The Archive

Roadmap for genuine recursive self-improvement in AI systems progressing from execution autonomy through meta-improvement, using Headroom-Closed Index.

Empirical study evaluates eight time-series foundation models on continuous glucose monitoring forecasting across diabetes populations with multimodal context.

DeBERTa trained on 10M words with visually grounded token embeddings shows ostensive learning persists through training but remains invisible to syntax benchmarks.

AdamX optimizer incorporates cosine similarity for adaptive update control with variance rectification, showing competitive convergence across benchmarks.

NLP analysis of US congressional speeches shows evidence-oriented language declined since 1970s and correlates with legislative effectiveness.

RetroThinker enables retrospective reasoning in streaming speech LLMs via asynchronous backchannels, improving complex task accuracy under latency constraints.

Conversational XAI interface for interpreting energy consumption ML models using LLM-based dialogue, reducing required technical expertise.

Layerwise intervention study reveals how Qwen, Llama, Gemma route query info and retrieve internal knowledge across layers during QA.

IndicTriMix dataset and models for token-level language identification in tri-lingual code-mixed text using XLM-RoBERTa and MuRIL.

Fiberwise optimal transport constructs model-aware diffusion/flow-matching schedules that minimize kinetic action and reduce prediction error.

Audit of cardiovascular screening models reveals target leakage (not model class) explains 0.89 AUROC; tabular foundation models don't resolve bias.

Multi-step transition lookahead in RL remains NP-hard at discount factors <1; paper provides near-optimal planning algorithms for finite horizons.

Survey of maritime professionals' trust in AI collision-avoidance decision assistants reveals generally positive attitudes toward MASS integration.

Logit Refiner adds sequential token sampling to visual autoregressive models to recover spatial dependencies lost in parallel decoding, improving image coherence.

Deploying a large language model is only the first step toward production-ready serving. Production teams also need to serve as many concurrent users as... Deploying a large language model is only the first step toward production-ready serving. Production teams also need to serve as many concurrent users as possible on available GPU infrastructure while preserving the interactivity that keeps applications responsive. That tradeoff matters even more for agentic AI workloads, where prompts can be long, context can be reused across steps… Source

Looped flows train recurrent inference models with local denoising objectives to enable multi-step reasoning without backprop-through-time bottlenecks.

SpecGuard detects backdoored LLMs at inference time via spectral analysis of activations without extra computation, addressing runtime trigger detection.

Language model embeddings improve multi-objective chemical reaction optimization by providing shared representations for diverse reactants beyond hand-crafted descriptors.

Switch-aware evaluation of 11 ASR and audio LM systems on English-Yoruba code-switched speech reveals performance gaps on low-resource diacritic-rich languages.

Weight spectral density metrics predict membership inference attack vulnerability without shadow models, enabling efficient large-scale privacy auditing.

Differential privacy framework anonymizes high-dimensional EEG clinical data while preserving utility for AI model training and healthcare research.

Whisper-based pipeline improves multilingual speech transcription from video for cross-cultural training data collection.

Medical LLM research evaluation lag widened 4.5× (2023–2026); 97.5% lack RCT/prospective designs as models cycle faster than clinical validation.

Controlled inversion test measures whether LLMs can undo news framing (lexis, agency, salience) while preserving facts, revealing generation-detection asymmetries.

Unified parameterization for per-token gating in on-policy knowledge distillation generalizes EOPD and ToDi methods.

Component-aware differential privacy fixes per-layer DP clipping for federated multilingual speech-LLMs with mismatched encoder/decoder norms.

RAG-Safety-Bench evaluates safety risks when LLMs retrieve from external knowledge bases, showing RAG can degrade response safety.

SIRF internalizes platform safety policies into LLM weights via continued pretraining for low-latency industrial content risk control.

Branch-and-bound algorithm for sparse portfolio optimization under ellipsoidal uncertainty; no LLM or AI model relevance.

py-kvcache characterizes external KV cache tradeoffs in vLLM across GPU/CPU/NVMe for long-context LLM serving optimization.