3148 articles
Hilbert-Geo introduces a unified formal framework for solid geometry via Parse2Reason: parsing into Conditional Description Language (CDL) then reasoning with theorem bank. Achieves 77.3% on SolidFGeo2k and 84.1% on MathVerse-Solid, outperforming Gemini-2.5-pro (54.2%) and GPT-5 (62.9%). Two expert-annotated datasets: SolidFGeo2k and PlaneFGeo3k.
ReTAMamba is a Mamba-based model for predicting irregular clinical time series. It estimates observation reliability from missingness and elapsed time, integrates multi-resolution information via Chronological Weaving, and uses a budgeted token router. On MIMIC-IV, eICU, and PhysioNet 2012, it improves AUPRC by 7.51%, 7.80%, and 10.15% respectively.
GPU-accelerated deep learning framework for next-day urban thermal prediction and heatwave risk assessment. ConvLSTM with mixed loss function achieves MAE=0.2293, RMSE=0.3089, R²=0.8877 using MODIS and Open-Meteo data in Sarajevo. Generates city heat risk maps.
ORACLE is an agentic framework for early scam anticipation from streaming app-usage trajectories. On a benchmark of 12 scam types (95 apps, 15 days average), the system uses a self-evolving context manager and on-policy self-distillation to detect early-stage signals from partial observations.
Critical audit of SAEBench, the de-facto standard evaluation suite for sparse autoencoders (SAEs). TPP and SCR metrics fail multiple reliability tests and should not be used. Other metrics show higher reseed noise and lower discriminability than assumed. Only sae-probes demonstrates acceptable reliability, but struggles to distinguish architecture variants.
EPIC (Efficient Preference-aligned Index Construction) optimizes on-device RAG by storing user preferences instead of raw data. Across 4 benchmarks, reduces indexing memory by 2404×, improves preference-following accuracy by 20.17 percentage points, and lowers retrieval latency by 33.33×. Memory footprint under 1 MB with 29.35 ms/query.
Comparative study of 29 data augmentation techniques for 2D echocardiography segmentation on U-Net. Anatomically plausible geometric transformations (affine, shift-scale-rotate, perspective, horizontal flip) improve cross-dataset generalization, while aggressive intensity augmentations degrade it. Pairwise combinations outperform individual augmentations.
Charon is a unified, modular simulator for predicting large-scale LLM training and inference performance. It achieves prediction error under 5.35% overall and 3.74% for training on GPU clusters. In inference deployment, it identified a configuration improving system throughput over engineering-tuned baselines.
MR-SLAM is a mixed reality system using Meta Quest 3 to teleoperate three TurtleBot3 robots in collaborative SLAM. The operator views the real world in passthrough with spatially anchored information panels. Three SLAM Toolbox instances merge occupancy grids in real time via ROS 2, achieving 94.7% cross-instance consistency and 8.83 Hz scan rate.
KVCapsule compresses KV cache in vision-language models during autoregressive decoding. The method exploits structured attention patterns in vision tokens to achieve 2x TPS improvement and 2.4x memory reduction at 60% compression ratio, without modifying the pretrained backbone.
Hierarchical two-stage framework for long-horizon vessel trajectory prediction under real ocean conditions. Combines long-term predictor with short-term Spatio-Temporal Graph Transformer on discretized maritime cells. Environmental module integrates currents, wind, wave height via cross-modal attention. Results: 25% improvement in ADE, 17% in FDE on Australian CTS data.
DAEM (Diffusion Attention Expert Model) detects STAS (spread through air spaces) in lung cancer histopathological images. Model achieves AUC 0.8946 on frozen sections and 0.9112 on paraffin sections. Validated across 8 external institutions. Semi-automatic localization and TME biomarkers identified.
PH-Dreamer embeds Port-Hamiltonian physical principles into world models to improve latent imagination. The framework models energy evolution, estimates the Hamiltonian from proprioceptive observations, and uses an energy-guided Actor-Critic. Results: 4.18-8.41% phase space volume reduction, up to 7.80% energy consumption decrease, up to 9.38% jerk reduction.
