model choice matters in causal drug safety detection
a study compares classification models in the inferbert framework for pharmacovigilance, finding that simpler models can suffice and domain-specific pretraining helps.
topic
machine learning
editing multimodal models fails when inputs split
knowledge edits in multimodal large language models often revert to old facts when image-text pairs are separated, due to modality-specific knowledge storage.
fixing shape-prior shortcuts in long-range fringe profilometry
a study uses mechanistic interpretability and uncertainty quantification to diagnose and repair neural network failures in single-shot 3d imaging at long distances.
kv cache as editable, composable notes
prefix caching fails when one field changes, but treating the kv cache as a notebook of memoized conclusions enables editing and composing precomputed skills.
llm engineer skills roadmap for 2026
a step-by-step guide covering foundations, prompting, retrieval, fine-tuning, and serving for building production llm applications.
policy regret for embedding model routing
a new algorithm tackles dynamic routing of queries to multiple embedding models under adversarial conditions and bandit feedback.
gpu workflow for hypersonic flow emulators
a fully gpu-based method speeds up building neural emulators for hypersonic flows with physics-aware refinement and uncertainty quantification.
mobile npu framework speeds up diffusion llm inference
llada.cpp is the first npu-aware framework for accelerating diffusion large language models on smartphones, using multi-block speculative decoding and other techniques.
olmo-eval: a workbench for iterative llm evaluation
olmo-eval is an open evaluation workbench for comparing model checkpoints during development with per-question analysis and flexible runtime policies.
7 ways to fund your startup idea
seven funding options for startups, from bootstrapping to venture capital, with pros and cons for each.
pytorch meetup singapore draws apac ai builders
eighty engineers gathered for the first pytorch meetup singapore, covering inference, distributed training, and community governance.
run claude code with local models
how to connect claude code to ollama, lm studio, or llama.cpp for zero-cost, rate-limit-free coding sessions using local models.
3 numpy tricks for faster numerical code
learn vectorization, in-place operations, and memory views to speed up numpy code and reduce memory use.
open source community backs openenv for agentic rl
openenv becomes a community-governed interoperability layer for reinforcement learning environments, backed by major ai organizations.
epistemic uncertainty is not always reducible
new research shows that the standard definition of epistemic uncertainty as reducible by more data is inconsistent with its common mutual-information measure, and proposes a three-part taxonomy.
build a minimal feature store from scratch
a hands-on guide to building a feature store with duckdb, parquet, redis, and fastapi, covering offline and online stores, materialization, and retrieval for both predictive ml and llm context.
pytorch profiling: from linear to fused mlp
a deep dive into pytorch profiling shows how nn.linear fuses bias into gemm kernels and how torch.compile removes cpu overhead in mlps.
steering away from sycophancy also cuts factual agreement
activation steering meant to reduce sycophancy in language models also lowers agreement with true statements, showing a structural overlap that current methods cannot separate.