---
title: "Machine Learning"
description: "622 posts about Machine Learning, summarised, each linking to the original."
---

# Machine Learning
> 622 posts about Machine Learning, summarised, each linking to the original.

## Articles

### [Up to 3.2x Faster Inference with LFM2.5-DSpark](https://yomu.fyi/post/up-to-3-2x-faster-inference-with-lfm2-5-dspark.md)
- Company: [Hugging Face](https://yomu.fyi/company/hugging-face.md)
- Author: xx, Leonie Monigatti, Fernando Fernandes Neto, Tarek Dakhran, nathan ranchin
- Published: Aug 20, 2026
- Type: Benchmark

Liquid AI released DSpark draft checkpoints for LFM2.5-1.2B-Instruct, LFM2.5-2.6B, and LFM2.5-8B-A1B, adding speculative decoding with a minimal memory increase. DSpark uses a DFlash-style parallel backbone conditioned on target context features, a lightweight sequential Markov-chain head, and a confidence-scheduled verifier that prunes low-confidence suffixes. The draft models are simplified attention-only networks with five layers and a block of nine, trained for 15 epochs on SFT, chat, code, and function-calling data; each contains roughly 300M parameters. Across five datasets, LFM2.5-2.6B averaged 2.67x speedup on an H100 and 2.27x on an M4 Max, while reducing multi-tool function-calling latency by 57% on average. Greedy decoding remains identical to baseline by construction, and integrations are available in llama.cpp and SGLang, with on-device gains for the 8B-A1B model limited by its current MoE Metal implementation.


### [How Generative Recommenders Are Redefining RecSys at Scale](https://yomu.fyi/post/how-generative-recommenders-are-redefining-recsys-at-scale.md)
- Company: [NVIDIA](https://yomu.fyi/company/nvidia-developer-blog.md)
- Author: Elizabeth Goodman
- Published: Aug 20, 2026
- Type: Explainer

Generative recommenders (GRs) recast recommendation as next-item or next-action sequence prediction conditioned on user histories, replacing traditional embedding-similarity objectives and potentially unifying retrieval and ranking. The post presents HSTU, which uses timestamp-ordered interleaved items and actions, SiLU-based attention weighting, relative attention bias, and gating, alongside Semantic IDs, which derive a smaller vocabulary through hierarchical clustering of item embeddings. For production, recsys-examples combines TorchRec, DynamicEmb, Megatron-Core, recommender-specific KV and decoding components, while nv-embedding-cache provides tiered GPU, CPU, and remote embedding storage. On an H100, the Semantic ID path was reported as 1.85x higher in online throughput and 46% lower median latency than SGLang under the stated long-context, beam-search workload, and DLRM v3 reached 99,997 queries/sec.


### [Multilingual Knowledge Transfer under Data Constraints via Lexical Interventions](https://yomu.fyi/post/multilingual-knowledge-transfer-under-data-constraints-via-lexical-int.md)
- Company: [Apple](https://yomu.fyi/company/apple.md)
- Published: Aug 20, 2026
- Type: Problem & solution

Cross-lingual knowledge transfer is critical for training multilingual language models when target-language data is scarce and downstream tasks demand reasoning or world knowledge. Existing transfer approaches often depend on scarce resources such as large parallel corpora, translation systems, auxiliary models, or extra training phases. To address this limitation, the authors introduce LINK, a data-level intervention method that operates directly during pretraining via bilingual vocabularies. LINK swaps randomly selected words in a portion of the high-resource English corpus with word-level translations according to a specified replacement ratio. Across evaluations spanning eight languages and five model sizes, the technique delivered notable performance gains on downstream target-language tasks while achieving up to a 2x speedup in training to reach equivalent performance.


### [Scaling Laws for Mixture Pretraining Under Data Constraints](https://yomu.fyi/post/scaling-laws-for-mixture-pretraining-under-data-constraints.md)
- Company: [Apple](https://yomu.fyi/company/apple.md)
- Published: Aug 20, 2026
- Type: Problem & solution

Scaling language models demands large volumes of data, but high-value target sources like low-resource languages and specialized domains remain inherently scarce. Practitioners frequently combine limited target data with abundant generic data, creating a balance between domain underexposure and excessive repetition leading to overfitting. An empirical evaluation across more than 2,000 language-model training runs demonstrates that mixture training tolerates substantially higher data repetition than single-source training, allowing scarce target corpora to be repeated 15 to 20 times. Researchers introduced a repetition-aware mixture scaling law that accounts for both the diminishing returns of repeated target tokens and the regularizing effect of generic data. Optimizing this scaling law offers a principled framework for calculating effective mixture configurations and practical pretraining recommendations under severe data constraints.