Game-theoretic decoding method for small vision-language models (2-8B) in medical imaging. Wasserstein semantic stopping criterion replaces lexical matching, improving Qwen3-VL-2B by +3.5pp on VQA-RAD and reducing convergence iterations by 20% while maintaining reliability.
ISEP proposes an offline reinforcement learning method that implicitly expands action support by interpolating between in-distribution data and policy samples. A stochastic mechanism alternates between conservative cloning and optimistic expansion signals, implemented via Conditional Flow Matching with classifier-free guidance.
TailedTS is a benchmark of 24.69 billion data points from Wikipedia 2024 page views, designed to test time series forecasting models under heavy-tailed and non-Gaussian conditions. The dataset shows 5% of pages account for 70% of traffic, and introduces a periodicity quantification framework revealing high-traffic pages have weaker periodic structure.
ProxyKV introduces a cross-model proxy pruning framework to accelerate long-context LLM inference. A lightweight in-family small model evaluates KV cache importance asynchronously via HybridAxialMapper and Multi-Granularity Hybrid Loss. On Llama-3.1, Qwen-2.5, and Qwen-3, recovers 98.7% of KVZip accuracy with up to 3.21× prefilling speedup (Llama-3.1-8B, dual-GPU) and sustains speedup at contexts up to 170k tokens.
Study of membership inference attacks (MIA) on masked diffusion language models (MDLM). Researchers extract 46-dimensional feature vectors from reconstruction loss at different masking ratios and train XGBoost and MLP classifiers. On MIMIR benchmark, XGBoost achieves AUC 0.878 (peak 0.930), outperforming SAMA baseline by 0.062 AUC. ELBO trajectory alone drives most of the signal.
arXiv paper proposing a formal statistical framework to combine LLM and human evaluations. Uses a doubly robust estimator (missing data approach) to determine optimal sample sizes of human ratings needed for benchmark validation, based on LLM judgment predictability.
PESD-TSF is a physics-inspired structured decomposition framework for long-term time series forecasting. It introduces a Multiplicative Periodic Gating mechanism, a multi-scale encoder with detrended attention, and Cross-Scale Collaborative Attention (CSCA) to preserve periodic structures and inter-variable dependencies across deep layers.
Peak-Detector leverages instruction-tuned LLMs for peak detection across physiological signals (ECG, PPG, BCG, BSG) with explainability. A "peak-representation" technique compresses time-series while preserving critical events. The model is optimized via supervised fine-tuning then multi-objective reinforcement learning, evaluated on 7 datasets (6 public benchmarks + 1 real-world cohort).
Transformer-based world model for video frame prediction. Formulates next-frame prediction as structured probabilistic inference with latent token correspondence variables. Each token either copies from previous frame or generates new. SOTA on 4 benchmarks: 72.5% return and 35.6% score on Craftax-classic (vs 67.4%/27.9% prior).
Video inference enhancement method using a fuzzy controller (FC-r) to dynamically adapt model sizes based on available device resources. Leverages spatiotemporal correlation across adjacent frames. Balances inference performance and resource efficiency without scaling up model complexity.
MINE (Mechanistically Interpretable Neural Encoding) applies mechanistic interpretability to neural encoding models to identify visual features driving activation in individual voxels of human visual cortex. Using language-aligned image representations and counterfactual editing, the approach causally validates fine-grained selectivity in category-selective brain regions.
LERA introduces a two-stage ad auction framework for LLM chatbots: coarse filtering via embeddings, then fine ranking using LLM prompts for relevance scores. Combines organic relevance with bids using critical-value payment rule. Improves ad selection accuracy and insertion diversity on synthetic benchmark.
CSRNet optimized with parameter-free attention mechanisms for crowd counting in public transport. Evaluation of PFCA, SA, and SimAM modules on ShanghaiTech dataset. Novel PFCASA combination (PFCA+SA) outperforms parameterized approaches while reducing model size, applicable to edge-deployed systems.
WhiteTesseract combines high-resolution XR and conversational AI to enhance museum visits. Tested on a Monet exhibition with 26 participants, the system increases viewing duration from 35.3 to 98.3 seconds (p<0.001). 60% of 529 interactions extend beyond factual queries to include analytical, emotional, and comparative inquiries.
