The Archive

HORST: modular optimizer composing L∞ and L1 geometries for stable, sparse transformer training.

Typed tensor language formalizing federated learning via client-local computation and mergeable aggregation.

VerbatimRAG system for hallucination-free QA over ACL Anthology via extractive retrieval and verbatim text spans.

WCXB: web content extraction benchmark with 2,008 pages across 7 content types for retrieval and LLM training.

UOTIP: unpaired inverse problem solver using unbalanced optimal transport for image reconstruction.

Error correction mechanisms for deep autoregressive time-series forecasting to mitigate long-term prediction drift.

https://reddit.com/link/1tik0qe/video/9bh6ypr3ca2h1/player A few hours, [Claude Code](https://www.claude.com/product/claude-code) \+ [Remotion](https://www.remotion.dev/), 4 black coffees, no design tools, no After Effects, no editor. **The whole trick:** Remotion is React for video. You write JSX, you get an mp4. Every animation is `interpolate(frame, [start, end], [from, to])`. That means **Claude Code can write the entire video for you** — it already knows React, animation is just numbers, and you can iterate the same way you iterate on a landing page. Change a value, re-render, see w...

LoCar introduces evaluation framework for in-vehicle LLM assistants with focus on Korean honorific stability and localization.

Decoupling Communication from Policy decouples latent representations in MARL to enable robust multi-agent coordination under bandwidth constraints.

AIMBio-Mat proposes AI-native FAIR framework linking materials provenance, knowledge graphs, and active learning for biomedical discovery.

AutoRPA distills LLM reasoning into efficient code synthesis for repetitive GUI automation tasks, bridging ReAct and traditional RPA.

User feedback on Claude Chat UI/UX: question popup blocks content and disrupts workflow.

Musical Attention Transformer incorporates meta-information (bar, key, tempo) into attention mechanism to reduce repetition in music generation.

GradeLegal evaluates LLM capability to automatically grade German legal exam solutions in criminal and public law domains.

Early-stage Gemma 4 MTP compilation work-in-progress shared on LocalLLaMA.

SpectralEarth-FM integrates hyperspectral imagery with multisensor Earth observation data via hierarchical transformer with spectral tokenization.

Fine-grained Claim-level RAG Benchmark for Law provides granular evaluation of legal RAG systems to detect hallucinations at claim level.

Self-Pretraining analysis investigates why masked token prediction pretraining on Transformers improves sequence classification without external data.

Robust Personalized Recommendation mitigates hidden confounding in MNAR observational data via novel causal inference approach for recommender systems.

APM benchmark for evaluating style personalization in LLMs using arbitrary preference mappings without reference responses.

Driving VLA redesigned via inverse kinematics framework to improve trajectory prediction by grounding visual tokens in dual boundary conditions.

Vector quantization-based multiclass calibration method for ML models addressing heterogeneous calibration errors across latent space.

Theoretical framework for training multimodal LLMs using only pairwise modality alignments instead of full joint multimodal datasets.

Position paper bridging causal representation learning and traditional representation learning via unified problem formulation.

Transformer-based mutation operator for Cartesian genetic programming applied to approximate circuit design optimization.

Multilingual whole-brain encoding study confirms LLM-brain alignment for language comprehension across Mandarin, English, French.

Qwen 3.6 35B MoE benchmark on RTX 5080: 56 tok/s at 128k context; Multi-Token Prediction offers no speed gain at scale.

OpenAI announces multi-year partnership in Singapore for AI deployment, talent development, and enterprise/public sector adoption.

Equivalence between Gaussian processes and linear diffusion models enabling likelihood-guided conditioning beyond conjugate settings.