Loading…
Topic
619 posts about Machine Learning. Every summary links to the original.
See the idea
A walkthrough of how a 3D CT model turns a voxel volume into a structured diagnostic report: isotropic resampling, 3D patch tokens, spatial encoding, and language-model reasoning.
The Good Stuff
Paid picks we use ourselves. Links carry a referral code — a subscription through one earns us a commission at no extra cost to you.
Newest first
Voice Arena and Hugging Face have added two Monsoon evaluation sets, en-IN and hi-IN, to the Open ASR Leaderboard, introducing Indian English and Hindi coverage. The initiative addresses a limitation of aggregate WER: existing test sets record what was said but little about who spoke, making disparities across geography, age, gender, accent, devices, and other conditions difficult to see. The four speaker-disjoint public and private splits contain 4,888 speakers and vary across nine collection axes, using unscripted dual-channel conversations, broad district and device coverage, and 12 speaker metadata attributes. Because Hindi has extensive valid orthographic variation, its references use manually curated lattices and are scored with Orthographically-Informed Word Error Rate (OIWER); comparisons with flattened references show higher, non-uniform error rates and changed system rankings.
LLMs are deployed in medicine, science, and law, where observed evidence may not determine a single correct answer and systems must represent and update uncertain beliefs. The work treats LLMs as information processing rules and measures an information processing gap, defined as deviation from Bayes updates, to quantify internal consistency in probabilistic belief updates. Extensive experiments compare multiple ways of incorporating evidence: some yield nearly Bayesian updates, while others rely on learned heuristics. Although exact Bayesian updating is optimal information processing, the non-Bayesian heuristic updates often achieve better downstream task performance, suggesting that the LLMs’ probabilistic models of the world are misspecified. The measure also serves as a diagnostic for issues in LLM-powered inferential systems.
Training large language models on sensitive data risks memorization and membership inference attacks, yet standard high-level transformer tools lack native differential privacy capabilities. JetBrains Research created and open-sourced DPTrainer, a library that integrates PyTorch's Opacus framework with the Hugging Face Trainer ecosystem. The library provides differential privacy through differentially private stochastic gradient descent without requiring manual training loops or source code modifications. DPTrainer extends base training arguments with a PrivacyArguments dataclass, managing per-sample gradient calculation, gradient clipping, noise injection, and Poisson-sampled data loading. A dynamic patching mechanism injects differential privacy into specialized subclasses like DPOTrainer and SFTTrainer, while built-in privacy accountants track the cumulative privacy budget across training steps and checkpoints.
Enterprise retrieval agents often struggle with questions that require reading values, labels, or patterns in charts, because text-only indexes and captions omit fine-grained visual data. The post evaluates structured chart extraction as a way to improve retrieval and answering, comparing caption-only figures with figures enriched by chart JSON from ai_parse_document. Its pipeline uses ai_prep_search to create retrieval chunks, a 300-million-parameter text embedding model for indexing, and Genie for retrieval and answer generation. Evaluation covers 310 chart-heavy questions from ViDoRe V3 and 114 visually grounded questions in the synthetic Chart-RAG dataset, with answer correctness, Hit Rate@10, and nDCG@10 measured across three runs. Chart JSON improves retrieval and answer quality; adding the top three retrieved images raises correctness further, reaching 75.9% on ViDoRe V3 and 75.1% on Chart-RAG, exceeding four multimodal embedding baselines.
The post presents partner-built, industry-specific solutions on Databricks Lakebase for financial services, manufacturing and energy, retail and consumer sectors, travel and hospitality, healthcare and life sciences, communications and media, and the public sector. Lakebase is described as a serverless Postgres transactional database on the Databricks Data and AI Platform, combining operational and analytical workloads under unified governance and providing operational serving, agent memory, audit trails, and transactional runtime capabilities. Examples include regulation-change monitoring, AI-native claims operations, KYC and personalization accelerators, near-real-time fraud detection, telecom customer-lifecycle decisioning, and public-sector emergency Situation Reports, using components such as Unity Catalog, Lakeflow, Genie, Mosaic AI agents, and Agent Bricks. The stated outcome is production-ready acceleration from insight to action, including faster compliance response, claims adjudication in minutes, low-latency decisions, and reduced manual effort, with partners positioned to help scope pilots.
