LLMs

Amazon Bedrock has made Z.ai's GLM 5.3 generally available, offering a mixture-of-experts model with 753B total parameters and 40B active per token. The model features a 1-million-token context window, up to 128K output tokens, and selectable reasoning effort levels to balance latency and performance. It is optimized for agentic coding and long-horizon software engineering tasks while leveraging AWS security and prompt caching capabilities.

• GLM 5.3 is a MoE model (753B total/40B active) optimized for agentic coding on Bedrock.
• Supports 1M input and 128K output tokens with selectable reasoning effort levels.
• Explicit prompt caching is available to reduce latency and cost for repeated system prompts.
• Leverages AWS security posture while providing access to Z.ai's latest open-weight flagship.

Periscope addresses the accuracy drop in frozen language models when processing long texts by factorizing the reading process into a finite decision task. It arranges N text chunks on a KxK grid and queries the model for local consecutive spans and global strided spans. By combining log-odds from these two perspectives, it generates an evidence map for each chunk without any additional training.

• Enables long-context reasoning with frozen models via training-free inference.
• Uses a KxK grid to split queries into local and strided span assessments.
• Combines local and global scores to build a per-chunk evidence map.
• Avoids quadratic cost penalties by factoring the read into finite choices.

AWS has made Amazon Nova 2.5 Sonic generally available as a speech-to-speech model designed for natural, real-time voice interactions. This update enhances reasoning, instruction following, and tool-calling accuracy while reducing latency. The model enables agents to maintain context and execute multi-step workflows like order lookups and returns within a single conversation.

• Nova 2.5 Sonic improves tool-calling accuracy for real-time voice applications
• Lower latency enables faster, more natural conversational interactions
• Enhanced reasoning supports complex multi-step instruction following
• Useful for customer service, personal assistants, and real-time chat bots

SYNLAT reduces token costs in long chain-of-thought traces by aligning compression boundaries with syntactic structure using non-overlapping Syntax-Aligned Units. Instead of fixed-length or token-level cuts that fragment phrases, this framework uses an answer-conditioned teacher to guide a student model in generating mixed reasoning. The student produces outputs based only on the question and requested compression level, preserving critical derivations while meeting budget constraints.

• Aligns compression boundaries with syntax to prevent fragmentation of formulas and phrases.
• Uses non-overlapping Syntax-Aligned Units (SAUs) instead of arbitrary token cuts.
• Teacher-student distillation creates mixed reasoning from question and compression targets.
• Preserves answer-critical information under strict output-token budgets.

Researchers identify 'absorption,' a phenomenon where LLMs treat trailing user comments as part of previously pasted text during editing tasks. Unlike existing benchmarks that explicitly separate instructions from data, this study focuses on unmarked seams where users paste content and add remarks. The authors introduce SEAM, a benchmark of 300 examples designed to measure how often models incorrectly merge these distinct text sources.

• LLMs fail to distinguish between pasted artifacts and immediate user commentary without explicit markers.
• Existing benchmarks miss this edge case by relying on marked instruction-data separation.
• SEAM benchmark provides 300 controlled examples to quantify absorption errors in editing tasks.
• Model outputs may inadvertently include user notes within edited pasted content.

Current uncertainty quantification methods for LLMs struggle with stability when inputs are semantically equivalent paraphrases. This paper introduces a framework that uses a lightweight proxy model on LLM hidden states to aggregate predictions across variants. By constructing label-wise nonconformity scores, the method maintains marginal coverage guarantees under score exchangeability while reducing variability caused by rewording.

• Fixes instability in LLM confidence scores caused by meaning-preserving paraphrases.
• Uses a lightweight proxy on hidden states to aggregate predictions across variants.
• Maintains formal conformal calibration guarantees under score exchangeability.
• Provides robust uncertainty quantification without requiring heavy model retraining.