LLMs

Research characterizes the Clinical Lost-in-the-Middle (CLitM) effect, where LLMs struggle to retrieve critical information located in the center of long electronic health records. Using MedAlign on 2,196 pairs across six models, the study found a 21.9 percentage-point accuracy gap between peak performance in the outer deciles and the trough in the 70-80% range. This central retrieval failure poses a significant risk for clinical reasoning tasks where key facts often reside in the middle of lengthy patient notes. The authors compare various context-selection strategies to mitigate this degradation in reliability.

• LLMs lose ~22% accuracy on facts in the center of long EHRs vs edges.
• CLitM is a distinct, measurable risk for clinical decision support systems.
• Context-selection strategies are being evaluated to fix central retrieval gaps.
• Validation on MedAlign suggests this affects multiple model architectures.

This paper introduces GRAFT, a method to adapt tree-based speculative decoding for diffusion language models (DLMs) like DFlash, which generate future-position distributions in a single pass. Unlike traditional builders that condition on parent tokens, GRAFT selects candidate nodes from high-probability tokens at each future position and optimizes edge selection using target-distilled scoring. This approach resolves the incompatibility between DLMs' parallel output structure and standard parent-conditioned draft tree construction.

• Enables speculative decoding for DLMs by decoupling draft construction from parent-conditioned expansion.
• Uses target-distilled edge scoring to improve draft quality beyond simple token probability.
• Constructs draft trees by selecting candidates from single-pass future-position distributions.
• Increases mean accepted tokens by optimizing paths across multiple future positions simultaneously.

Research shows that LLM safety mechanisms often fail to address harmful knowledge embedded during pretraining, leaving models vulnerable to semantic camouflage attacks. These adversarial inputs wrap malicious intent in benign narratives, effectively bypassing standard input and output filters. By analyzing latent activation trajectories in small language models, the study identifies a critical 'Intent Horizon' depth where harmful intent becomes detectable before generation completes.

• Surface-level refusal mechanisms do not erase pretraining knowledge of harmful concepts.
• Adversarial semantic camouflage can bypass guardrails by embedding harm in benign contexts.
• Latent activation analysis reveals a universal 'Intent Horizon' for early detection.
• Models remain vulnerable despite final-stage safety alignment protocols.

Large language models often struggle with latency when generating long reasoning traces for complex tasks. This paper introduces Self-Speculation for Reasoning Models (SSR), a training-free decoding method that uses partial chain-of-thought outputs as draft tokens. By leveraging the structure of reasoning workflows rather than just token-level generation, SSR accelerates inference for latency-sensitive applications like voice assistants and coding agents.

• SSR is a training-free decoding method that reduces inference latency for reasoning models.
• It uses partial chain-of-thought sequences as draft tokens to speed up verification.
• Targeted at latency-sensitive interactive apps like voice assistants and coding agents.
• Leverages reasoning workflow structure, improving on standard token-level acceleration.
• Addresses the bottleneck of long reasoning traces in complex multi-step decision making.

A new study compares zero-shot LLMs against fine-tuned classifiers for intent detection, finding performance hinges on the intent schema. On compact datasets like ATIS, fine-tuned RoBERTa significantly outperforms Claude Haiku while being vastly cheaper and faster. However, on broad schemas like CLINC150, the two approaches achieve statistically tied results.

• Zero-shot LLMs do not universally replace fine-tuned NLU; context matters.
• Fine-tuned RoBERTa beats LLMs by 11.8 points on compact ATIS data.
• LLMs and fine-tuned models tie on broad 150-intent CLINC150 schemas.
• Fine-tuning remains superior for cost and latency on specific domains.

Amazon Bedrock has expanded its built-in Web Search tool to include an external_web_access parameter, enabling models to fetch content directly from the public web. This update allows for grounding responses in the freshest information, such as live sports scores, while maintaining data security within the AWS environment. Users can enable this feature by granting the bedrock-websearch:ExternalWebAccess IAM permission.

• Enable external public web retrieval by setting external_web_access to true.
• Grant bedrock-websearch:ExternalWebAccess IAM permission to the request identity.
• Access live public web data for use cases requiring real-time information.
• Maintains zero data egress from your secured AWS environment.
• Extends existing Bedrock Web Search capabilities to broader public sources.

OpenAI has lowered API prices for GPT-5.6 Sol, with input tokens now at $4 per million and output tokens at $20 per million. These rates represent a 20% reduction in input costs and a 33.3% drop in output costs compared to previous tiers. The promotional pricing is valid through November 21, 2026, aiming to support high-volume agentic and research workloads on AWS Bedrock.

• Input pricing drops to $4/1M tokens, a 20% decrease from prior rates.
• Output pricing falls to $20/1M tokens, reflecting a 33.3% reduction.
• Promotional rates are locked in through November 21, 2026.
• Optimized for sustained, high-volume agentic coding and analysis tasks.
• Check AWS Regions page for current GPT-5.6 Sol availability.