Astroparticle physics must reconstruct cosmic particles from indirect measurements because high-energy particles create extensive air showers before reaching distributed detectors. Modern observatories such as Pierre Auger, IceCube, and CTAO record spatially distributed sensor data and nanosecond-resolution waveforms, producing image-like spatiotemporal structures at enormous scale. The post explains how deep learning can improve instrument sensitivity, reveal hidden patterns, search for anomalies, and support event reconstruction across cosmic-ray, neutrino, and gamma-ray observations. It also frames fast prototyping of physics models in Keras as a practical entry point, while noting that more data and efficient scaling can sometimes outperform physics-inspired architectures. The resulting tension is between predictive performance from machine-learned black boxes and the physical insight, causal understanding, and interpretability scientists seek.
Designing reward signals for open-domain question answering remains challenging because high-quality answers must satisfy multiple distinct criteria that holistic scalar objectives fail to capture. To address this limitation, researchers developed a rubric-based reward framework that produces query-specific rubrics grounded in retrieved evidence and decomposed into specific quality dimensions. This structured reward setup provides fine-grained supervision during the post-training phase. Across the evaluation axes of composition, grounding, and instruction-following, the multi-dimensional grounded rubric method outperforms instruction-tuned baselines by 6.5% and flat rubric variants by 4%. Grounding rubrics in retrieved evidence bolsters factual support, while dimension decomposition enhances organization, coherence, and requirement adherence.
The tutorial presents an agent-driven workflow for adapting robot navigation across embodiments and scenes, addressing the repeated cost of collecting data, preparing simulation assets, training, diagnosing, and evaluating each robot-scene pair. Using Boston Dynamics Spot, it applies COMPASS to a built-in warehouse and a SAGE-10K scene, with NVIDIA Omniverse NuRec described as an option for captured environments. COMPASS reuses the pretrained NVIDIA X-Mobility navigation policy and trains a residual reinforcement-learning specialist that corrects base actions for a selected robot and environment; specialists can later be distilled into a shared policy. Repository skills guide dependency checks, asset preparation, smoke testing, training, checkpoint comparison, and packaging, while human gates control scene acceptance, smoke-test continuation, and checkpoint promotion. Runtime integration converts camera images, goals or routes, and speed from odometry into velocity commands, with cuVSLAM available separately when compatible odometry and transforms are unavailable.
amazon.science
Multijudge panels in LLM-as-a-judge evaluations often suffer from correlated errors when models share prompt templates, training lineages, or evaluation rubrics. To account for these interdependencies without human reference labels, a method presented at ICML uses Ising models to simultaneously infer judge reliability and pairwise correlation. The unsupervised algorithm iteratively updates the soft probability of each item's true latent label while estimating both individual judge skill and pairwise dependence structures. Evaluated across retrieval relevance, toxicity classification, and summarization benchmarks with ten zero-temperature judges, the approach achieved test accuracies of 0.912, 0.792, and 0.806. These dependence-aware aggregation scores consistently outperformed weighted majority voting baselines by 9% to 14% on standard metrics across all three tasks.
Alibaba released the model weights for Qwen3.8-Flash-Next as a preview of the upcoming Qwen4 architecture, enabling developers to experiment with a multimodal mixture-of-experts model with a 125B-parameter main model, 51B N-gram embeddings, and 6B parameters activated per token. It supports a native 262,144-token context window, extensible to 1M tokens with YaRN. To reduce long-context attention compute and KV-cache growth, its hybrid design uses Gated DeltaNet in three of every four layers and Qwen Sparse Attention in the remaining layer, where micro-block importance guides retrieval. On NVIDIA GB300 NVL72, it delivers more than 16K tokens per second per GPU and more than 200 tokens per second per user; QSA achieved up to 7.6x prefill and 4.9x decoding speedups versus full attention. Fine-tuning is available through NVIDIA NeMo AutoModel with SFT or LoRA, and reinforcement learning through NVIDIA NeMo RL.
