LLMs

This study evaluates baking documents into 4-bit Gemma-4-e4b weights via LoRA to enable closed-book question answering without retrieval. Across 100 training runs, the authors found that adapter capacity acts as a hard gate, but once sufficient, training data quality becomes the dominant factor for accuracy. A single curation pass focused on canonical answer spans improved accuracy by nearly 30 percentage points, outperforming architectural changes or hyperparameter tuning.

• Data quality outweighs LoRA rank, learning rate, and architecture changes in closed-book QA.
• Adapter capacity is a hard gate; insufficient size renders data interventions ineffective.
• Curation (shortening answers, dropping trivia) yields larger gains than architectural tweaks.
• Closed-book QA is viable on 4-bit models if internalized data is highly curated.

This paper introduces a framework for accurately predicting LLM inference latency on heterogeneous edge hardware. It models requests by combining static descriptors with dynamic telemetry, separating prefill and decode phases. A gated prediction model adaptively fuses these inputs to account for factors like DVFS and thermal variations. The approach facilitates faster, more reliable model selection during deployment.

• Separates prefill and decode phases for granular latency modeling.
• Fuses static config with dynamic telemetry via a gated model.
• Accounts for DVFS and thermal variations in predictions.
• Validated on Pixel mobile devices for edge deployment scenarios.
• Enables faster LLM screening based on accurate runtime estimates.

Traditional benchmarks fail to capture nuance when multiple answers are acceptable, relying instead on static ground truths. This research introduces a framework where a panel of diverse LLMs ranks anonymized responses to gauge relative preference. Aggregate inter-model agreement serves as a proxy for perceived quality, focusing on clarity and usefulness rather than binary correctness.

• Moves beyond static datasets to measure relative preference among model outputs.
• Uses a panel of diverse LLMs to rank anonymous responses for the same prompt.
• Inter-model agreement acts as a proxy for response quality and usefulness.
• Better captures nuances like clarity and completeness than absolute correctness metrics.

AWS has launched Claude Opus 5 via Amazon Bedrock, marking the latest iteration of the Opus series with enhanced capabilities for coding and autonomous agents. The model supports zero data retention and demonstrates improved accuracy in navigating large codebases and analyzing complex, document-heavy enterprise tasks. It is designed to sustain long-running agent sessions, recovering from errors to complete multi-hour objectives.

• Accessible via Amazon Bedrock with zero data retention support for secure enterprise use.
• Excels at understanding large codebases and generating production-quality code autonomously.
• Capable of running dependable agents for hours, recovering from errors to meet goals.
• Delivers significant reasoning improvements for complex analysis of long documents.