### [Progressive Refinement: An Iterative Pseudo-Labeling Approach for Mandarin-English Code-Switching ASR](https://yomu.fyi/post/progressive-refinement-an-iterative-pseudo-labeling-approach-for-manda.md)
- Company: [Apple](https://yomu.fyi/company/apple.md)
- Published: Aug 20, 2026
- Type: Problem & solution

Code-switching automatic speech recognition faces major performance challenges due to the scarcity of code-switching training data. To mitigate this data limitation, an iterative pseudo-labeling framework leverages a large unlabeled audio corpus to generate semi-supervised speech datasets. The system structure executes across three sequential phases: pseudo-label generation, two-stage bilingual model pre-training with supervised code-switching fine-tuning, and progressive iterative improvements. Successive refinement cycles further enhance the neural model's recognition accuracy when processing complex code-switching scenarios across alternating languages. When evaluated on the SEAME benchmark, this training pipeline achieved Mix Error Rate reductions of 6.35 percent on the devman subset and 8.29 percent on the devsge subset.


### [Building Menu Vision: Real-Time Dish Recognition](https://yomu.fyi/post/building-menu-vision-real-time-dish-recognition.md)
- Company: [Yelp](https://yomu.fyi/company/yelp.md)
- Author: Arpitha Dudi, Growth Tech Lead
- Published: Aug 20, 2026
- Type: Problem & solution

Yelp developed Menu Vision to help diners recognize menu items and view corresponding user-generated photos and reviews in real time using their phone cameras. The feature evolved from a two-day hackathon Android prototype into a production system deployed across iOS and Android. Client devices perform all optical character recognition, text cleanup, and dish matching locally using native machine learning frameworks, while dish metadata is prefetched upon opening a restaurant page. Backend pipelines combine partner data, reviews, and photo captions into a Cassandra database, filtering exclusively for dishes with photos. To handle optical character recognition errors and name variations, the matching pipeline runs exact matching, bidirectional substring matching, and Jaro-Winkler similarity matching.


### [How we think about text classification in the LLM era](https://yomu.fyi/post/how-we-think-about-text-classification-in-the-llm-era.md)
- Company: [Medium](https://yomu.fyi/company/medium.md)
- Author: Raphael Montaud
- Published: Aug 19, 2026
- Type: Problem & solution

Medium revisited its aging NSFW text-classification model, which had outdated code, infrequent retraining, and poorly documented training and performance history. The replacement uses an LLM prompt containing NSFW guidelines and story content, requesting a 0-to-100 score, structured output, and a short reason rather than requiring a training set. To evaluate it, the team created a curator-labeled dataset with deliberate oversampling of NSFW stories, then compared several LLMs with the legacy model using false-positive and false-negative metrics. Some models matched or exceeded the legacy model; the team selected Mixtral 8×7b based on cost and performance and ran inference in Snowflake. The post presents this as a strong fit where training data is costly, trends are relatively stable, and only distribution-eligible stories need scoring, while topic and spam models still benefit from continuously updated datasets.


### [Building Federated Multimodal AI Workflows with NVIDIA FLARE](https://yomu.fyi/post/building-federated-multimodal-ai-workflows-with-nvidia-flare.md)
- Company: [NVIDIA](https://yomu.fyi/company/nvidia-developer-blog.md)
- Author: Tanya Lenz
- Published: Aug 19, 2026
- Type: Explainer

Federating vision-language models (VLMs) across institutions requires keeping images, text, and prompts local while coordinating training and combining updates from sites with different task and modality mixes. NVIDIA FLARE addresses this workflow through server-scheduled rounds, client-local execution, and an explicit update contract, while supporting large-object externalization, pull-based tensor streaming, and disk-backed aggregation for large payloads. The post presents FedUMM as an adapter-based design in which clients train LoRA adapters over a frozen BLIP3o backbone and send only adapter updates, while full-model training remains possible when required. In reported eight-client experiments, adapter-only federation reduced per-client communication from 28.6 GB to 0.094 GB per round, improved VQA v2 by 0.7 points, and retained about 97% of centralized performance on VQA v2 and GenEval; the evaluation used simulated sites and does not establish clinical performance or formal privacy guarantees.


