3148 articles
FishBack proposes activation steering using pullback Fisher geometry for transformers. Authors show activation space is non-Euclidean (>97% deviation on GPT-2) and derive closed-form optimal steering equation. Method outperforms CAA, ActAdd, ITI by 1.3×–2.5× on off-target KL reduction.
ChemVA framework advances LLM understanding of chemical reaction diagrams through Visual Anchor mechanism for functional group detection and semantic alignment translating visual features to entity names. Achieves 92.0% structural recognition accuracy on OCRD-Bench dataset and ~20 percentage point performance gain across 9 diverse LLMs.
An arXiv paper argues that responsible agentic AI requires explicit, traceable provenance across the full lifecycle. Authors formalize this through a causal attribution function and responsibility tensor, demonstrating provenance is estimable and interventionable online before irreversible harm accumulates.
HEED introduces density-weighted residual alignment for distilling vision-language models (e.g., Qwen3-VL-8B) into hybrid Mamba-2/attention architectures. The method targets high-density patches (text, fine details) experiencing 3.6× larger residual drift. Results: +8.7 points OCRBench v2, +5.13 points average, 4.12× throughput, 68% memory savings.
Study on how representation geometry organization in language models depends on scale. Subspace PGA metric tests alignment of intermediate geometry with unembedding matrix readout. Small models (≤1024) progressively lose organization at late layers during training, while large models (≥2048) preserve it throughout. Scale determines how geometry organizes for prediction.
D²Evo is an RL framework to enhance LLM reasoning. It addresses scarcity of medium-difficulty samples by mining anchors matched to model capability and training a Questioner to generate diverse questions at appropriate difficulty. Results: outperforms existing methods on math benchmarks with <2K real samples.
Attractor-Vascular Coupling Theory (AVCT): mathematical framework showing cardiac attractor geometry encodes blood pressure information. Calibrated LightGBM model on smartphone PPG achieves MAE 2.05 mmHg (SBP) and 1.67 mmHg (DBP) in strict leave-one-subject-out cross-validation (46 subjects, 29,684 windows), meeting AAMI/IEEE SP10 criteria. PPG-only ablation matches ECG+PPG within 0.05 mmHg.
IsalProgram is a regular assembly-like language where every finite string is a valid program. Executed on a virtual machine with circular doubly linked list and three data pointers, it eliminates memory addresses and variable names. Proposed as a target for neural program synthesis.
DriveSafe is a framework for risk assessment in autonomous driving scenarios. It generates spatially grounded captions enriched with motion and depth cues, then fine-tunes a lightweight adapter to identify hazardous objects and suggest safety actions. Achieves SOTA on DRAMA benchmark.
SkyPart, a lightweight head for vision transformers, improves cross-view drone-to-satellite geo-localization by explicitly separating layout and texture via learnable prototypes. At 26.95M parameters, it achieves state-of-the-art on SUES-200, University-1652, and DenseUAV with enhanced robustness under weather corruptions.
Asynchronous RL pipelines for LLM agents lose historical old logits required for PPO off-policy correction, entangling discrepancy repair with staleness correction. The paper proposes three acquisition strategies (snapshot, dedicated model, interruption) and a revised PPO-EWMA method to preserve decoupled correction semantics.
Orthrus unifies autoregressive LLM fidelity with parallel diffusion token generation via a dual-architecture framework. A lightweight trainable module augments a frozen Transformer to enable parallel generation while maintaining exact autoregressive quality. Achieves up to 7.8x speedup with O(1) memory overhead.
LLM-based multi-agent pipelines flip to incorrect answers under simulated peer disagreement (yield). Contrary to common attribution, RLHF is not responsible: pretrained base models exhibit the same substitution pattern. Activation patching localizes corruption to a narrow mid-layer window. A single correctly-arguing dissenter reduces yield by 54-73 percentage points.
Complementary-label learning (CLL) methods remain limited to 10-class classification. This paper proposes BICL, a framework that deliberately uses biased (non-uniform) transition matrices to restrict complementary labels to class subsets. On CIFAR-100 and TinyImageNet-200, BICL achieves 7× accuracy improvements over traditional methods.