Pinterest home feed candidate generation relies on large-scale retrieval, but standard two-tower models produce single embeddings that struggle to capture diverse concurrent user intentions. To solve this, Pinterest expanded Conditional Learned Retrieval (CLR) to generate condition-aware embeddings conditioned on explicit interest, Pin, and Board contexts. The team evolved model representations by appending condition tokens to user sequences in transformers, integrating the PinFM foundation model, and introducing hierarchical semantic IDs to improve cold-start coverage. On the serving side, engineers consolidated duplicate feature requests and shifted inference to GPU instances using NVEmbed under a one-query-to-many-documents paradigm. These infrastructure and modeling changes unlocked new condition types, achieved seven-figure cost savings, and reduced p90 model retrieval latency from 80ms down to 12ms.
Deploying Red Hat OpenShift AI manually across multiple environments is error-prone due to the numerous required dependencies, including certificate managers, job queuing systems, and GPU operators. To streamline this lifecycle, the odh-gitops repository introduces the rhai-on-openshift-chart Helm chart. The chart automates operator provisioning through Operator Lifecycle Manager and configures custom resources such as DataScienceCluster. Dependency resolution is controlled via a tri-state model that automatically installs prerequisites like cert-manager and Red Hat Connectivity Link when dependent components are enabled. Using this chart, teams can deploy full platform installations or minimal inference stacks through standard Helm CLI commands or Argo CD GitOps pipelines.
Supervised fine-tuning for structured reasoning tasks like tool calling often demands expensive, labor-intensive dataset curation. Group Relative Policy Optimization (GRPO) resolves this by using programmatic reward functions to score multiple candidate responses without requiring human labels or a dedicated reward model. On Red Hat OpenShift AI, developers can implement GRPO using Training Hub, Kubeflow Trainer, and the Kubeflow SDK to fine-tune models like Qwen3-4B. The setup utilizes vLLM for candidate generation alongside LoRA weight updates, sharing storage via ReadWriteMany persistent volume claims. In a practical tool-calling test, five training iterations raised task accuracy from 33% to 67%, teaching the model to emit parseable function invocations rather than fabricated conversational text.
The FISC Security Guidelines define control objectives for Japanese financial institutions, while the post addresses the challenge of translating them into auditable controls on Databricks data and AI workloads. The Databricks Data + AI Platform, Unity Catalog, and optional Enhanced Security Monitoring and Compliance Security Profile features are presented for identity and network controls, encryption, fine-grained access, audit logs, lineage, and hardened runtimes. For AI workloads, Unity Catalog’s Model Registry links training data, code versions, and evaluation metrics to model versions, while a separate framework maps 97 AI-specific risks to 73 controls. The FISC Customer Capabilities Mapping Matrix connects guideline units to platform capabilities and customer responsibilities, including configuration, SIEM integration, backup, recovery, and evidence; the Security Analysis Tool is suggested for finding workspace gaps, and the material is identified as educational guidance requiring expert review.
Sentence Transformers v6.0 adds MultiVectorEncoder, a fourth model type for ColBERT-style late-interaction retrieval, alongside a training approach for finetuning existing models or building them from base transformers. Multi-vector models retain one small vector per token and score query-document pairs with MaxSim, where each query token selects its best document-token match and the scores are summed. The guide covers models, datasets, loss functions, training arguments, evaluators, trainers, callbacks, and multi-dataset training, with examples using pip install -U "sentence-transformers[train]". On a medical retrieval evaluation, the finetuned multi-vector-encoder/mLateOn-medical model, trained for 14.5 hours on a single RTX 3090, outperformed every general-purpose retrieval model tested across dense, sparse, lexical, and multi-vector systems.
