GuardedRepair is a guarded best-of-N repair framework for LLM mathematical reasoning that selectively fixes incorrect traces while preserving correct answers. On GSM8K (95.60% → 96.89%), it fixes 17 of 58 errors with no measured broken-correct cases. On weak-reasoner ASDiv, accuracy improves from 78.40% to 87.60%.
Researchers replicate Picbreeder (interactive image evolution platform) by replacing human users with Vision Language Models (VLMs). Results show qualitative differences from human baseline. Study of causal factors: exploratory noise, behavioral diversity between agents, memory of past actions.
Preregistered study shows current LLMs are overconfident: confidence exceeds accuracy on average. A hard-easy effect moderates this bias: overconfidence peaks on difficult tasks, while easy tasks show substantial underconfidence. Introduces LifeEval, a benchmark for evaluating model calibration across difficulty levels.
arXiv paper analyzing latency-reliability-cost tradeoffs in LLM-enabled multi-agent workflows. Introduces performance models for LLM and non-LLM agents, proposes water-filling token allocation policy, and characterizes optimal workflow reliability via shadow prices under latency and cost constraints.
Paper on runtime enforcement of Reconstructive Authority (RAM) in autonomous agent systems. Introduces execution model with three states (admit/deny/halt), dynamic dependency resolution, and Recovery Loop integrating drift detection with execution control. Guarantees no action executes without constructible authority.
Survey paper at the intersection of human-body communication (HBC) and federated learning for wearable sensor networks. Proposes taxonomy of FL deployments (intra-body, body-hub, cross-user, clinical-cloud) and introduces BODYFED-HBC reference architecture with scheduling algorithm and reproducible simulation combining public datasets with empirical HBC signal-loss models.
Learning-assisted hyper-heuristics for Job Shop Scheduling (JSSP). Proposed selector uses regret-normalized rollout labels, contextual KNN uncertainty estimation, and a gate that acts only when predicted gain exceeds uncertainty-adjusted margin. Reduces Random-HH mean RPD by over an order of magnitude on synthetic instances.
Thinking-learning interaction model for autonomous robots in changing environments. Thinking guides learning (change identification, evidence selection, planning), learning improves thinking (knowledge updates, action strategies). Results: recognition accuracy 0.419→0.845, action length 13.0→4.0, evidence selection rate 0.272→0.965.
MAPLE, a tree search method, extends AlphaZero to imperfect-information games by aggregating policy and value evaluations from multiple sampled world states. Tested on Phantom Go and Dark Hex, MAPLE outperforms PIMC-AlphaZero baseline with Elo gains of 291 and 136.
Neuro-inspired framework for embodied planning and control. The Inverter uses Inverse Learning (IL) to generate multi-step action sequences. Outperforms offline-RL and diffusion-planner baselines on D4RL (+24.2% average) with 100-1000x less inference compute. Application: single-qubit quantum gate synthesis matching GRAPE fidelity at 1000x faster per-gate compute.
MODIAD framework for multimodal industrial anomaly detection in distributed online settings. Introduces SMG algorithm for multi-class scheduling and REC-LoRA strategy reducing computational overhead. Validated on MVTec 3D-AD and Eyecandies datasets.
CAFD is a DNN fault detection method combining model signals, distance features, and a novel Concept Failure Ratio (CFR) leveraging Vision-Language Models. Evaluated on ImageNet and three models, CAFD outperforms 5 baselines with average 18.3% improvement in Fault Detection Rate.
Verifiable Transformers framework converts task-localized Transformer circuits into solver-checkable formal claims. Extracts circuits and verifies functional equivalence, edge necessity, invariance, and robustness via SMT encoding. Demonstrates direct verification on symbolic tasks and surrogate-mediated verification at GPT-2 scale with SMT-representable operators (Signed L1 BandNorm, sparsemax, LeakyReLU).
arXiv paper proposing an online aggregation mechanism to align LLMs with human feedback in mobile crowdsourcing. The system incentivizes truthful preference reporting from strategic workers via a dynamic Bayesian game, reducing regret from O(T) to O(√T) over T time slots.