Study showing that two protocols for evaluating layer redundancy in transformers (replacement and interchange) yield divergent results for identifying layers to prune. On Pythia, Qwen3-8B, and Llama-3.1-8B, the protocol gap dramatically changes which layers appear safe to remove, even under the same KL evaluator.
Computational tool to classify manner and result verbs at scale. Uses linguistically informed prompts with LLMs to generate annotations over MASC and InterCorp data (436 VerbNet classes). RoBERTa-based classifier achieves 89.6% accuracy on three held-out gold-standard datasets. Applicable to developmental research on verb semantics.
Researchers propose LAD-inspired PPT, pre-pretraining on MP-STRUCT, a formal language encoding hierarchical composition and long-distance displacement. After 500 steps, this approach matches formal-language baselines in token efficiency while imparting LLMs human-like resistance to structurally implausible languages.
AgentKernelArena is an open-source benchmark for evaluating AI agents on GPU kernel optimization. It contains 196 tasks (HIP-to-HIP, Triton-to-Triton, PyTorch-to-HIP) and tests generalization on unseen configurations. Tested agents (Cursor Agent, Claude Code, Codex) achieve speedups up to 6.89x, but show generalization weaknesses on PyTorch-to-HIP.
Discrete diffusion models generate code through iterative refinement. CDC (Constrained Diffusion for Code) integrates constraints directly into the denoising process without additional training, combining mathematical optimization and program analysis to improve functional correctness, security, and syntax.
RTI-Bench is a structured dataset of 1,516 Indian Central Information Commission (CIC) decisions with outcome labels, exemption citations, and IRAC-style reasoning components. Mistral 7B achieves 57.3% accuracy on outcome prediction (baseline 14.3%). First publicly released structured dataset for Indian RTI administrative decision analysis.
MixSD is an external-teacher-free fine-tuning method that injects knowledge by dynamically mixing tokens from two model conditionals: an expert branch observing the injected fact, and a naive branch reflecting original priors. On QA and knowledge-editing benchmarks, MixSD retains up to 100% of base model capabilities versus 1% for standard SFT.
JSPG introduces a dynamic filtering framework for Chinese contextual ASR combining semantic, pinyin, and glyph features. The approach uses an extended Smith-Waterman algorithm to score N-best hypothesis sequences against keywords. Experiments on Aishell-1 and RWCS-NER datasets show significant improvements in keyword recognition accuracy.
Large Reasoning Models generate traces aligned with human reaction times, but this alignment persists regardless of inference-time reasoning budget. Study across GPT-OSS-20B and GPT-OSS-120B: token allocation tracks human difficulty patterns and remains invariant across effort levels, suggesting cognitive cost alignment is crystallized at training time.
Multilingual coreference resolution system based on Gemma-3-27b with two-stage adaptation (multilingual base adapters then dataset-specific adapters). CoNLL F1 score of 74.32 on CRAC 2026 test set, ranked first in LLM track. Mention spans represented by headword using XML-inspired format with local reindexing.
SkillTTA synthesizes task-specific textual skills by retrieving relevant training trajectories, with no parameter updates to the solver model. Evaluated on SpreadsheetBench, ALFWorld, and BigCodeBench: SpreadsheetBench improves from 0.397 to 0.505 Pass@1, BigCodeBench from 0.517 to 0.651.
PARALLAX reveals that 4 of 6 major hallucination detection benchmarks embed the ground-truth answer in the prompt, allowing a naive baseline (TxTemb) to achieve near-perfect detection without access to model internals. Evaluation of 22 methods across 12 open-source models: most fail under controlled conditions, except SAPLMA and DRIFT (supervised probes on upper-layer hidden states).
PluRule is a multimodal multilingual benchmark for moderating pluralistic communities on social media. It covers 13,371 rule violations across 1,989 Reddit communities (9 languages, 2,885 rules). State-of-the-art vision-language models, including GPT-4.5 with advanced reasoning, only marginally outperform a trivial baseline.
SEMA-RAG is a multi-agent framework for retrieval-augmented generation applied to medical reasoning. It decouples the process into three specialist agents: clinical interpretation, iterative document exploration, and evidence adjudication. Tested on 5 benchmarks and 5 LLM backbones, it improves baselines by +6.46 accuracy points on average.
