#machine-learning
Every summary, chronological. Filter by category, tag, or source from the rail.
Advances in Data Center Inference Engineering
Inference engineering is shifting from post-training optimization to a cycle where dedicated training processes—specifically in quantization, KV compaction, and speculative decoding—are essential for production performance.
AI EngineerMoving Beyond Academic Benchmarks: The Shift to Task-Based AI Evaluation
Vals is replacing static, public AI benchmarks with private, task-specific evaluations that measure real-world performance in high-stakes industries like law, finance, and cybersecurity.
Self-Improvement via Fast Tree-Search
The paper presents a methodology for enhancing AI model performance by integrating fast tree-search algorithms, enabling models to iteratively improve their reasoning and output quality through structured exploration.
Architecting Long-Horizon AI Agents via Cascaded Intelligence
The paper proposes a hierarchical architecture for long-horizon AI agents that decouples high-level strategic planning from low-level execution using 'levels' and 'ticks' to manage complex, multi-step tasks.
Risks of Agent-Mediated Hiring: Access and Recurrence Bias
Multi-agent résumé screening systems can inadvertently amplify hiring biases, creating 'recurrence' where specific candidate profiles are consistently excluded due to agent-to-agent feedback loops.
Mapping the Design of LLM Benchmarks
Current LLM benchmarks often lack transparency in design, leading to misaligned evaluations. This research provides a taxonomy to categorize benchmark construction, helping developers better understand what models are actually being tested for.
Detecting LLM Harm via Latent States
Rather than relying on output filtering, this research proposes monitoring internal latent states of LLMs to detect harmful intent before it manifests in generated text.
The Strategic Silence of World Model Startups
World model companies are intentionally obscuring their product roadmaps to avoid early competition, leveraging current funding abundance to remain in a 'research-only' phase.
The Inference Engineering Pareto Atlas: Optimizing LLM Performance
The paper provides a systematic framework for navigating the trade-offs between cost, quality, and latency in LLM inference, identifying which optimization techniques dominate the performance frontier.
OBC-Prune: Outcome-Based Calibration for Reasoning Model Pruning
OBC-Prune introduces a calibration method for pruning large reasoning models that prioritizes final output accuracy over intermediate token probability, preserving complex reasoning capabilities during compression.
NeMo Data Designer: Framework for Multimodal Synthetic Data
NeMo Data Designer provides an extensible, modular framework for generating high-quality synthetic data across multiple modalities, addressing the critical bottleneck of data scarcity in training large-scale AI models.
Detecting Sensor Attacks in Urban Flows with Physics-Constrained AI
This research introduces a framework for securing urban pedestrian flow data by combining physics-based digital twins with conformal prediction to detect stealthy false data injection attacks.
Evidence Masking as a Driver for Compositional Generalization
Research confirms that masking specific evidence during training forces models to learn compositional rules rather than memorizing patterns, significantly improving generalization.
PrismML's Ternary Compression for On-Device LLMs
PrismML is shrinking high-performance LLMs to fit on consumer hardware by using 'ternary' weight compression, achieving 98% benchmark parity with original models.
Treble: Scaling Voice AI via Physics-Based Simulation
Treble has raised $18 million to scale its acoustic simulation platform, which replaces internet-scraped audio data with physics-based synthetic data for training and testing voice AI, wearables, and robotics.
Scalable Verification for Nonlinear Neural Feedback Systems
The paper introduces a branch-and-bound verification framework for nonlinear neural feedback systems, enabling formal safety guarantees by systematically partitioning the state space to handle nonlinear dynamics.
CLEAR: Cross-Source Evidence Adjudication for Medical LLMs
The CLEAR framework improves medical LLM reliability by systematically adjudicating conflicting evidence across multiple sources, reducing hallucinations and improving clinical accuracy.
Safe Error Correction for LLMs via Frozen-Base Adjustment
The paper introduces a method to correct specific model errors by adjusting parameters while keeping the base model frozen, ensuring targeted fixes without degrading general capabilities.
Optimizing Disaggregated LLM Serving via Calibrated Routing
Efficient LLM serving in disaggregated clusters requires a two-stage approach: calibrating model latency predictions to account for system noise, followed by intelligent request routing to minimize total cluster latency.
Optimizing KV Cache Placement for Long-Lived LLM Sessions
To handle long-lived LLM sessions, move the KV cache beyond GPU VRAM into CPU memory and SSDs using tiered placement policies to balance latency and memory capacity.
The Data Bottleneck in Physical AI and Robotics
Robotics lacks a 'ChatGPT moment' because it lacks a massive, internet-scale dataset for physical interaction, forcing the industry to rely on synthetic data and simulation.
Building Whistleblowing Infrastructure for AI Agents
New reporting tools allow AI agents to flag misbehaving peers, but experts warn that fostering collaboration through positive models is more effective than building an automated surveillance state.
The Andrej Karpathy Blog: A Decade of AI Engineering
Andrej Karpathy's blog serves as a foundational archive of practical AI engineering, emphasizing 'from-scratch' implementations, deep learning fundamentals, and the importance of hands-on experimentation.
KuaiRP: Technical Report on Role-Playing Model Optimization
The KuaiRP technical report details specialized training methodologies for enhancing LLM performance in role-playing scenarios, focusing on character consistency and narrative depth.
Grounding Agent Memory via Environment-Probing Curation
Enterprise AI agents often fail due to stale or irrelevant memory. This paper introduces 'Environment-Probing Curation,' a method that actively validates and filters memory stores against real-time environment states to ensure high-fidelity decision-making.
The Agent Incident Registry: A Framework for Preventing AI Failures
The Agent Incident Registry (AIR) proposes a standardized, community-driven database to catalog and analyze AI agent failures, enabling developers to learn from past errors and prevent recurring systemic vulnerabilities.
Quantifying the Memorization-to-Generalization Transition in Grokking
The paper provides a quantitative framework for understanding 'grokking'—the phenomenon where neural networks suddenly shift from memorizing training data to generalizing—by identifying specific scaling laws and phase transitions in model learning.
Training Nemotron for Olympiad-Level Mathematics
The paper outlines a systematic recipe for training LLMs to achieve gold-medal performance in Olympiad-level mathematics, emphasizing high-quality synthetic data generation and iterative reinforcement learning.
Automating QUBO Formulation from Natural Language
This paper introduces a method to bridge the gap between human-readable optimization problem descriptions and the mathematical rigor of Quadratic Unconstrained Binary Optimization (QUBO) using LLMs.
Task-Agnostic Environment Preprocessing for AI Agents
The paper introduces a method for AI agents to learn from environments without predefined task syllabi, focusing on task-agnostic preprocessing to improve generalization and performance.
Showing 30 of 610