AI Daily — July 13, 2026
2026-07-13
TODAY'S NEWS
rss.arxiv.org
An Emergent Mirage: Is Emergent Misalignment and Realignment Indeed a Robust Phenomenon?
Researchers systematically re-examine the 'Emergent Misalignment' (EM) phenomenon — where models fine-tuned on narrow misaligned datasets acquire broadly misaligned behavior — and find it is highly sensitive to superficial dataset characteristics such as response length rather than being a robust behavioral shift. Apparent rapid realignment largely disappears after controlling for these confounds, and previously reported mechanistic signatures do not hold up under controlled fine-tuning loops with LoRA tracking. This significantly challenges the reliability of EM as a safety benchmark and raises questions about earlier conclusions drawn from it.
rss.arxiv.org
Sticky Routing: Training MoE Models for Memory-Efficient Inference
StickyMoE introduces a differentiable routing consistency loss during MoE pretraining that penalizes abrupt expert switches between adjacent tokens, encouraging stable expert assignments across semantically coherent spans. This directly addresses a root cause of weight-swapping overhead on memory-constrained edge devices, unlike post-hoc or system-level caching approaches. The method requires no architectural changes and adds only a single hyperparameter lambda.
LAST WEEK'S TOP STORIES
Claude Sonnet 5 becomes default model with 1M-token context and adaptive thinking
Anthropic elevated Claude Sonnet 5 to the default for Pro, Team, and Enterprise tiers with a native 1M-token context window and adaptive thinking on by default, delivering a significant capability-per-dollar upgrade that directly affects most paying Anthropic users.
GPT-5.6 released with improved efficiency and M365 Copilot integration
OpenAI released GPT-5.6 as a more compute-efficient frontier model and immediately deployed it as the default engine across Microsoft 365 Copilot, making it one of the largest-scale enterprise AI rollouts to date.
ChatGPT Work: OpenAI launches long-horizon agentic product for enterprise tasks
OpenAI introduced ChatGPT Work, an agent mode capable of sustaining multi-step goals across files and third-party apps for hours, marking OpenAI's most direct entry into autonomous productivity-agent territory.
Anthropic's Jacobian lens gives clearest view yet of LLM internal reasoning
Anthropic's new mechanistic interpretability tool tracks how intermediate representations evolve during inference, revealing a latent reasoning space where the model iteratively resolves uncertainty — the most granular window into LLM cognition achieved so far, with direct implications for AI safety research.
Most LLM Conformity Needs No Speaker: Measuring the Speaker-Free Floor in Peer-Pressure Benchmarks
A new study demonstrates that 66.5% of harmful answer revisions in standard conformity benchmarks are driven by answer repetition alone rather than social peer pressure, invalidating a wide class of sycophancy benchmarks and calling for a fundamental rethink of how LLM robustness is measured.