Realtime AI News
A comprehensive survey finds that while LLMs can generate fluent peer review critiques, their reliability, robustness, and security as decision-support systems remain poorly understood.
A new dataset of 32,534 double-marked real student GCSE responses shows top LLMs agree with examiner consensus nearly as closely as two examiners agree with each other.
New research analyzes token-level predictive entropy trajectories across LLM layers to detect jailbreak attacks encoded in the model's internal representations.
A new framework called G-SPIN uses graph structures to correct phonetically-similar residual errors in ASR output, going beyond naive token-level fixes.
A lightweight RAG approach uses phonetically-aware TF-IDF retrieval to correct ASR hallucinations of rare entities and domain-specific terms.
Dustin introduces draft-augmented sparse verification to overcome the KV cache loading bottleneck in long-context speculative decoding.
New research shows that the direction detecting a behavior in LLM activations differs significantly from the direction that causes it, challenging a key interpretability assumption.
A new arXiv study uses reasoning tokens in vision-language models as an analog to reaction time in human visual search, finding behavioral similarities across four classic paradigms.
A new comprehensive practitioner's reference titled 'The Hitchhiker's Guide to Agentic AI' published on arXiv, covering the full stack from transformer architecture to production deployment.
A new survey paper reframes industrial continual learning for LLMs as a closed-loop update-and-release problem in an ecosystem, shifting focus from static benchmarks to real industrial needs.
A new study questions the reliability of learned fault-tolerant control methods for spacecraft, proposing a stricter benchmark that requires sustained pointing accuracy on unseen faults.
A new arXiv paper finds that multi-agent debate systems in subjective tasks like persuasion suffer from severe problem drift and sycophantic conformity, identifying semantic leakage in standard RAG as a reproducible trigger.
A new arXiv paper presents a method for computing Asymmetric Shapley Values using causal graphs, achieving polynomial-time computation in scenarios where SHAP is #P-hard.
A new benchmark, InvestPhilBench, evaluates large language models' procedural reasoning across 8 cognitive tiers in expert investment decision frameworks.
A new arXiv paper proposes using large language models to automate benchmarking for relational reasoning, addressing the core problem of unknown instance difficulty in evaluating neural generalization.
Researchers propose BrainAgent, a multi-agent LLM framework for autonomous brain signal understanding that lowers the technical barrier to brain-computer interface applications.
A new arXiv paper introduces Heuresis, a framework that abstracts the research pipeline into composable primitives, enabling open-ended scientific exploration while optimizing for quality, diversity, and novelty.
New benchmark RWGBench evaluates LLM performance in generating related work sections from a citation-level scholarly positioning perspective, going beyond traditional summarization metrics.
A new arXiv paper proposes per-player style embeddings for human chess that measure stylistic similarity via inner products while being approximately disentangled from playing strength (Elo).
New framework ReviewGuard uses a two-stage architecture to align LLM-generated peer reviews with citation-based estimates of long-term scientific impact.
New paper proposes using confidence sequences for online statistical model checking of MDPs, addressing the unrealistic assumption of exact probability knowledge.
New research proposes WinDOM, combining self-family distillation with reinforcement learning to achieve breakthrough GUI grounding performance in ~2B parameter small models.
New research proposes extracting and reusing skills from multi-agent offline datasets to address catastrophic forgetting and plasticity loss in sequential task scenarios.
New paper warns that widespread AI agent deployment could be abused for user surveillance, and proposes methods to evade such agentic surveillance.
New research introduces Autodata, a method enabling AI agents to act as data scientists building high-quality training and evaluation data, with self-optimization through Agentic Self-Instruct.
Researchers build a physically grounded multi-agent discovery engine that autonomously designs hardware-compliant computing systems, addressing the hallucination problem in generalist AI agents.
A new arXiv paper reveals a sharp dichotomy in how multi-agent RAG models benefit from document assessment — per-document filtering vs. holistic scoring — and proposes model-adaptive strategies to reduce computational costs.
New research introduces position graphs, a formal graph-based reasoning framework using two strict partial orders to model horizontal and vertical alignment of discrete tokens.
New research presents a hybrid quantitative-qualitative method based on Answer Set Programming for computing constrained branching trajectory modes for moving objects in real-world settings.
A new arXiv paper proposes OPPO, a reinforcement learning framework that explicitly optimizes multimodal perception for emotion reasoning, addressing the underutilization of multimodal cues and hallucination in current Omni-MLLMs.
A new arXiv paper proposes using cross-domain transferability of reasoning skills to dynamically adjust multi-domain RLVR training curricula, addressing inefficiency in fixed sampling strategies.
Researchers introduce the cliff token concept, identifying the precise single token where an LLM's reasoning shifts from correctness toward failure in mathematical tasks.
A new paper systematically categorizes four failure modes of LLMs on research-level mathematics, where models are confidently but fluently wrong, based on the First Proof benchmark.
A new paper adopts a 'compression is intelligence' viewpoint, proposing to quantify AI agent system intelligence in bits.
New research presents a versatile framework for evaluating fuzzy quantification queries over standard and fuzzy ontologies as well as knowledge graphs, agnostic to quantifier type and evaluation method.
A new arXiv paper examines autonomous AI systems transitioning from advisory to prescribing roles, noting that US bill H.R. 238 and Utah's prescription-renewal pilot authorize AI to prescribe, while identifying critical gaps in current regulatory guidelines.
Researchers propose the TSJ (Theater-Stage-Judge) framework, a longitudinal evaluation approach that reveals cumulative risks from prolonged AI companion interactions with cognition-developing users including children and adolescents.
New research addresses the problem of LLM-driven GUI agents encountering user-sensitive information screens, proposing a guided exploration approach that enables user takeover when needed.
New paper proposes an 'Unfireable Safety Kernel' concept for execution-time AI alignment, addressing the fundamental vulnerability of safety controls inside agent runtimes.
New research reveals that low-bit post-training quantization causes reasoning models to generate longer chains of thought even when answers remain correct, adding hidden inference costs.