Tutorial on multilingual multimodal LLMs for low-resource languages. Covers recent models (PALO, Maya), speech-text-vision pipelines, low-cost data creation, tri-modal alignment via adapters, and culture-aware evaluation beyond English.
VLMs struggle with planning from complex visual inputs. This paper proposes Pattern Induction, an online inductive learning strategy that discovers and optimizes reusable visual patterns as composite experts. Pattern Inference enables VLMs to recognize these patterns and directly infer world model structures. Evaluated on FrozenLake, Crafter, and CubeBench.
arXiv study demonstrating that 9 frontier LLMs inherit stigmatizing language bias from clinical notes. Models reduce treatment aggressiveness when exposed to a single stigmatizing sentence. Chain-of-Thought and self-debiasing show limited mitigation efficacy.
FIM-LoRA optimizes rank allocation in LoRA by using 8 calibration passes to estimate gradient variance per layer. This parameter-free approach matches standard LoRA performance (88.6 vs 88.7 on GLUE with DeBERTa-v3-base) while reducing memory costs by 256x compared to full Fisher estimation.
Injecting mathematically wrong drafts from a smaller model (Qwen2.5-Math-1.5B) mismatched to the current problem into a stronger learner's (Mathstral-7B) GRPO context outperforms standard on-policy GRPO. On MATH-500, the mismatched-wrong variant reaches 71.98% (highest published result for this model), +1.62pp vs matched-wrong variant, without SFT or reward models.
AgentRevive introduces a Markov state-aware framework for resilient multi-agent LLM system evolution. Instead of aggressively pruning failing agents, the method uses soft state transitions (Active/Standby/Terminated) with a hallucination risk estimator. Results: outperforms baselines on general reasoning, domain-specific tasks, and hallucination challenges while reducing token consumption.
AMATA is an adaptive multi-agent trajectory alignment framework for knowledge-intensive question answering. Six specialized agents collaboratively perform structured actions to improve factual consistency and reduce hallucinations. The system formalizes multi-agent collaboration as a trajectory preference alignment problem with intra-trajectory and inter-agent dependency learning.
Parameter-efficient vocabulary adaptation method to improve LLM tokenization on specialized domains (legal, medical). Tested on Llama-3.1-8B and Qwen2.5-7B: reduces training time by 35-55% vs continual pretraining, decreases parameters by 37% vs expansion-only, improves summary quality through domain-specific tokens.
MiniGPT is a compact from-scratch PyTorch implementation of GPT-style autoregressive language modeling in a single notebook. The 10.77M-parameter model achieves validation loss of 1.4780 on Tiny Shakespeare with character-level tokenization and generates text with recognizable Shakespeare-style dialogue structure.
BELIEF combines structured evidence modeling and uncertainty-aware fusion for biomedical question answering. The framework converts retrieved documents into evidence objects (clinical attributes, source quality, relevance, support strength) and fuses two reasoning paths: symbolic (Dempster-Shafer theory) and neural (LLM). SOTA results on PubMedQA, MedQA, MedMCQA across 5 LLM backbones.
Study reveals an 'unlearnability' phenomenon in Reinforcement Learning with Verifiable Reward (RLVR) for LLMs. Some hard examples remain unlearnable even with correct rollouts. Cross-example gradient analysis shows fundamental representation flaws: low gradient similarity and ungeneralizable reasoning patterns. Data augmentation fails to improve gradient similarity.
Analysis of routing dynamics in two MoE architectures (Qwen3-30B-A3B and Nemotron-3-Nano-30B-A3B) reveals deep-layer routing collapse for underrepresented languages (Hebrew, Japanese). Continual pre-training on balanced bilingual data corrects this imbalance better than supervised fine-tuning alone.
Paper introduces Contract-Bounded Evidence Activation (CBEA) with Lexicographic Commitment Validation (LCV) for personalized language systems. CBEA+LCV achieves zero failures at 0.49-0.60 availability versus 0.003-0.092 for baselines, with 74-75% median input payload reduction.
