Depth-Wise Activation Steering for Honest Language Models

French AI startup Mistral has launched the Devstral 2 coding model, which includes a laptop-friendly version optimized for software engineering tasks. This release follows the introduction of the Mistral 3 LLM family, aimed at enhancing local hardware capabilities for developers.

A new benchmark called Know-Show has been introduced to evaluate the spatio-temporal grounded reasoning capabilities of large Video-Language Models (Video-LMs). This benchmark consists of five scenarios that assess how well these models can reason about actions while grounding their inferences in visual and temporal evidence, highlighting significant gaps between current models and human reasoning.

Recent research has demonstrated the potential of Large Language Models (LLMs) in automating the generation of discharge summaries, which are critical documents in patient care. The study evaluated five models, including proprietary systems like GPT-4 and Gemini 1.5 Pro, and found that Gemini, particularly with one-shot prompting, produced summaries most similar to gold standards. This advancement could significantly reduce the workload of healthcare professionals and enhance the accuracy of patient information.

A new online structured pruning technique called Token Filtering has been introduced for large language models (LLMs), allowing pruning decisions to be made during inference without the need for calibration data. This method measures token redundancy through joint key-value similarity, effectively reducing inference costs while maintaining essential information. The approach also includes a variance-aware fusion strategy to ensure important tokens are preserved even with high pruning ratios.

The introduction of Graph-Regularized Sparse Autoencoders (GSAEs) aims to enhance the safety of large language models (LLMs) by addressing their vulnerabilities to adversarial prompts and jailbreak attacks. GSAEs extend traditional sparse autoencoders by incorporating a Laplacian smoothness penalty, allowing for the recovery of distributed safety representations across multiple features rather than isolating them in a single latent dimension.

Recent research indicates that large language models (LLMs) demonstrate biases in evaluation tasks, particularly favoring self-generated content. However, a study exploring retrieval-augmented generation (RAG) frameworks found no significant self-preference effect, suggesting that LLMs can evaluate factual content more impartially than previously thought.

Large language models (LLMs) have revolutionized automated software development, enabling the conversion of natural language into functional code, as highlighted in a comprehensive survey on code intelligence. This evolution is exemplified by tools like Github Copilot and Claude Code, which have significantly improved coding success rates on benchmarks like HumanEval.

Recent research has introduced a controlled evaluation framework to assess the generalization capabilities of large language models (LLMs) like BERT, Qwen2, and LLaMA under various logical perturbations, including rule deletion and contradictory evidence. The findings indicate that these models maintain high accuracy despite structural changes in reasoning tasks.