### [How we knew COVID was over (and what our models had to unlearn)](https://yomu.fyi/post/how-we-knew-covid-was-over-and-what-our-models-had-to-unlearn.md)
- Company: [Airbnb](https://yomu.fyi/company/airbnb.md)
- Author: Harrison Katz
- Published: Aug 19, 2026

Airbnb's forecasting team outlines their framework for managing production models that experience drift or external shocks. Rather than treating model updates as a generic retraining task, they distinguish between refitting with new data, respecifying the underlying model structure, and intentionally holding without changes. This discipline prevents teams from chasing temporary noise, carrying obsolete crisis assumptions, or over-engineering solutions during disruptions.


### [Post-Train NVIDIA Cosmos 3 Edge for On-Device Robot Control](https://yomu.fyi/post/post-train-nvidia-cosmos-3-edge-for-on-device-robot-control.md)
- Company: [NVIDIA](https://yomu.fyi/company/nvidia-developer-blog.md)
- Author: Michelle Horton
- Published: Aug 19, 2026
- Type: Tutorial

Deploying large world models on robots is constrained by onboard memory and the control latency needed for continuous manipulation. The tutorial post-trains NVIDIA Cosmos 3 Edge, a 4B omni-model with a 2B NVIDIA Nemotron-based reasoner, into a DROID action policy using 76,000 successful teleoperated trajectories, then serves it through an OpenPI-compatible WebSocket server on Jetson Thor. The configuration predicts 32 future 8-dimensional joint-position actions at 15 Hz, while a receding-horizon loop executes a prefix and replans from fresh camera and robot-state observations. On an NVIDIA Jetson AGX Thor T5000, inference takes about 1.53 seconds per action chunk covering roughly 2.13 seconds of motion, and the policy reaches 22.9% success in closed-loop RoboLab evaluation; the source contrasts this with Cosmos 3 Nano’s 36.8%.


### [Evaluating AI Agent Skill Performance with NVIDIA SkillEvaluator](https://yomu.fyi/post/evaluating-ai-agent-skill-performance-with-nvidia-skillevaluator.md)
- Company: [NVIDIA](https://yomu.fyi/company/nvidia-developer-blog.md)
- Author: Michelle Horton
- Published: Aug 19, 2026
- Type: Benchmark

NVIDIA SkillEvaluator measures whether packaged agent skills improve agent performance by comparing task runs with and without a skill installed. It combines three evaluation tiers: static safety and structure checks, embedding-based distinctiveness analysis, and live evaluations in isolated Harbor sandboxes. In the August 12, 2026 benchmark snapshot, more than 300 verified skills spanning over 30 NVIDIA products were tested across Claude Code and Codex, with macro-averaged Skill Lift of +31 points across all dimensions and +39 excluding Security. Correctness rose from 46 to 87 and Effectiveness from 39 to 78, while Security moved from 97 to 98; token and runtime effects varied by skill, with one reducing both and another increasing them.


### [Gisting: Compressing LLM Agent context to ↑ throughput and ↓ cost](https://yomu.fyi/post/gisting-compressing-llm-agent-context-to-throughput-and-cost.md)
- Company: [Shopify](https://yomu.fyi/company/shopify.md)
- Author: 2023-10-18
- Published: Aug 19, 2026
- Type: Explainer

Long system prompts improve agent behavior but increase inference time, cost, and the GPU capacity needed to serve a given traffic level. Shopify compresses the Sidekick GraphQL agent’s roughly 6,000-token system prompt into about 1,500 gist tokens, using special vocabulary tokens whose embeddings are learned by knowledge distillation while the model weights remain frozen. Training compares teacher logits from the full prompt with student logits from the gist-token prompt using KL divergence; at inference, the prompt is replaced with the gist-token string, requiring no custom serving path. In load tests at 350 requests per minute, compression reduced median time to first token from 438ms to 354ms and end-to-end latency from 6.8s to 4.2s, while raising throughput from 20.2 to 23.4 QPS and enabling 14% fewer GPUs in production without losing prediction quality.


### [Examining Human-Like Behaviors in LLMs: A Multi-Dimensional Analysis of Model Behaviors, User Factors, and System Prompts](https://yomu.fyi/post/examining-human-like-behaviors-in-llms-a-multi-dimensional-analysis-of.md)
- Company: [Apple](https://yomu.fyi/company/apple.md)
- Published: Aug 19, 2026
- Type: Benchmark

Large language models frequently exhibit human-like behaviors such as expressing emotions, building relationships, and setting boundaries, yet empirical guidance on appropriate usage remains limited. To address this gap, researchers analyzed 21,000 multi-turn conversations across four models: gpt-4o, gpt-4.1-mini, claude-sonnet-4.6, and gemini-2.5-flash. The study combined LLM-as-a-judge techniques with human evaluation to measure behavior prevalence, perceived appropriateness, and prompt controllability across diverse user goals and profiles. Human evaluators judged LLM self-referential and relationship-building behaviors as less appropriate than identical human actions, whereas boundary-maintaining behaviors were rated as more appropriate for models. System prompting effectively steered these behaviors, though the authors emphasize that rigorous evaluation remains necessary to prevent unintended side effects.