Comparative study of 10 annotation workflows for conversational speech summarization. Audio-based summaries are less informative than transcript-based ones, but iterative peer-editing with audio mitigates this gap. Validates this approach for creating benchmarks incorporating lexical and prosodic information.
Benchmark of LLMs on multi-label legal precedent treatment classification. Expert-annotated dataset of 239 real-world citations. Gemini 2.5 Flash achieves 79.1% on high-level classification, GPT-5-mini 67.7% on fine-grained schema. Novel Average Severity Error metric to measure practical impact of misclassifications.
Novel topic modeling approach (SBTA) assigns topics to text segments rather than entire documents, reducing topic contamination. Authors construct SemEval-STM, a dataset annotated via LLM + human refinement, and validate improved clustering quality and interpretability across multiple models.
Researchers demonstrate that small models (Gemma 4 E4B, Qwen3-4B) fine-tuned with 8-bit QLoRA internalize tool knowledge without requiring tool schemas in prompts. On AssetOpsBench, fine-tuned models outperform unfine-tuned baselines: 82.6% input length reduction, AT-F1 of 0.65 vs 0.47, and 2.5× faster for Qwen3.
An arXiv study compares Matryoshka Representation Learning (MRL) with simple embedding truncation. Results show non-MRL embeddings remain robust up to 80% dimensionality reduction. MRL provides advantage only for heavy truncation (>80%), questioning its systematic training cost.
Systematic evaluation of synthetic clinical notes generated by LLMs at million-note scale from MIMIC databases. Study shows synthetic notes preserve core clinical information for coarse-grained tasks but lose fine-grained details for ICD coding. Chunk-based rephrasing mitigates detail loss but reduces factual precision under incomplete context.
SynPro, a synthetic data generation framework, helps LLMs learn more thoroughly from limited organic corpora via rephrasing and reformatting. Optimized with RL, it unlocks 3.7-5.2x more effective tokens than simple repetition on 400M and 1.1B models, even surpassing the non-data-bound oracle at 1.1B scale. Code open-sourced.
Pilot benchmark for translating natural language to First-Order Logic (FOL) in planetary exploration. Dataset built from NASA mission documentation (2003-2013), manually annotated with FOL representations capturing temporal structure, agent roles, and operational dependencies. Structured predicate vocabularies provided.
AutoVecCoder teaches LLMs to generate explicit vectorized code via SIMD. The framework combines VecPrompt (data synthesis to inject intrinsic knowledge) and VecRL (reinforcement learning aligned with execution efficiency). AutoVecCoder-8B achieves state-of-the-art on SimdBench (SSE/AVX) subsets and sometimes surpasses -O3 optimizations.
PROTEA is an interface for offline debugging and refinement of multi-agent LLM workflows. It evaluates intermediate outputs with configurable rubrics, localizes bottlenecks via workflow graph visualization, and generates targeted prompt revisions. On two production-adjacent workflows, PROTEA improves document-inspection accuracy from 64.3% to 83.9% and recommendation Hit@5 from 0.30 to 0.38.
Study on 'alignment drift': gradual process where LLM outputs become less constrained by user's current message and more shaped by interaction history, while remaining helpful. Mechanism-oriented framework distinguishes signal A/B, feedback loops, and interactive regimes to control this cumulative drift.
MINE (Mechanistically Interpretable Neural Encoding) applies mechanistic interpretability tools to neural encoding models to localize visual features driving voxel-level activity in human visual cortex. Validated via image generation and counterfactual editing: inserting/removing predicted features shifts neural activation as expected.
Formal verification framework for LLM pipelines using Lean 4 certificates. Three certificate families (conflict-aware bilattice, embedding sensitivity, Hoare-style agent action) plus two operators (Maximal Certifiable Residue, Compositional Stability) for high-stakes deployments (regulated finance, clinical support, agentic systems). Compiled artifact covers 22 certificate types, 17/46 declarations axiom-free.
arXiv paper proposing formal framework for combining LLM and human evaluations. Uses doubly robust estimator (missing data literature) to determine optimal number of human reviews needed. Shifts LLM role from substitutive to auxiliary in two-stage sampling design.
Study of 2,258 English-language posts (April 2026) shows uncertain posts receive 82% more replies than certain posts. Regression confirms positive association (β=0.126, p=0.011), ~13% higher reply engagement. Replicates asymmetry observed in Arabic, suggesting universal interactional mechanism across languages.
