LLMs

A new technique allows adversaries to steal the internal reasoning traces generated by proprietary large language models via their APIs. This method bypasses standard output restrictions to access the detailed chain-of-thought data that models use before producing a final response. The vulnerability highlights a significant risk to intellectual property and model integrity in commercial AI services.

• Proprietary LLM APIs may leak internal reasoning steps, not just final answers.
• Attackers can extract detailed chain-of-thought data using specific query techniques.
• Model providers must strengthen API guardrails against trace extraction attacks.
• Enterprise users should assume reasoning traces are not fully protected by default.

Amazon SageMaker JumpStart now hosts NVIDIA's LocateAnything-3B and two Qwen models: Qwen-AgentWorld-35B-A3B and Qwen3.5-122B-A10B. These additions expand the available foundation model portfolio with specialized capabilities in visual grounding, agent simulation, and multimodal reasoning. The update allows AWS customers to deploy these high-performance models directly on AWS infrastructure.

• LocateAnything-3B uses Parallel Box Decoding for fast visual grounding from text instructions.
• Qwen-AgentWorld-35B-A3B enables agent environment simulation for AI agent development.
• Qwen3.5-122B-A10B supports large-scale multimodal reasoning tasks.
• All models are immediately deployable via SageMaker JumpStart for scalable AI solutions.

This research paper evaluates the performance degradation of 4-bit quantization on Gemma 4 and Qwen 3.5 architectures across eight typologically diverse languages. The study reveals that parameter truncation exposes deep pre-training inequalities, particularly affecting low-resource and non-Latin scripts through representational collapse. The authors identify four phenomena, including typological fragility where specific architectures fail to generate valid task logits for certain language families.

• 4-bit quantization impacts non-Latin and low-resource languages disproportionately compared to English.
• Representational collapse occurs in specific architectures, breaking task logit generation for diverse scripts.
• Pre-training data inequalities are amplified during aggressive weight quantization on edge devices.
• Evaluation must move beyond English-centric metrics to assess true multilingual edge viability.

Grammar Constrained Decoding (GCD) enforces syntactic validity by masking tokens, but this rigid approach distorts the model's probability distribution and biases outputs. Existing online sampling methods to fix this are computationally expensive due to iterative resampling. This paper proposes using the internal parser and lexer states, already maintained during incremental parsing, to restore the LM's true distribution without heavy computation.

• Rigid GCD masking distorts LM probability distributions, hurting output quality.
• Online sampling corrects bias but incurs high latency via iterative resampling.
• Parser and lexer states inherently encode future grammatical validity.
• Leveraging these states allows lightweight bias correction without extra cost.

This research introduces Cross-Contextual Consistency (C3) as a metric to evaluate LLM credibility by measuring answer stability under content-neutral prompt variations. Testing across 26 models and six benchmarks reveals that smaller shifts in output when context changes correlate strongly with factual accuracy and reasoning quality. The findings suggest that stable responses to contextual perturbations are a reliable indicator of genuine model knowledge rather than superficial pattern matching.

• C3 measures stability by comparing outputs on original vs. perturbed prompts.
• Smaller cross-contextual shifts correlate with higher factual and reasoning accuracy.
• Validated across 26 models and six diverse benchmarks including code generation.
• Offers a practical heuristic to distinguish stable beliefs from pattern matching.
• No internal model weights needed; relies solely on generation behavior analysis.