The Archive

BiasFlow diagnostics toolkit monitors spurious feature reliance via centroid alignment metrics and paired regularization penalty.

Framework interprets Multimodal Diffusion Transformers' contextual tokens by mapping them into LLM input space via lightweight bottleneck.

Recursive Video In-Context Learning hierarchically summarizes robot demonstration videos for LLM-agent planning without full-frame overhead.

Sequential Monte Carlo diffusion sampler initializes at intermediate timesteps via pulled-back posteriors and Gaussian bridges.

Looped language models exploit fixed-point structure for 1.79x faster prefill distillation, KV-cache sharing, and 2x RL gradient speedup.

UniSlider maps image-editing strength parameters to perceptually uniform slider ranges via monotonic trajectory enforcement.

MemPilot orchestrates on-demand, query-aware multimodal memory curation for LLM agents with tunable latency-cost-performance tradeoffs.

CLIFT trains web agents via conformal self-verification on natural-language rollout questions; reduces expensive frontier-model judge calls.

PlotGround benchmark for evaluating plot digitization accuracy using real scientific figures paired with source data.

TasteVal benchmark measures experimental research taste of frontier LLMs via iterative experiment design and compute efficiency.

SeqSmoother uses transformers to estimate sleep heart rate from wrist accelerometers without optical sensors.

Study of differential privacy in online learning with separation bounds between online learning and prediction under Littlestone classes.

Paradee distills Kokoro-82M TTS to 8M parameters maintaining voice quality with 10x fewer params and 15x less compute.

TAPDreamer demonstrates transferable adversarial patch attacks against world action models using public encoders.

Efficient protocols for agnostic tomography of pure bosonic Gaussian states with polynomial runtime bounds.

Block Disentanglement framework advances identifiability guarantees in causal representation learning for visual state estimation.

H-JEPA introduces hierarchical world models with multi-timescale planning via stacked action-conditioned JEPAs for visual navigation.

On-policy distillation technique trains LLMs to sample from power-distribution sharpened outputs without search-time compute overhead.

MC-Sparse reduces latency in diffusion transformers via training-free sparse attention, addressing quality degradation from token grouping and interaction selection errors.

Linear response theory framework validates ML emulators of chaotic systems via forced-response testing, not just forecast skill or long-run statistics.

Round-Trip KNN Clustering detects hierarchical clusters across scales without prespecifying cluster count, using bidirectional nearest-neighbor graphs.

Back-to-the-Future agentic framework automates chip verification beyond RTL generation, tackling post-simulation debugging and waveform analysis.

Systematic procedure derives robust scaling laws for HEP ML models using ATLAS JetSet2 (~11B jets), enabling budget-aware architecture comparisons.

T-Search open-weight agentic retriever on Qwen3.6-35B uses GSPO training for multi-step search with evidence ranking and justifications.

IdeaLens detects idea provenance (human vs. AI origin) in long-form writing via outline representation, addressing policy questions beyond prose authorship.

Learned heuristics (CNN, GATv2 with DQL) select optimal agent corners in orthogonal guard art galleries while preserving formal coverage guarantees.

Analysis of missing limits in neural PDE solvers shows unbounded parameters prevent attaining best loss; kernel derivatives recover missing functions.

ARBOR parameter-efficient medical LLM adapter selects rank-one components per question using specialty/operation tags, avoiding fixed-rank approximation floors.

Founded by two sisters, Hot Girl Hotline uses AI to give young women personalized dating and relationship advice, with an emphasis on safety and avoiding emotional dependency.

MatrixFormer pre-trained transformer predicts full distributions for missing matrix entries in single forward pass, trained on synthetic low-rank matrices.