### [Databricks Document Intelligence: pushing the frontier for complex document extraction](https://yomu.fyi/post/databricks-document-intelligence-pushing-the-frontier-for-complex-docu.md)
- Company: [Databricks](https://yomu.fyi/company/databricks.md)
- Author: Jane Zhang, Arnav Singhvi, Ivan Zhou, Archika Dogra, Matthew Ding, Nihit Desai
- Published: Aug 18, 2026
- Type: Announcement

Databricks introduces Precision Mode for its ai\_extract document extraction API, targeting enterprise workloads that challenge existing LLM and rules-based systems. The mode addresses long documents with cross-references, large nested outputs such as thousands of line items, and schemas requiring reasoning across financial statements or applying discounts. It combines custom-trained, efficient extraction models with an agent harness that decomposes jobs semantically, runs smaller tasks in parallel, preserves intermediate results, and reconciles them into a final structured output. On benchmarks covering roughly 9,000 documents, including files up to 2,000 pages and schemas with more than 300 deeply nested fields, Precision Mode reached 94.7% accuracy, seven points above the strongest frontier-model chunk-and-merge baseline, GPT-5.6 Sol; the source says it is now available through ai\_extract and the Information Extraction UI.


### [How Much Memory Does Your Agent Actually Need?](https://yomu.fyi/post/how-much-memory-does-your-agent-actually-need.md)
- Company: [Hugging Face](https://yomu.fyi/company/hugging-face.md)
- Author: Vatche Isahagian, Gaodan Fang, Jayaram Radhakrishnan, Punleuk Oum, Ashwath Vaithinathan Aravindan, Evelyn Duesterwald, G Thomas, Vinod Muthusamy, Merve Unuvar, Ayhan Sebin
- Published: Aug 18, 2026
- Type: Benchmark

The post examines how much agentic memory different models can use effectively, evaluating ALTK-Evolve's self-distilled guidelines across eight models on AppWorld's 585 multi-step tasks. Memory is produced by extracting and consolidating lessons from successful and unsuccessful trajectories, then delivering either the full guideline set on every ReAct step or a fixed core plus task-relevant retrieval, without updating model weights. Results identify three patterns: strong models with headroom benefited from the full set, gpt-oss-120b gained 16.1 percentage points in task completion from curated retrieval, and GLM-5 showed no measurable gain; stricter scenario completion sometimes improved more. Curated retrieval gave gpt-oss-120b that gain with only 5% more tokens versus 51% for its full set, while prompt caching is presented as a way to reduce recurring input cost, supporting calibration to model capability rather than simply accumulating memory.


### [How AI Coding Agents Can Unlock Materials Simulation with NVIDIA ALCHEMI Toolkit](https://yomu.fyi/post/how-ai-coding-agents-can-unlock-materials-simulation-with-nvidia-alche.md)
- Company: [NVIDIA](https://yomu.fyi/company/nvidia-developer-blog.md)
- Author: Elizabeth Goodman
- Published: Aug 18, 2026
- Type: Benchmark

Atomistic simulation combines scientific judgment, compute-efficient implementation, and an accessible interface to the simulation stack, but the MLIP ecosystem lacks the mature interfaces available for classical force fields. NVIDIA ALCHEMI Toolkit addresses implementation with composable, PyTorch-native, GPU-accelerated building blocks and in-flight batching, while its agent skills and reference files supply API patterns to coding agents. The post benchmarks 45 generated pipelines across silicon equation of state, copper adsorption, and lithium self-diffusion, checking code features and running representative workloads on NVIDIA H200 GPUs. All 45 scripts used batched GPU execution, 38 completed screening, and all 15 production representatives completed; however, agents did not assess physical validity, and thermostat choices could suppress measured lithium diffusion by 3–5×, so independent validation against DFT, experiment, or another reference remains necessary.