Cascade-KDE is a training-free restoration method for time series corrupted by Gaussian noise and impulse outliers. It estimates temporal-amplitude density, applies Density-Truncated Robust Expectation to limit anomaly influence, then refines via exponential cascade. Tested on ECG and battery degradation, it preserves derivative peaks better than classical filters.
arXiv paper showing consecutive transitions in on-policy RL are redundant and cause hidden instability. Randomly dropping 25% of transitions in PPO stabilizes training (KL divergence, entropy, value estimates) without degrading rewards, across CartPole, Acrobot, LunarLander, HalfCheetah, Hopper.
Heteroscedastic uncertainty-aware PINN framework for flood extent mapping from SAR data. Attention-Gated FNO-UNet with dynamic Warm-Start protocol and aleatoric uncertainty modeling prevents gradient divergence ("Physics Shock"). On Sen1Floods11: +25% relative IoU improvement over deterministic baselines, with calibrated confidence bounds for disaster response.
Riemannian archetypal analysis using data-driven pullback geometry on deformed star distributions. Combines interpretability of classical archetypal analysis with non-linear model expressiveness. Riemannian archetypal mapping (RAM) projects onto manifolds of geodesically convex archetype combinations. Experiments on MNIST demonstrate meaningful geodesics and geometry-aware denoising.
Unified framework (FPMC) modeling denoising functions in diffusion models. Consolidates existing approaches through query precision vectors, response weights, and source distributions. Improves performance via soft relaxations and distribution augmentations.
PrivFusion is a privacy-preserving multi-agent framework that automates harmonization of heterogeneous datasets before federated training. Tested on four COVID-19 datasets, it uses agents to analyze local data, cluster semantically similar features, and recommend iterative transformations without centralizing sensitive health information.
Decision support framework for digital therapeutics modeling both recommendation and adherence effects using linear dynamical systems. UCB-BOLD algorithm proposed for online treatment selection with sublinear regret guarantees. Evaluation on micro-randomized trial data: 2-3x lower conditional value-at-risk regret than benchmarks.
ChainzRule replaces standard activations with learnable polynomial layers governed by Differential Regularization (DREG), a Jacobian penalty computed analytically during forward pass. Tested across tabular, NLP, and vision: 85.71% on Pima Diabetes, 46.20% on SST-5 with frozen encoder (5% of RNTN training data), 55.79% on SST-5 fine-tuned BERT, +2.32% on CIFAR-10-C. Improves robustness and sample efficiency.
Theoretical study of scaling laws for sketched linear regression with mini-batches. Comparative analysis of one-pass SGD, multi-pass SGD with and without replacement. Key result: variance O(min(M,(T_eff*γ)^(1/a))/(B*T_eff)), 1/B reduction in multi-pass without-replacement regime, zero fluctuation at B=N.
Study of 'omissive bias': LLMs systematically omit religious perspectives in responses to everyday ethical questions. AllFaith Benchmark with 150 questions (grief, forgiveness, relationships) evaluated across 27 models. Finding: asymmetric religious underrepresentation, more pronounced for practical situations than existential questions.
An LLM-based framework extracts segment disclosures from Form 10-K filings to improve completeness and comparability of financial data. The system uses RAG to integrate information across multiple periods and firms, demonstrating effectiveness for longitudinal analysis and cross-firm geographic alignment.
Modular pipeline for educational analogy generation in four stages (source finding, sub-concept generation, explanation, evaluation). Evaluation of 12 LLMs across two annotated datasets (SCAR, ParallelPARC). Sub-concepts improve explanation quality and retrieval precision. Claude Sonnet 4.6 aligns better with human rankings than absolute scores.
Dialect-aware phonetic framework for Vietnamese speech recognition. Decomposes syllables into structured phonetic components mapped to dialect-specific IPA representations. On UIT-ViMD dataset, matches wav2vec2-base-vi-250h performance with fewer parameters and no external pretraining.
