucination auditing lowers the baseline leader's score by 7.58 points, and harness differences explain more variation in evidence grounding than model differences. The benchmark aims to advance financial reasoning and factual verification for reliable agent responses.

Why it matters: Provides a concrete, recent benchmark with novel evaluation methodology and actionable findings for financial AI agents, useful for researchers and practitioners.

arxiv1d agoScore 82

financial AIagent evaluationhallucination auditing

Research Editor selected

LoCo: Local Preference Learning for Reasoning Compression

LoCo is a new method that reduces the inference cost of large reasoning models by learning when to stop reasoning. It detects overthinking states, regenerates shorter verified continuations, and trains with DPO on state-conditioned preference pairs. On math and code benchmarks, LoCo cuts completion length by up to 45% while maintaining or improving accuracy, and transfers from math to code.

Why it matters: This paper presents a concrete, novel technique with strong empirical results and cross-domain generalization, offering practical value for reducing AI inference costs.

arxiv1d agoScore 82

reasoning modelsinference efficiencypreference learning

Research Editor selected

Natural Textual Unlearnable Examples for Large Vision Language Models

This paper introduces the first unlearnable example (UE) method specifically for large vision-language models (LVLMs), called Natural Textual Unlearnable Examples (NTUE). NTUE perturbs image labels subtly to create a shortcut between an identity and a target, misleading LVLMs without adding misinformation or harming readability. Experiments show NTUE outperforms traditional UE methods, achieving high success with a poison rate as low as 5%.

Why it matters: Provides a novel, practical defense against data misuse in multimodal AI, with concrete results and a low poison rate, valuable for privacy and security.

arxiv1d agoScore 82

Large Vision Language ModelsUnlearnable ExamplesData Privacy

Research Editor selected

PaySeqGen: Interpretable Inverse Reinforcement Learning for Fraud Sequence Generation

PaySeqGen is an interpretable inverse reinforcement learning framework that reconstructs fraud behavior policies from observed payment trajectoriesholistically, generating new fraud sequences for detector augmentation. It learns interpretable behavioral rewards capturing fraudsters' preferences across transaction contexts)Skip, and experiments at Ant International show it outperforms state-of-the-art generative augmentation methods. The framework also enables reward-based trajectory analysis to identify multistep fraud tactics in real payment systems.

Why it matters: This paper presents a novel application of inverse RL to payment fraud, with real-world deployment and interpretability benefits, offering practical value for fraud detection and analysis.

arxiv1d agoScore 82

fraud detectioninverse reinforcement learningpayment security

Research Editor selected

HTCE: Topological Cross-Scene Experience Reuse for Object Navigation

HTCE is a homeomorphism-inspired framework that canonicalizes and reuses navigation experience across scenes for embodied object-goal navigation. It compresses stored route descriptions by 17.7–23.6% without performance loss and achieves state-of-the-art success and SPL on HM3D and MP3D benchmarks under a frozen vision-language policy. The method improves success rate by up to 4.9 points and SPL by up to 2.8 points.

Why it matters: This paper presents a novel topological approach to experience reuse in embodied AI, with concrete benchmark gains and open-source code, offering practical value for navigation research.

arxiv1d agoScore 82

Object Goal NavigationTopological LearningExperience Reuse

Research Editor selected

MambaPFN: Amortized Nonlinear Variable Selection with Prior-Data Fitted Networks

The paper introduces MambaPFN, a Transformer-based method for model-agnostic variable selection that identifies variables with nonlinear dependence on a response. It uses a functional-mapping perspective, pretraining on generated data to learn posterior inclusion probabilities, then applies to new datasets without updates. Experiments show strong performance, including a scaling law and a 20% improvement over baselines on a Facebook comments dataset.

Why it matters: Presents a novel, concrete AI method with empirical gains and scaling behavior, useful for researchers and practitioners in interpretable ML.

arxiv1d agoScore 82

variable selectionTransformer architecturesamortized learning

Research Editor selected

GroupAssistBench: Benchmarking Real-World Group Memory Assistance Beyond Recall

GroupAssistBench is a new benchmark for evaluating memory use in group assistance, built from real-world group conversations with a deployed assistant, covering 80,299 messages across 50 groups. It tests both recall (L1) and use (L2) with service requests lacking fixed answers, across single, distributed, and conflicting memories. Evaluations show frontier models and memory systems lag on use, especially for distributed and conflicting memories, but adding retrieval-relation and constraint-binding techniques improves performance.

Why it matters: This introduces a novel benchmark addressing a gap in AI memory evaluation, with real-world data and actionable insights for improving group assistant systems.

arxiv1d agoScore 85

AI memorygroup chat assistantsbenchmarking

Research Editor selected

ScamWorld: Interactive Environment for Evaluating Web Agents Against Online Fraud

ScamWorld is a new interactive environment that reconstructs reported scam websites to test how web agents handle online fraud. In tests with six frontier LLMs, baseline deception rates ranged from 11% to 35%, with agents more likely to proceed with payments than with credential or wallet requests. The environment enables controlled, repeatable evaluation of agent behavior against real-world fraud scenarios.

