The Archive

Robust CurveMoE uses mixture-of-experts with mode connectivity to defend Mixture-of-Experts models against multi-norm adversarial perturbations efficiently.

Trace Integrity framework audits LLM data agents by verifying computation traces are explicit, executable, schema-valid, and replayable—not just answer-correct.

Dialogue model treats referential uncertainty conservatively via Bayesian belief revision, preferring preservation over immediate update on local mismatch.

Alibaba released the model weights for Qwen3.8-Flash-Next as a preview of the upcoming Qwen4 architecture for developers to experiment with and evaluate. It’s... Alibaba released the model weights for Qwen3.8-Flash-Next as a preview of the upcoming Qwen4 architecture for developers to experiment with and evaluate. It’s a multimodal mixture-of-experts (MoE) model with a 125B-parameter main model supplemented by an additional 51B N-gram embeddings, with 6B parameters activated per token. It has a native 262,144-token context window, extensible to 1M tokens… Source

DualOPSD adapts both student and privileged teacher policies during training to make supervision responsive to learner distribution shifts.

Google DeepMind launches Gemini 3.5 Transcribe, an upgraded speech-to-text model with improved contextual understanding.

Google has updated Gemini Audio with some new Gemini 3.5 models, introducing new transcription capabilities that automatically detect specialized jargon and more than 85 languages. Gemini 3.5 Live, 3.5 Live Experimental, and 3.5 Transcribe are designed to provide better precision for Google's voice-controlled AI features, without struggling with background noise or when your speech is interrupted. Gemini 3.5 Transcribe is a completely new addition to the Gemini family, and its introduction comes as we're still waiting for Google to release the Gemini 3.5 Pro model that it promised to roll out...

VISA: agentic self-evolving framework for multimodal instruction synthesis that iteratively filters constraints and improves data quality via feedback loops.

Graph neural network trained via imitation learning for automated theorem prover proof construction in connection-tableau calculi.

Self-evolving multi-agent test-time defense against LLM jailbreak attacks using persistent rule memory to adapt to emerging attack strategies.

VoiceMem: dual-brain streaming memory architecture for conversational speech language models with separate informational and emotional processing.

AsymSpec: asymmetric speculative decoding for agentic LLMs that compresses verifier context while drafting on full input to reduce inference latency.

CardioFusion-AI: ECG-PPG sensor fusion framework for wearable physiological monitoring robust to signal degradation and sensor dropout.

Analysis of conceptual vs. referential disruptions in human reading and LLM processing reveals distinct temporal dynamics in linguistic meaning grounding.

ProgRouter: online progress-guided routing for multi-agent LLM workflows that adapts to dynamic task state and cost-quality tradeoffs.

Spectral analysis reveals why Muon optimizer outperforms Adam for LLM pretraining via anisotropic, stable spectral allocation across loss landscape.

Multi-granularity context enhancement for RAG over multimodal knowledge graphs to reduce LLM/MLLM hallucination.

FRAME framework separates sampling variation from representational bias in medical imaging fairness audits.

Layer-wise MBTI personality analysis of quantized LLMs reveals trait changes across precision levels.

ASR models reveal phonological traces of first language persist in pre-phonemic layers during second-language training.

SciMIF benchmark evaluates MLLMs' complex instruction-following across 22 tasks in 5 scientific domains.

PANDA framework transfers auxiliary modality signal to primary-modality medical models without paired data at inference.

Theoretical analysis of robust MDPs with linearly-defined uncertainty sets and ω-regular parity objectives.

LivingRAG adds writable experience store to Graph RAG, enabling reuse of verified reasoning across queries.

Lovable extends AI web app builder with MCP-powered capabilities for agent integration in SaaS workflows.

Spectral analysis detects AI-generated text by capturing signal fluctuations absent in confidence-based metrics, offering training-free scalability.

Study shows magnitude pruning degrades sparse autoencoders' interpretability of LLMs; activation-aware methods like Wanda preserve SAE robustness.

Audit of 51 physical AI models on 12 benchmarks reveals redundancy in benchmark suites, affecting model ranking reliability.

Multi-agent systems fail despite correct candidate answers; LLM judges improve selection under high candidate diversity.

On-policy self-distillation trains models on own generations with privileged teacher information; critical review of feasibility and scaling.