MoleCode is an LLM-native molecular language representing molecules as explicit graphs with typed entities and direct relations, replacing implicit SMILES strings. Training-free, it improves frontier LLMs on molecular reasoning, editing and generation tasks, especially for unfamiliar molecules, topology-sensitive operations and larger structures.
Visual Agentic Memory (VAM) is a training-free framework for long video understanding. It combines online selective indexing, hierarchical memory, and agentic retrieval. On OVO-Bench, VAM achieves 68.41 (vs 67.46 for Gemini 3 Flash alone) and 17.11% on MM-Lifelong (105.6h over 51 days).
HGPM (Hypergraph Pattern Machine) models higher-order interactions by tokenizing compositional subsets and using an inclusion-aware Transformer. On 10 hypergraph benchmarks, the method matches or exceeds state-of-the-art, notably in adverse-event prediction where it correctly identifies inhibitory drug combinations that existing methods miss.
Paper proposing an experiment to establish case law on inventorship in AI-assisted inventions. Given the lack of legal precedent and rapid AI development, authors suggest a structured approach involving stakeholders to identify how to measure human contribution and determine inventor recognition.
Post hoc calibration method for Deep Cox models using isotonic regression. Improves calibration of predicted survival probabilities without affecting discriminative power. Theoretical guarantees including double-robustness and asymptotic calibration, validated on synthetic and real-world clinical data.
FedNL reformulates federated learning as a three-level nested optimization system. Embeds Titans-based linear attention to enable zero-shot test-time adaptation without additional training. Experiments on MMLU and long-context benchmarks show competitive performance and constant inference memory.
Study of geometric structure in Mixture-of-Experts (MoE) architectures using Jacobian-PCA-Grassmann framework. Analysis of Mistral and Qwen reveals asymmetry: strong functional decorrelation between experts but partially overlapping representations. Sparse routing (top-k) strengthens functional separation.
HPC-LLM is a specialized LLM assistant for HPC workflows (Slurm, MPI, GPU, filesystems). Built on Llama 3.1 8B fine-tuned with QLoRA on ~9k-24k HPC examples, it combines RAG and domain adaptation. The model achieves performance comparable to Qwen 2.5 14B with lower GPU resource requirements.
GRASP is an agentic system for multi-hop question answering that decomposes queries into dependency-aware plans and uses a three-layer hierarchical graph (entities, propositions, passages). On MuSiQue, 2WikiMultihopQA, and HotpotQA, GRASP achieves highest accuracy while consuming 40-50% fewer tokens than IRCoT+HippoRAG2.
PromptDecipher is an authoring system for AI tutoring chatbots that restructures workflow around direct corrections rather than abstract system prompts. Teachers interact with a live chat preview, edit undesirable bot responses, and an automated pipeline proposes targeted prompt rewrites validated across pre-defined test scenarios.
LoopQ is a post-training quantization (PTQ) framework for recursive language models (LoopLMs) that reuse Transformer blocks. It addresses three challenges: distribution shift across roles, state reuse between loops, and recursive error accumulation. Under W4A4 quantization, LoopQ improves downstream accuracy by 68.8% and reduces perplexity by 87.7% versus static PTQ baseline.
Graph Hierarchical Recurrence (GHR) is a novel framework for GNNs and Graph Transformers that captures long-range dependencies through hierarchical abstraction via pooling. GHR outperforms existing models on long-range benchmarks using only 1% of SOTA parameters, and improves out-of-distribution generalization.
LLMs used as autonomous agents select unauthorized tools despite explicit instructions. Study across Qwen 2.5 7B, Llama 3.1 8B, and Claude Haiku 3.5 shows an MCP proxy with attribute-based access control (ABAC) reduces unauthorized invocation rate to 0%, versus 11-18% for prompt-based restrictions. Architectural enforcement, not prompting, is required for reliable tool access control.
DACA-GRPO improves reinforcement learning for diffusion language models by addressing temporal credit assignment and mean-field likelihood bias. It introduces Denoising Progress Scores and Stratified Masking Likelihood, achieving gains up to 7.4pp on code generation and 36.3pp on constraint satisfaction across seven benchmarks.
