Engineering
Per-tensor layout maps for GGUF quantization
A Hugging Face contributor describes a data-driven quantization scheme for GGUF models that empirically measures per-tensor sensitivity rather than relying…
Engineering
From Parameters to Answers: How LLMs Retrieve and Use Their Internal Knowledge
The paper studies how large language models retrieve and use internal knowledge by performing layerwise interventions on hidden states across…
Engineering
When Agents Disagree: Bayesian Backward Reasoning as a Label-Free Anchor for Multi-Agent Collective Decision-Making
The paper proposes a Bayesian backward-reasoning method for aggregating decisions from multiple LLM agents, constructing reverse posteriors via explicit likelihood…
Engineering
Beyond Confidence: Stability-Aware Test-Time Adaptation for LLM Reasoning
The paper introduces TASCO, a test-time adaptation framework that improves LLM reasoning by optimizing lightweight task-level prefixes while keeping the…
Engineering
A Unified Per-Token Gating Family for On-Policy Distillation: FKL/RKL Mixing with Multi-Channel and Bias Coefficients
The paper proposes a unified four-coefficient parameterization for on-policy knowledge distillation that generalizes existing per-token gating methods by combining forward…
Engineering
Memory Compression for High-Fanout Agent Sandboxes
Researchers introduce AgentZip, a memory compression system for AI-agent sandboxes that exploits redundancy across sandbox instances and between sandboxes and…
Engineering
COBRA-Skills: Contextual Bandit-Guided Evolution for Agent Skill Optimization
COBRA-Skills is a framework that optimizes reusable skills for LLM agents by framing skill development as budgeted sequential optimization over…
Engineering
Train Smarter, Not Harder: Switching Signal-Guided Training in Active Learning
The paper studies when to retrain an active-learning model from scratch versus fine-tune it, finding retraining wins early and fine-tuning…
Engineering
Co-Evolving Harnesses and Models: On-Policy Correction Helps Weaker Models Catch Up Where Imitation Fails
The paper studies agent harnesses (system prompt, tools, execution hooks, context management) together with model fine-tuning, finding that naively imitating…
Engineering
Diffs vs. Whole Files: An Empirical Comparison of Iterative Edit-Based and Direct Generation for Flutter/Dart Code Models
The authors train two small code models (a 100M-parameter Rainbow-Pony-100M and a fine-tuned Qwen2.5-Coder-0.5B) on Flutter/Dart code in both direct…
Engineering
A Three-Layer Caching Architecture for Low-Latency LLM Web Search on Commodity CPU Hardware
The team describes OreoLook, an open-source AI answer-engine (browser-agent-driven, provider-routed LLM inference) whose local caching stack runs entirely on commodity…
Engineering
EvoSafeHarness: Evolving Model- and Domain-Specific Harnesses for Securing Agents
Researchers introduce EvoSafeHarness, a framework that automatically generates customized safety policies for LLM agents by jointly optimizing a natural-language policy…