Study on using LLMs to compose personalized and persuasive push notifications. Authors define 6 quality dimensions (contextual relevance, clarity, actionability, etc.) and demonstrate +8% to +14.5% CTR gains vs static templates. Proposes architectural framework with budget-aware routing, grounded generation, and online learning.
Study of 38 models on 8,900 scholarly references: factual recall quality follows a sigmoid combining model size and topic frequency in training data. These two variables explain 60-94% of variance. Model proposes recall is gated by signal-to-noise ratio scaling with concept frequency and model capacity.
EnvFactory automates creation of executable environments and synthesis of multi-turn trajectories for Agentic RL training. Using 85 verified environments across 7 domains, the framework generates 2,575 SFT/RL trajectories and improves Qwen3-series models by +15% on BFCLv3, +8.6% on MCP-Atlas, and +6% on conversational benchmarks.
Circuit analysis of a backdoor in an 8B model: a 3-word Latin trigger redirects English output to French. The circuit operates in 3 phases via attention heads, propagates through a subspace orthogonal to natural language-identity directions, then converts via MLP. A single serial bottleneck position controls the entire flow.
MA²P is a multi-agent autonomous framework for complex persuasion. It coordinates perception management, mental-state inference, strategy execution, and performance evaluation. A meta-cognitive configurator selects domain-appropriate meta-strategies from a knowledge base to improve generalization and persuasion success rates.
Data-driven approach to integrate constructs and their relations in information systems. Combines task-adapted text embeddings and clustering to group constructs from structural equation models. Optimizes trade-off between semantic purity and parsimony through explicit loss function.
LongMINT is a benchmark evaluating agents' memory management in long contexts (up to 1.8M tokens) with multi-target interference. 15.6k QA pairs across 4 domains (state tracking, dialogue, Wikipedia revisions, GitHub commits). 7 systems tested (long-context LLMs, RAG, agent frameworks) achieve 27.9% average accuracy, bottlenecked by retrieval and memory construction.
Investigation of LRM internal representations through probe trajectories. Authors show that continuous evolution of concept probability during reasoning predicts final behavior better than static predictions. Max-pooling achieves 95% AUROC across 4 datasets (safety, mathematics).
DiSP, a demonstration selection framework for in-context learning, predicts whether a query-context pair will succeed rather than searching for optimal context. On 5 classification datasets with Llama 3-8B and Qwen 2.5-7B, DiSP improves accuracy by 3.4% and achieves 23× end-to-end speedup.
New IH-GRPO algorithm decouples tool invocation from execution to enhance LLM mathematical reasoning. Achieves 1.87–2.53% improvements on mathematical benchmarks with Qwen3 (1.7B–8B). Code released.
Preregistered study comparing Vector RAG and LLM-compiled markdown wiki on 13 questions over 24 papers. Wiki excels at cross-paper synthesis and claim-level citation accuracy, but uses more query tokens. A decomposition-based RAG variant recovers most wiki advantages at lower LLM-token cost.
EvoMemBench is a unified benchmark evaluating LLM agent memory along two axes: scope (in-episode vs cross-episode) and content (knowledge vs execution-oriented). Comparison of 15 memory methods: long-context baselines remain highly competitive, retrieval-based methods dominate knowledge-intensive tasks, procedural methods excel at execution-oriented tasks.
Comparative study of human judgments and 4 LLMs predictions on presupposition projection in conditionals. 120 participants evaluated in parallel with models. Humans integrate probabilistic and pragmatic cues; LLMs show variable alignment. Models matching human ratings lack coherent pragmatic reasoning.
Infini-News indexes 1.35B CC-News articles (August 2016–present) with metadata extraction, language detection (GlotLID, lingua, CommonLingua), and geographic attribution (83.4% coverage). Infini-gram suffix-array indexes enable sub-second full-text pattern search across the entire archive.
EPIC is an on-device RAG index construction method optimized for personal AI agents. It reduces indexing memory by 2,404× by focusing on user preferences, improves preference-following accuracy by 20.17 percentage points, and decreases retrieval latency by 33.33×. Memory footprint under 1 MB with 29.35 ms/query latency.