CrossView Suite introduces CrossViewSet (1.6M multi-view samples), CrossViewBench (evaluation benchmark), and CrossViewer (three-stage framework: Perception → Alignment → Reasoning) to enhance cross-view spatial reasoning in MLLMs. A multi-agent data engine generates annotated data covering 17 fine-grained task types.
Controlled study comparing human vs synthetic soft-labels on MNIST. Human labels improve calibration by regularizing predictions on difficult samples and aligning model uncertainty with human uncertainty, beyond simple mislabeling correction.
PIMSM embeds physical constraints into a multi-scale Mamba architecture to improve representation stability under distribution shift. The model aligns discretization parameters to characteristic frequencies in temporal series (fMRI, weather). Results: improved robustness on Human Connectome Project and Weather-5K with minimal MAE in out-of-distribution forecasting.
PopPy is a system that automatically uncovers parallelization opportunities in Python compound AI applications (multiple ML model calls). On real-world applications, PopPy achieves up to 6.4× speedups in end-to-end execution time, combining an ahead-of-time compiler and runtime to handle language complexity, dynamic dispatch, and variable mutation.
Tabular foundation models (TFMs) for healthcare are distilled into lightweight models using context-leakage-aware distillation. Across 19 medical datasets and 6 TFMs, student models retain ≥90% of teacher AUC while running 26× faster on CPU, preserving calibration and fairness critical for health applications.
Theoretical paper developing a proof-theoretic semantic account of information grounded in inferentialist reasoning. Replaces truth with inferability in Dretske's framework, introduces the 'inferon' as a primitive unit, and applies proof-theoretic tools to distributed systems modelling.
Prediction interval method for financial markets combining spatio-temporal graphs and bi-level chaotic fusion. Model with volatility-aware gating mechanism and LUBE objective. Tests on 43 NSE companies (2016-2026): Winkler score 0.0778, coverage 96.6%, significant improvements vs LSTM/GRU/GCN (p<0.001).
Systematic study of code as AI agent infrastructure. Three layers: harness interface (code connects reasoning, action, environment modeling), mechanisms (planning, memory, tool use, feedback), and multi-agent scaling. Applications: coding assistants, GUI/OS automation, embodied agents, scientific discovery, DevOps.
AI agent model with three human-inspired memory systems (short-term, episodic, semantic), each modeled as a knowledge graph. Evaluated in custom RL environment « the Room ». Deep Q-learning agent learns to encode, store, and retrieve memories to answer questions. Outperforms agent without this memory structure.
Action-Gradient MCTS (AGMCTS) combines global tree search with local gradient-based action refinement for online planning in continuous spaces. Three theoretical contributions: action score gradient theorem, Multiple Importance Sampling Tree for sample reuse, tractable gradients via Area Formula. Outperforms state-of-the-art sample-based solvers on continuous MDP/POMDP benchmarks.
Alignment evaluation across 75 LLMs benchmarked against 376 humans. Qualitative analysis derives 6 themes of optimal AI functioning (Performance, Adaptive Capacity, Social Good, Ethics and Responsibility, Relational Integration, Agency). Models reproduce human value ordering but systematically exaggerate differences. Profile fidelity does not correlate with model size or recency.
BPO, a three-stage framework (bootstrapping, extrapolation, refinement), creates a self-improving data flywheel to train robust reasoning models for long-horizon sparse-reward planning. Uses planning quaternions, long-short chain-of-thought fusion, and complexity-stratified curriculum learning. SOTA on ALFWorld, ScienceWorld, WebShop with significant token efficiency.
GVGAI-LLM is a video game benchmark for evaluating spatial reasoning and problem-solving in LLMs. Built on the General Video Game AI framework, it contains 118 ASCII games testing planning and logical reasoning. Zero-shot evaluations reveal persistent limitations of current models in spatial reasoning, partially improved by structured prompting.
R-AIRL (Reasoning Adversarial Inverse Reinforcement Learning) infers process-level reward functions from expert Chain-of-Thoughts without explicit reward definitions. Tested on GSM8K, MMLU-Pro, and MedReason: improves pass@1 by 17.4 points via inference-time reranking, outperforms SFT in post-training, localizes reasoning failures with 86.1% accuracy.
