The Archive

GEASS steering mitigates object hallucination in vision-language models by asymmetric caption weighting without retraining.

EGAD entropy-guided knowledge distillation improves token-level transfer by weighting per-token importance in student model training.

GFlowNet training stability improved via loss-to-TV bounds that provide probabilistic guarantees against mode collapse.

LLMs exhibit genre-dependent credibility assessment bias, misclassifying entertainment news as fake more often than hard news.

FEDIN captures latent periodic patterns in user behavior for click-through prediction via frequency-domain spectral entropy analysis.

Motion-aware caching accelerates autoregressive video generation by adaptive denoising step allocation based on pixel motion density.

SignVerse-2M dataset adds 2M pose-annotated sign language clips across 25+ languages for improved recognition and generation.

TCDA improves conversational sentiment analysis by modeling dialogue threads with positional awareness instead of flat sequences.

CoAction framework enables single multi-task Pareto set learning model instead of separate models per task.

Route receipts propose audit trails for model routing decisions in adaptive AI systems to ensure transparency and accountability.

SplitZip achieves ultra-fast KV cache compression for disaggregated LLM serving, addressing bottleneck in prefill-decode architecture.

Federated learning framework for 5G jamming detection preserves privacy vs. centralized approaches.

Reasoning Trap formalizes why multi-agent debate preserves answer accuracy but degrades reasoning via information-theoretic bounds.

Floating-point networks under automatic differentiation can approximate differentiable functions and gradients despite finite precision.

Stability analysis of decentralized SGD/SGDA under Markov chain sampling characterizes generalization with dependent data.

Probe-Geometry Alignment surgically removes memorization traces from unlearned LLMs via cross-sequence detection without capability loss.

LLM-based agent uses adaptive code exploration to extract information from heterogeneous BIM models at runtime, evaluated on ifc-bench v2.

Reddit discussion on using Claude for project management and meeting tracking over multi-year timelines.

Empirical comparison of Qwen3.6-27B and Coder-Next models across 40 test cases shows statistical parity with task-dependent tradeoffs.

Functional taxonomy of world models organizes latent state design by downstream task (prediction, control, planning, grounding) rather than architecture.

Complex Diffusion Maps framework uses ω-parameterized kernels to extract harmonic structure from high-dimensional data.

GRAVITY module injects relational, temporal, and thematic structure into conversational memory retrieval for long-horizon agents.

MultiBreak benchmark evaluates LLM safety via scalable multi-turn jailbreaks using active learning to generate diverse adversarial prompts.

Reddit user expresses privacy concerns about Claude's Cowork feature and data handling when connecting to cloud/email services.

Engineer describes autonomous agent system with self-generating tools that can write, test, and register new capabilities without user intervention.

Large-scale evaluation of DiffDock, AutoDock-GPU, GNINA, and NMDN on LIT-PCBA library (15 targets, 578K pairs) for molecular docking.

Class-Aware Adaptive Differential Privacy framework for sensor-based fall detection using 3D CNN and BiLSTM with per-class noise tuning.

MissBGM method imputes missing data via Bayesian generative modeling with explicit missingness mechanism and posterior uncertainty quantification.

CP-SynC multi-agent workflow translates natural language to MiniZinc constraint models using synthesized checkers for semantic validation.

FPGA-based inference of 4,192-parameter MicroGPT achieves 50k tokens/sec throughput using onboard ROM weight storage.