### [Run Massive-Scale UMAP in Minutes Using Multiple GPUs—Without Losing Accuracy](https://yomu.fyi/post/run-massive-scale-umap-in-minutes-using-multiple-gpus-without-losing-a.md)
- Company: [NVIDIA](https://yomu.fyi/company/nvidia-developer-blog.md)
- Author: Tanya Lenz
- Published: Aug 18, 2026
- Type: Explainer

UMAP workflows on tens to hundreds of millions of vectors are constrained by the cost of constructing an all-neighbors k-nearest-neighbor graph, especially when repeated runs support exploration and parameter tuning. The NVIDIA cuML and cuVS 25.06 feature distributes this graph-construction stage across multiple GPUs while retaining the out-of-core design that partitions data into balanced clusters, overlaps vectors across nearby clusters, computes local graphs, and merges them into a global graph. knn\_n\_clusters and knn\_overlap\_factor expose the memory, runtime, and graph-quality trade-off: more clusters reduce per-GPU data, while more overlap can improve nearest-neighbor recall at higher computational and memory cost. On eight NVIDIA H100 GPUs, the reported evaluation found up to 74x end-to-end speedup over a projected CPU runtime for the 106M-vector MIRACL dataset, completing it in 8 minutes. The post states that performance continues to improve from one to eight GPUs while preserving comparable embedding quality across configurations.


### [Evaluating AI Agents Live at the Grounded Reasoning Cup](https://yomu.fyi/post/evaluating-ai-agents-live-at-the-grounded-reasoning-cup.md)
- Company: [Databricks](https://yomu.fyi/company/databricks.md)
- Author: Databricks AI Research Team
- Published: Aug 18, 2026
- Type: Benchmark

The Grounded Reasoning Cup evaluated whether AI agents could generalize grounded reasoning improvements from OfficeQA to a newly released, enterprise-style document benchmark. Eleven academic teams had about two months to build agents with a partner lab’s model family, then applied them live to OfficeQA Pro V2, based on U.S. Treasury Accounts of Receipts and Expenditures, after a 36-hour release window. Stanford won with 63.3% accuracy, while the top systems combined document preprocessing, targeted or hybrid retrieval, specialized tools, parallelism, and answer verification; 18.8% of questions remained unsolved by every team. The results show that benchmark gains did not reliably transfer, and that parsing, retrieval, tool use, verification, and operational infrastructure can matter as much as model choice in end-to-end enterprise reasoning.


### [Multi-Vector (Late Interaction) Embedding Models with Sentence Transformers](https://yomu.fyi/post/multi-vector-late-interaction-embedding-models-with-sentence-transform.md)
- Company: [Hugging Face](https://yomu.fyi/company/hugging-face.md)
- Author: Tom Aarsen, Antoine Chaffin, Raphael Sourty
- Published: Aug 18, 2026
- Type: Tutorial

Sentence Transformers v6.0 adds MultiVectorEncoder, bringing ColBERT-style late-interaction retrieval into the library’s existing API for dense, sparse, and reranker models. The encoder retains one projected, typically 128-dimensional vector per token instead of pooling a document into one vector, allowing documents to be encoded independently and indexed offline. At scoring time, MaxSim finds each query token’s highest dot-product match among document tokens and sums those maxima; contextual embeddings support both paraphrase matching and exact-token evidence. The guide covers checkpoint loading, encoding, scoring, semantic search, retrieval and reranking, indexing, token pooling, inference speed, evaluation, and retrieval from page images, audio, and video. It also quantifies the storage trade-off: 4,874 passages produced 608,414 token vectors and a 311.5 MB float32 index, while fast-plaid compression reduced that example to 92 MB.


### [How Databricks Feature Store serves features with sub-second freshness](https://yomu.fyi/post/how-databricks-feature-store-serves-features-with-sub-second-freshness.md)
- Company: [Databricks](https://yomu.fyi/company/databricks.md)
- Author: Ian Ackerman, Nick Joung, Abhay Bothra
- Published: Aug 17, 2026
- Type: Explainer

Databricks Feature Store addresses the lag between batch-computed machine-learning features and decisions that depend on events from seconds earlier. It lets a feature definition drive both offline batch processing and online streaming pipelines, orchestrating Kafka ingestion, Spark Real-Time Mode (RTM), Lakebase, and Model Serving. For a rolling 10-minute transaction sum, RTM updates per entity state in local RocksDB, expires contributions per event, and writes revised values to Lakebase through a streaming JDBC sink. The described path reports end-to-end p99 200ms latency from Kafka arrival to online feature availability, while Model Serving retrieves declared dependencies automatically at inference. Tumbling and sliding windows remain options when fewer, cheaper updates are acceptable.


[Newer posts](https://yomu.fyi/topic/machine-learning/page/4.md) · [Older posts](https://yomu.fyi/topic/machine-learning/page/6.md)