Scales++ proposes benchmark subset selection based on intrinsic task properties rather than model-specific failure patterns. Using 0.25% of data on Open LLM Leaderboard, it predicts full scores with 3.2% mean absolute error, reducing selection cost by 18x.
New method to estimate Precision-Recall curves for generative models using binary classification framework. Statistical analysis with minimax bounds and extension of existing PR metrics. Experimental analysis across various settings.
MemOCR is a multimodal memory agent that compresses long interaction histories into structured images with adaptive information density. Trained via RL with budget-aware objectives, it outperforms text-based baselines on multi-hop and single-hop QA benchmarks under extreme context constraints.
Systematic study of JEPA world models for physical planning. Authors characterize technical choices (architecture, training objective, planning algorithm) and propose a model outperforming DINO-WM and V-JEPA-2-AC on navigation and manipulation tasks. Code, data and checkpoints released.
PersonaDual is an LLM framework that adaptively switches between objective and personalized reasoning based on context. Trained via SFT then optimized with RL using DualGRPO, it preserves personalization benefits while maintaining factuality and reducing interference.
Researchers propose UALM (Unified Agent Lifecycle Management), a governance framework for agentic AI systems in healthcare. The model addresses agent sprawl through five control-plane layers: identity registry, orchestration, PHI-bounded context, runtime policy enforcement with kill-switch triggers, and lifecycle management linked to credential revocation and audit logging.
R2M (Real-Time Aligned Reward Model) introduces a lightweight RLHF framework to mitigate reward overoptimization. Instead of relying solely on semantic representations, R2M leverages evolving hidden states from the policy model to align with real-time distribution shifts during reinforcement learning training.
New DDR-Bench benchmark evaluates investigatory intelligence of LLMs: autonomous ability to explore databases and extract insights without explicit queries. Frontier models show emerging agency but struggle with long-horizon exploration. Study distinguishes investigatory intelligence (setting own goals) from executional intelligence (completing assigned tasks).
AgentArk distills multi-agent intelligence into a single LLM through three hierarchical strategies: reasoning-enhanced fine-tuning, trajectory-based augmentation, and process-aware distillation. The resulting model preserves single-agent computational efficiency while retaining multi-agent reasoning and self-correction capabilities.
RaBiT introduces residual-aware binarization training for 2-bit LLM quantization. It solves inter-path feature co-adaptation by sequentially deriving each binary path from a shared full-precision weight, ensuring error correction across layers. Results: SOTA performance, 4.49× inference speedup on RTX 4090.
LiTS is a modular Python framework for LLM reasoning via tree search. It decomposes search into three reusable components (Policy, Transition, RewardModel) compatible with MCTS and BFS. Tests on MATH500, Crosswords, and MapEval demonstrate component-algorithm orthogonality. Key finding: LLM policy diversity, not reward quality, is the bottleneck in infinite action spaces.
Researchers show that combining acoustic and expectation-related ANN representations improves music identification from brain activity (EEG). Models pretrained on both representation types outperform baselines, with complementary gains exceeding seed ensembles. This approach enables general-purpose EEG models grounded in cortical encoding principles.
Causal inference framework for financial stress testing in panel data with continuous treatment and no control group. Proposes closed-form confounding envelope parameterized by two sensitivity parameters, combines partial identification with importance-weighted conformal prediction. Shows standard predictive models remain causally biased on US unemployment data.
Explicit Logic Channel method runs parallel to MLLMs to validate and enhance multimodal models on zero-shot tasks. Combines LLM, VFM and probabilistic logical reasoning (factual, counterfactual, relational) with Consistency Rate metric for model selection without ground-truth labels. Evaluated on MC-VQA and HC-REC with 11 open-source MLLMs.
EnterpriseArena, a 132-month CFO simulator, benchmarks LLM agents' ability to allocate resources over long horizons under uncertainty. Tests across 23 models and 4 frameworks: only 15.4% of trials complete the full horizon. Larger models do not reliably outperform smaller ones. Reveals critical capability gap in managing binding commitments under partial observability.
arXiv paper introducing spatiotemporal robustness (STR) for temporal logic specifications over discrete-time signals. Formalized as multi-objective reasoning problem via partial order, STR captures joint spatial and temporal perturbations. Applicable to multi-agent systems, smart cities, air traffic control. Monitoring algorithms proposed.