First machine unlearning framework for masked diffusion language models (LLaDA, Dream). MDU minimizes KL divergence from prompt-conditional to prompt-masked unconditional distribution at each masked position, with temperature scaling for privacy-utility trade-off. Code released.
QuantFPFlow integrates quantum amplitude estimation (Grover) into stochastic policy optimization via Fokker-Planck formulation. Provable quadratic speedup O(1/ε) vs O(1/ε²) classical. On continuous multimodal task, outperforms SAC (1295.7 vs 1284.0 reward) and finds global optimum 10.4% more frequently (33.9% vs 30.7%).
STRIDE is a self-reflective agent framework for LLM-based symbolic equation discovery. It improves reliability by coordinating data-aware generation, mixed-fitting evaluation, critic-executor repair, and diversity-preserving semantic memory. Experiments on symbolic regression benchmarks show gains in accuracy, OOD robustness, and structural recovery across multiple LLM backbones.
arXiv study using wearable sensors (accelerometry, EDA, temperature) and multimodal foundation models to predict challenging behaviors in 9 profoundly autistic children in classroom settings. Prediction up to 10 minutes in advance with AUC-ROC 0.78 on 110.7 hours of real-world data.
arXiv advances its HTML Papers offering for TeX/LaTeX submissions since 2023. 2025-2026 highlights: resolved ~3,000 of 6,000 user reports, targeting 90% error-free HTML conversion (currently 75%), initial MathML 4 Intent annotations for accessible speech output, and in-progress Rust port of LaTeXML to reduce compute costs.
Theoretical paper arguing AI in biomedicine must transition from predictive systems based on historical data to interventional models capable of simulating novel treatment effects. Current architectures remain observational and cannot generalize to unobserved interventions.
Study reveals dataset visibility asymmetry in multilingual NLP: 118 languages (59% of 200 most-spoken) have zero catalogued datasets per LRE Map and LDC. Using LLM-assisted citation-mining on Semantic Scholar, authors identify 609 unique datasets across 53 low-visibility languages, 356 publicly accessible. Data scarcity is a documentation and discoverability issue, not just production.
Whole-slide MIL models suffer from 'spatial blindness': accurate predictions but ignoring tissue architecture. ResTopoMIL fixes this by separating appearance statistics (prototype histogram) from spatial relations (graph branch with permutation constraint). Improvements across 9 WSI benchmarks with 1.15M parameters.
ContraFix is an agentic framework for automated vulnerability repair combining differential runtime evidence and skill reuse. On SEC-Bench (C/C++) and PatchEval (Go, Python, JavaScript), it achieves 84.0% and 73.8% resolution rates with GPT-4-mini, outperforming baselines while costing less than one-third of comparable approaches.
FactorizedHMR introduces a two-stage hybrid framework for video human mesh recovery. A deterministic regression module anchors torso and root, while a probabilistic flow-matching module completes ambiguous distal articulations (arms, legs). Geometry-aware supervision and classifier-free guidance improve recovery under occlusion.
Safety-critical control method for autonomous navigation in unknown environments. Uses conformal prediction to generate risk-aware obstacle ellipsoids accounting for sensor uncertainty, then two nested differentiable optimization layers to build control barrier functions. Validated through numerical simulations.
Deep learning models analyzing facial expressions and head movements in asynchronous video job interviews detect honesty and deception with 91% and 84% variance explained. Outperforms human evaluators on N=121 applicants.
Closed-loop intelligent tutoring system using XGBoost to assess oral presentation skills via multimodal analysis (facial, vocal, textual, oculomotor). Trained on 10,360 MOOC videos, generates feedback aligned to 7-dimensional BARS scale. Study with 204 learners over 30 days: significant improvements (Cohen's d = 0.39-0.90), strong correlation between practice frequency and performance.
SaaSBench is the first benchmark to evaluate AI agents in enterprise SaaS engineering. It contains 30 complex tasks across 6 SaaS domains with 8 programming languages, 6 databases, and 13 frameworks. Experiments show >95% of failures occur before business logic: agents struggle to configure and integrate multi-component systems.