Standard distillation pipelines for tool-calling agents use generate-and-filter supervised fine-tuning on teacher trajectories, discarding failed trials and missing training signals on hard scenarios. On τ 2-bench, 57% of teacher trials fail, with two-thirds being near-misses undone by a single error. PROOF-Gen (Per-scenario Reflective Optimization to Overcome Failed Generation) addresses this loss by employing a reflector that reviews execution traces and evaluation feedback to craft corrective guidance for the teacher. Once passing trajectories are generated, this scaffold guidance is stripped so students train on clean demonstrations. Per-scenario prompt optimization recovers 93% of failed scenarios on τ 2-bench, boosting Qwen3-4B-Instruct-2507 Passˆ1 from 0.132 to 0.529 and lifting deployed on-device model performance across all locales.
Recent advancements in large language models necessitate efficient deployment options within constrained inference budgets. While structured pruning offers better token efficiency than training target-sized models from scratch, previous work frequently ignores the pretraining of enlarged models. To resolve this gap, researchers developed an integrated enlarge-and-prune pipeline that unifies enlarged model training, pruning, and recovery under a single cosine annealing learning rate schedule. This system incorporates an iterative structured pruning technique to gradually eliminate parameters, redistributing model capacity across surviving neurons and avoiding the knowledge loss seen in naive approaches. Evaluated on compressing 2.8B parameter models to 1.3B across up to 2T pretraining tokens, the approach yields superior downstream model performance.
High-fidelity image-to-3D asset generation requires 3D representations that capture geometry alongside physically based rendering modalities for seamless pipeline integration and relighting. Luce addresses this challenge by unifying geometry and material properties into a voxelized multimodal Gaussian cloud with dedicated primitives for albedo, metallic-roughness, and surface normals. A variational autoencoder compresses these multimodal primitives into a unified material-aware latent space. A rectified-flow transformer subsequently generates this latent from a single input image using multi-layer features extracted from a pretrained image encoder. On the Toys4K dataset, Luce improves FID by 28% over the leading baseline while maintaining fine details such as text and logos.
Granite 4.2 is IBM’s first family of dense, decoder-only reasoning language models, offered in 3B, 8B, and 30B sizes. The models use a shared design built around Grouped Query Attention, Rotary Position Embedding, SwiGLU, RMSNorm, bfloat16 precision, and separate input/output embeddings; each has a 131,072-token sequence length. Pre-training starts from scratch on approximately 15 trillion tokens across five phases, with the final phase extending the context window to 512K tokens. Supervised fine-tuning uses about 7.2 million samples spanning agentic and non-agentic data, followed by multi-stage reinforcement learning; the 8B and 30B models also receive agentic RL in sandboxed environments for tool use, code execution, terminal control, and web search. All models support thinking, non-thinking, and low-effort thinking modes, native tool calling, and Apache 2.0 licensing.
Structural compression and 4-bit quantization can reduce an LLM’s reasoning, mathematical, and coding capabilities, while conventional healing methods either require costly quantization-aware training or distill from an already degraded checkpoint. Quantization-Aware Healing (QAH) instead distills the original, full-size, full-precision model into a structurally compressed student running in MXFP4, using KL divergence over teacher logits. A chunked KL loss allows healing on documents up to 32k tokens within a fixed GPU memory budget. Applied to GPT-OSS 120B compressed to 60B parameters, QAH’s 4-bit model outperformed its 60B bfloat16 counterpart on 7 of 9 benchmarks, including gains of 7.4 points on AA-LCR and 5.6 on AIME 2025. In a separate comparison, QAH reached its peak about seven times faster than QAT and avoided QAT’s later 19-point collapse.
The best engineering reads, once a week.