HeteroFusion merges heterogeneous language models (Llama, Qwen, Mistral) by aligning functional module structures rather than raw weights, and suppressing incompatible transfer signals. Outperforms fusion, merging, and ensemble baselines on heterogeneous transfer, multi-source fusion, and cross-family generalization.
Framework enabling dialogue with non-human systems (gene regulatory networks, RL agents) by treating communication as a game. An LLM routes prompts to the system whose internal dynamics best match, without altering parameters. Shows behavioral convergence across disparate agents and reveals inductive biases of systems themselves.
LEAF is a living benchmark to evaluate LLM forecasting capabilities using multidimensional events. The system uses recursive retrieval agents and dual-agent cross-validation to provide relevant auxiliary text. Testing shows LLMs leverage signals from complex events to improve predictions, particularly on equities.
EmergentBridge improves unified multimodal embedding models for unpaired modality pairs (audio↔depth, infrared↔audio). The method learns a mapping producing a 'noisy bridge anchor' and enforces alignment in the orthogonal subspace, preserving existing anchor-alignment structure. Results across 9 datasets: outperforms baselines on zero-shot classification and retrieval.
LEGO is a modular platform for LLM-based digital front-end design generation. It decomposes the flow into 6 steps and extracts 42 reusable circuit skills. On 41 hard VerilogEval v2 problems where GPT-5.2-codex fails, LEGO achieves 80.5% Pass@1 vs 0% baseline, outperforming hierarchy-verilog (+14.6%) and VerilogCoder (+2.5%).
DataClawBench is a benchmark for exploratory real-world financial data analysis agents, containing 2.06 million raw records and 492 cross-domain tasks. Evaluation of 8 advanced LLMs shows that increased exploration does not reliably translate to task progress or correct answers.
RAC (Robust Agent Compensation) is a log-based recovery paradigm integrated into agent frameworks (LangGraph, LangChain) to prevent unintended side effects. Requires no code changes. Results: 1.5-8X better latency and token efficiency vs state-of-the-art LLM recovery approaches on τ-bench and REALM-Bench.
UxSID introduces a framework for modeling ultra-long user sequences using Semantic IDs and dual-level attention. The architecture balances computational efficiency with semantic awareness, achieving state-of-the-art performance and 0.337% revenue lift in large-scale advertising A/B testing.
MarsTSC, an agentic framework for VLMs, addresses few-shot multimodal time series classification. Three collaborative roles (Generator, Reflector, Modifier) iteratively refine a knowledge bank via reflective reasoning. Tested on 12 benchmarks across 6 VLM backbones: consistent gains, interpretable rationales.
TruthMarketTwin, a multi-agent simulation framework, studies strategic behavior of LLM agents in e-commerce markets. Agents autonomously exploit weaknesses in reputation-based governance, while warrant enforcement reduces deception and reshapes strategic reasoning.
GCPO (Group Cooperative Policy Optimization) replaces competitive rollout optimization with team-level credit assignment. Rollouts are rewarded by contribution to valid solution coverage (determinant volume over semantic embeddings), not individual accuracy. Results: improved reasoning accuracy and solution diversity across benchmarks.
LegalCheck automates drafting of objection response letters in Dutch municipalities using RAG and context-augmented generation. Deployed in Amsterdam, the system generates near-final legal drafts in minutes instead of hours, capturing 80–100% of essential content while maintaining expert review and legal consistency.
arXiv paper showing LLMs exhibit a knowing-doing gap in tool use: recognition of tool necessity vs. actual invocation diverge. Testing 4 models on arithmetic and factual QA reveals 26.5-54% mismatches. Hidden state probing shows cognition and action signals become nearly orthogonal in late layers, with most failures at the cognition-to-action transition, not in recognition itself.
OpenDeepThink uses Bradley-Terry aggregation to select best solutions from multiple parallel candidates. The system randomly compares answer pairs, aggregates votes, and preserves top candidates for mutation. On Codeforces, Gemini 3.1 Pro gains +405 Elo points across 8 LLM calls (~27 min). Authors release CF-73, 73 expert-annotated problems.