Why it matters: Provides a concrete, novel testbed and empirical results on LLM-based web agents' vulnerability to online fraud, directly relevant to AI safety and agent evaluation.

arxiv1d agoScore 85

AI safetyweb agent evaluationonline fraud

Research Editor selected

Modelling lock-in of AI adoption through skill erosion

This paper presents a mathematical model showing that generative AI can improve immediate performance but erode human skills, leading to bistability and hysteresis between low-adoption/high-skill and high-adoption/low-skill states. The model derives exact skill distributions for users and non-users, and shows that AI adoption can become locked in due to history-dependent switching costs. It also finds asymmetric transitions: a single AI user can spread among experts if AI pay exceeds expert pay, while a single non-user can only reverse adoption if they regain sufficient skill before returning to AI.

Why it matters: Provides a rigorous theoretical framework for understanding AI adoption lock-in and skill erosion, with practical implications for policy and workforce planning.

arxiv1d agoScore 82

AI adoption dynamicsskill erosionhysteresis

Research Editor selected

Fault-tolerant shallow recurrent decoders for sparse sensing with missing sensor measurements

This paper introduces fault-tolerant sparse sensing strategies for shallow recurrent decoders (SHRED) to handle missing, asynchronous, and heterogeneous sensor data. The authors propose transformer-based masking mechanisms and a multi-step decoding training objective, demonstrating improved reconstruction fidelity and robustness across synthetic and real-world hydrodynamic/ocean modeling benchmarks without added inference overhead.

Why it matters: Provides concrete, recent advances in robust AI-based state estimation for real-world sensing, with clear practical value for operational environments.

arxiv1d agoScore 82

SHREDsensor fault tolerancetime-series reconstruction

Research Editor selected

VisSearch-Bench: New Benchmark Evaluates Long-Horizon Visual Search in Multimodal Agents

VisSearch-Bench introduces 204 human-written tasks across 7 domains to evaluate multimodal agents on visual context management during long-horizon search. It covers five task types requiring agents to combine visual details with external information. Benchmarking 16 models shows the best system solves only 26.47% of tasks, highlighting significant room for improvement.

Why it matters: Provides a concrete, recent benchmark with clear results that quantify current limitations in multimodal agent visual search, valuable for researchers and developers.

arxiv1d agoScore 85

multimodal agentsvisual searchbenchmarking

Research Editor selected

Artifact-to-Capability: Diagnosing Breaks in Capability Formation of Self-Evolving Agents

This paper introduces a formal definition of capability formation from artifacts in self-evolving agents, proposing five sequentially dependent properties and a three-scale diagnosis method. In freestyle Gomoku, Reflexion, Voyager, and DGM all show standard proxy signals but exhibit distinct capability-formation breaks, such as legalization gain, compositional override, and loss-only drift. The findings demonstrate that signal-based evaluation cannot reliably determine whether a capability has formed.

Why it matters: Provides a concrete diagnostic framework and counterexamples that challenge common evaluation practices in self-evolving AI agents, offering practical value for researchers and developers.

arxiv1d agoScore 85

self-evolving agentscapability evaluationagent diagnostics

Research Editor selected

Shuffled OPSD: When Semantic Consistency Hurts On-Policy Self-Distillation?

This paper introduces Shuffled OPSD, a variant of on-policy self-distillation that shuffles the words of the shared problem statement while keeping the teacher-only worked example intact. Across three math reasoning benchmarks and three Qwen3 model scales, it improves mean accuracy over both intact-target and Standard OPSD, with the largest gain of 3.52 points on Qwen3-1.7B. Controls show that shuffling the teacher-only example hurts accuracy, while word-level shuffling of the problem statement helps, challenging the assumption that semantic coherence is always beneficial.

Why it matters: Provides concrete, recent experimental evidence that disrupting input structure can improve self-distillation, offering a novel insight for AI training methodology.

arxiv1d agoScore 85

on-policy self-distillationknowledge distillationmathematical reasoning

Research Editor selected

ARPG: Dependency-Aware Agentic Framework for Automated Robot Policy Generation

The paper introduces ARPG, a framework that uses LLMs as designers to automatically generate executable robot reinforcement learning policies from natural-language task descriptions. It decomposes the process into sequential MDP formulation and modular simulator implementation, with localized verification and correction to ensure semantic alignment of state, action, reward, and termination. Experiments in Isaac Sim show improved modeling correctness, environment executability, and policy-generation success over baselines.

Why it matters: Provides a concrete, recent AI development with a novel framework and experimental validation, useful for researchers and practitioners in robot learning and LLM agents.

arxiv1d agoScore 82

automated policy generationLLM as designerdependency-aware framework