Unveil is a visual-textual embedding framework for multi-modal document retrieval. It integrates textual and visual features through knowledge distillation, transferring semantic capabilities from a visual-textual model to a purely visual model. Results: improved retrieval accuracy and efficiency without parsing.
BoxLitE is a knowledge base embedding model for DL-Lite^H using convex optimization. It maps concepts to convex regions in vector space to represent ontological hierarchies. For any satisfiable DL-Lite^H KB, BoxLitE produces a weakly faithful embedding.
Survey on reproducibility of AI systems in regulated finance (credit, fraud, AML). Identifies three sources of nondeterminism: post-hoc explanation variance (tabular models), stochastic sampling (graph networks), batch-dependent divergence (LLM agents). Proposes evaluation framework with RBO, D_cos, TDI, PSD metrics for audit readiness.
arXiv paper arguing LLMs fail at causal reasoning and long-horizon planning due to lack of world models. Authors introduce Latent Dynamics Inference (LDI) and Flux, a sequential reasoning environment specified in natural language. RL agents with explicit latent state access achieve 79% win rate vs 11% for LLMs, revealing failures in persistent state tracking.
EvoSci is a bio-inspired multi-agent framework for scientific discovery using LLMs. It integrates evolution, knowledge graphs, and specialized agents (mentor, researcher, reviewer) to iteratively generate, evaluate, and refine research ideas. On real-world topics, EvoSci achieves ICLR peer-review score of 4.90 and Top-10 ranking of 54%.
Novel training method for input-convex neural networks (ICNNs) using an unconstrained hypernetwork that emits inter-layer weights. Approach inspired by parameter-extension lifts from PDE-constrained inverse problems, circumvents limitations of projected gradient descent and softplus reparametrization. Results on log-concave density estimation and convex-potential normalizing flows show improved convergence.
Study on optimal model selection in LLM ensembles. Authors reframe proposer selection as a combinatorial problem based on complementarity rather than accuracy alone. Greedy algorithms tested on small labeled sets to balance performance and computational cost.
KREPE, a generative representation learning method for hyper-relational knowledge graphs, uses masked discrete diffusion to generate complete facts from partially observed queries. Unifies link prediction and fact generation in a single framework, outperforming LLM-based baselines on standard benchmarks.
Agent-ToM is a learning-to-monitor framework using Theory-of-Mind reasoning to detect covert malicious behavior in autonomous LLM agents. It infers agent beliefs, intent hypotheses, and behavioral deviations from task-consistent baselines. Evaluated on SHADE-Arena and CUA-SHADE-Arena benchmarks, it outperforms ensemble monitoring baselines with a two-call reasoning pipeline.
New arXiv paper proposing DINOSaur, a training-free method for continual anomaly detection in industrial settings. Combines frozen DINOv3 backbone, spatially-indexed coreset memory, and neighborhood-restricted anomaly scoring. Achieves zero forgetting, outperforms all baselines across 5 protocols, runs <100ms inference on Jetson Orin Nano with on-device adaptation <30s.
Anubis OSS v3.6, macOS app for benchmarking local LLMs (Ollama, LM Studio, MLX), adds direct model downloads from UI. Available via Homebrew and direct download. Call for testing on Apple Silicon. GPL-3.0, open-source, leaderboard with 400+ runs.
Firecrawl now available on Vercel Marketplace. Vercel teams can power AI agents and applications with structured web data without managing crawling infrastructure. Key features: scrape pages to markdown/HTML/structured data, search and retrieve full page content, interact with dynamic websites via AI prompts.
A lawyer shares experience running a 12 V100-SXM2 32GB cluster for local legal document drafting. After abandoning vLLM due to GPU Volta incompatibility with MoE models, he switched to llama.cpp with Gemma-4-26B and Qwen3.5-122B. Dense models on V100 are inefficient (~20-28 tok/s); MoE models achieve 50-113 tok/s decode on long-context legal prompts.