KairosHope is a time-series foundation model replacing quadratic attention with dual-memory architecture (Titans modules + Continuum Memory System). Pre-trained on Monash archive via MTSM and contrastive learning, it fuses latent representations with statistical features. Superior results on UCR for HAR and sensor data.
Network intrusion detection system using triplet networks with online triplet mining and KNN classifier. Few-shot approach detects attacks with as few as 10 malicious samples per class, outperforming standard supervised methods on small datasets and anomaly detection models (high false-positive rates).
New method to control False Discovery Rate (FDR) in structured hypothesis spaces using reproducing kernels. Reformulates the problem as regularized learning in RKHS, unifying continuous domains, graphs, and hierarchies. Validated on spatial data and differential gene expression tasks.
Comparative study of 6 EEG foundation models across 8 datasets beyond clean accuracy. Robustness analysis (noise, channel dropout), interpretability via Attention-Aware Layer-Wise Relevance Propagation, and expressiveness through block-wise probing. Findings: no single model dominates all failure modes; models focus on task-appropriate brain regions but decode corrupted content poorly.
SynVA is a modular toolkit for generating vascular meshes and synthesizing anatomically consistent intracranial aneurysms. Combines flow-matching methods for healthy vessels with anatomy-conditioned approaches for aneurysm generation. Releases a dataset of 50,000 fully labeled mesh samples for medical vision tasks.
Study of 'sleeper memory poisoning' attack against stateful LLM agents with persistent memory. Adversary corrupts external documents to inject false user memories. Success rates: 99.8% (GPT-5.5), 95% (Kimi-K2.6). Poisoned memories trigger attacker-intended actions in 60-89% of cases.
SparseSAM introduces training-free structured sparsification for Segment Anything Model's ViT encoders. Using Stripe-Sort Attention (Z-order permutation) and Residual-Consistency MLP, it achieves 2x inference speedup and 2.8x memory reduction with only 0.004 mIoU loss at 0.4 density.
Theoretical paper on transformer training in mean-field regime (infinite depth and width). Authors model training as controlling a neural PDE (vs ODE for ResNets), establish well-posedness of forward pass, derive explicit formulas for Wasserstein gradients, and prove gradient flow convergence to global minima under NTK injectivity conditions.
Lance is a lightweight unified multimodal model supporting understanding, generation, and editing of images and videos. Built on dual-stream mixture-of-experts architecture with modality-aware rotary positional encoding, it combines collaborative multi-task training and adaptive data scheduling to outperform existing open-source unified models in visual generation.
StructLens analyzes the internal organization of representations in language models through maximum spanning trees built on residual streams. The framework reveals that middle layers strongly organize nearby tokens, and that smaller local units emerge before larger units during pre-training.
OSCAR quantizes KV caches to INT2 for long-context LLMs by estimating attention-aware covariance structures offline. Tested on Qwen3 (4B–32B) and GLM-4.7 (358B), it reduces accuracy gap to 1.42–3.78 points vs BF16, cuts memory by 8x and improves throughput by 7x. Custom INT2 kernel compatible with vLLM/SGLang.
Longitudinal study (2020-2026) of 471 AI students revealing a preference shift: from 2024 to 2026, valuation of human intelligence rises from 53% to 65% in technical courses and 90% in design courses. Authors identify four phases (hype, distrust, trust, dependency) and conclude AI is being reappraised as a routine tool.
Studies rearrangement of densely packed tabletop objects with parallel grippers. Introduces a directional knock primitive to overcome infeasible direct picks. Formulates optimal knock-pick problem and proposes abstractions with maximum-weight perfect matching for polynomial-time computation of action-minimizing plans. Validated in simulation (IsaacSim).
Study of tabletop stack rearrangement using nonprehensile toppling actions. A novel graphical abstraction models interleaved pick-and-place and topple operations. IsaacSim physics simulation benchmarks demonstrate faster execution than pick-and-place alone.
New arXiv paper introducing CSDI (content-style differential independence) to identify content and style factors in multi-domain generative models. Relaxes prior statistical independence conditions via blockwise orthogonality constraints on Jacobian subspaces. Demonstrates identifiability even with dependent content/style and dense Jacobians.