Study showing diverse ensembles of monitors detect misaligned AI agent actions better than homogeneous ensembles. 12 GPT-4.1-Mini monitors (prompting + fine-tuning) evaluated on coding tasks: best 3-monitor ensemble achieves 2.4x greater detection gain versus identical-monitor ensemble, with strong generalization on independent data.
Paper proposes hybrid hyperparameter tuning method (random grid search) for cardiovascular disease classification. Combines global exploration (random search) with focused exhaustive search (grid search). Experimental results show reduced training time and improved performance versus traditional tuning methods.
Researchers formalize activation steering (LLM control without retraining) as budget-constrained optimization over intervention layer and coefficient. They introduce concept granularity to measure directional heterogeneity and present GRACE, a framework using activation geometry to diagnose steering difficulties and reduce required evaluations by 39.8% on average.
ML study on pre-test risk stratification for Chlamydia trachomatis infection. 93 urine samples analyzed using patient-reported data and urine biomarkers. Combined models reach AUC 0.72; urine biomarkers provide reliable predictive signal complementary to clinical data.
arXiv survey on trustworthiness of RAG (Retrieval-Augmented Generation) systems. Proposes Trust-RAG Compass framework assessing 6 dimensions: factuality, robustness, fairness, transparency, accountability, privacy. Introduces TRC Bench evaluation benchmark comparing proprietary and open-source models.
Study on extrinsic gender bias in Bangla pretrained language models. Four manually annotated task-specific datasets constructed (sentiment analysis, toxicity detection, hate speech, sarcasm detection) with minimal-pair gender perturbations. RandSymKL debiasing strategy proposed, combining symmetric KL divergence and cross-entropy loss, reducing bias while maintaining competitive accuracy.
AdaGraph is a graph-native clustering algorithm that overcomes the curse of dimensionality by operating on kNN graph topology instead of Euclidean distances. Tested on 10 synthetic benchmarks (d=10 to 5000) and three scientific domains (genomics, NLP, materials science), it outperforms HDBSCAN, WGCNA, and other methods without requiring k specification.
ReAlignFit, a molecular relational learning method, improves model stability by aligning substructure representations using chemical Induced Fit-based inductive bias. Tested on 9 datasets, it outperforms state-of-the-art models and enhances robustness to distribution shifts (rule-shifted and scaffold-shifted data).
Research paper proposing to integrate AI advances into cognitive science methods to build generalizable theories. Authors argue that naturalistic experimental paradigms and computational models capable of accommodating them are necessary to understand natural intelligence and ensure theory generalization.
COLSON applies diffusion-based reinforcement learning to social navigation for mobile robots in dynamic environments. The method outperforms Gaussian policy approaches in action flexibility and adapts to unseen scenarios (static obstacles, variable objectives) without retraining.
UniversalRAG is a multi-modal RAG framework that retrieves and integrates knowledge from heterogeneous sources (text, images, videos) at variable granularities. It introduces modality-aware routing to avoid intra-modal bias and organizes each modality into granularity levels. Validated on 10 benchmarks, it outperforms single-modality and unified baselines.
arXiv paper addressing Catastrophic Overfitting (CO) in fast adversarial training. Authors propose controlling the lp training norm instead of noise injection or regularization. They quantify gradient concentration via Participation Ratio and entropy measures, developing an adaptive lp-FGSM that automatically tunes the training norm based on gradient information.
DexWild introduces a low-cost system for collecting dexterous manipulation data through natural human hand gestures. The framework co-trains on human and robot demonstrations, achieving 68.5% success rate in unseen environments (4× better than robot-only) and 5.8× improved cross-embodiment generalization.
Study shows unlearning in LLMs merely suppresses information: minimal fine-tuning restores original behavior. Representation-level analysis framework (PCA, CKA, Fisher information) reveals four forgetting regimes by reversibility and catastrophicity. Identifies cases of seemingly irreversible targeted forgetting.
FinTagging is a benchmark for evaluating LLMs on extracting and tagging financial data with XBRL. It decomposes the task into two stages: FinNI (financial numeric identification) and FinCL (mapping to full US GAAP taxonomy). Testing shows models generalize well in extraction but struggle significantly with fine-grained concept linking.