The Archive

Analysis shows CIR benchmarks can be solved with single-modality embeddings, questioning necessity of multimodal composition.

Websites can fingerprint LLM browser agents with 96% F1 accuracy via UI interaction traces, enabling targeted exploits.

Study identifies systematic imbalanced forgetting patterns in class-incremental learning with rehearsal-based mitigation.

NVIDIA releases quantized NVFP4 versions of Moonshot AI's Kimi-K2.6 and Kimi-K2.5 models via Model Optimizer with benchmark results.

User reports unexpected Claude API usage reset timing discrepancy; potential billing or rate-limit system bug.

Conservative Peng's Q(λ) algorithm for offline RL using multi-step value estimation in fixed behavior distributions.

DDPG-based approach for criminal identification in complex datasets with reduced false positives.

Oscillatory data-volume scheduling method that dynamically adjusts training data selection ratios for efficiency.

BioHuman10M dataset enables muscle activation inference from video via simulation-based biomechanical annotation.

MediaClaw multimodal agent platform unifies fragmented AIGC capabilities with pluginized architecture and workflow orchestration.

Concept-based compositional framework for controllable de novo crystal generation via vector-quantized VAE.

Real-time streaming speech-to-text translation system combining speech recognition and translation in SpeechLLM architecture.

Persian MusicGen adapts MusicGen to Persian tonalities and Dastgah systems using 900-hour culturally-specific dataset.

Scenema Audio releases open-weights model for zero-shot expressive voice cloning, decoupling voice identity from emotional performance via separate control prompts.

Information Filtering Networks and Homological Neural Networks combined to study compositional sparsity as structural prior for DNN design.

Anthropic launches Claude Certified Architect exam covering evals, RAG, multi-agent orchestration, and LLM integration pitfalls.

Reddit thread on daily Claude usage patterns, from document analysis to agent building workflows.

LLM-based preference interviews paired with semantic feature extraction outperform human judges on personalized image aesthetic assessment.

Crys-JEPA addresses stability-novelty trade-off in crystal generation via embedding screening and generative refinement for materials discovery.

RNN-ProVe probabilistically verifies RNN-based policies in partially observable RL without restrictive assumptions or coarse approximations.

XDomainBench diagnostic benchmark stress-tests LLM compositional reasoning across interdisciplinary scientific knowledge with interactive workflows.

Two-stage knowledge distillation framework addresses student misconception classification via cognitive uncertainty guidance on edge devices.

EVA model editing defense mitigates textual and visual jailbreak attacks on LLMs and VLMs without safety-utility trade-off via targeted edits.

Non-linear intervention framework extends LLM mechanistic understanding beyond Linear Representation Hypothesis to implicitly encoded features.

Video2GUI extracts GUI interaction trajectories from unlabeled Internet videos for large-scale GUI agent pretraining without manual annotation.

Value-filtered decoding selectively applies safety steering at test-time, avoiding unnecessary interventions that degrade helpfulness and coherence.

Study shows LLM-based financial governance lacks behavioral compliance; proposes five rationale-level metrics and mechanical enforcement approaches.

Reinforcement learning method combining Goal-Space Planning and DDPG for demand response scheduling with terminal constraints.

Task-aware layer pruning improves OOD generalization but not ID accuracy in LLMs; geometric explanation via norm/distance profile divergence.

Audio-visual speech extraction system IsoNet uses spatial cues and face embeddings on compact 4-microphone arrays with curriculum learning.