AI Daily — May 11, 2026
2026-05-11
TODAY'S NEWS
arxiv.org
More Thinking, More Bias: Length-Driven Position Bias in Reasoning Models
A study across 13 reasoning-mode configurations finds that longer chain-of-thought trajectories correlate positively with increased position bias in multiple-choice QA, with partial correlations ranging from 0.11 to 0.41 (all p < 0.05) after controlling for accuracy. This contradicts the common assumption that extended reasoning reduces shallow heuristic biases. The finding holds across DeepSeek-R1 distilled 7–8B models, base models with CoT prompting, and the full 671B DeepSeek-R1, tested on MMLU, ARC-Challenge, and GPQA.
arxiv.org
RateQuant: Optimal Mixed-Precision KV Cache Quantization via Rate-Distortion Theory
RateQuant applies rate-distortion theory to mixed-precision KV cache quantization, addressing the problem that existing methods assign uniform bit-widths across attention heads despite large variance in head importance. The paper identifies a critical pitfall: each quantizer follows a distinct distortion curve D(b)=alpha*beta^{-b} with beta varying from 3.6 to 5.3, meaning cross-quantizer distortion models invert optimal allocation. The proposed framework derives theoretically optimal per-head bit allocation, directly targeting the primary memory bottleneck in LLM serving.
LAST WEEK'S TOP STORIES
GPT-5.5 Instant: smarter, clearer, and more personalized
OpenAI replaced ChatGPT's default model with GPT-5.5 Instant across all tiers, claiming meaningful gains in accuracy and reduced hallucinations alongside new long-term personalization capabilities.
Testing ads in ChatGPT
OpenAI has begun showing ads to free-tier ChatGPT users, marking a fundamental shift in its monetization strategy that could reshape the product's long-term incentives and independence.
AlphaEvolve: How our Gemini-powered coding agent is scaling impact across fields
Google DeepMind reports that AlphaEvolve is now delivering measurable real-world impact across scientific research, engineering, and business optimization, representing one of the most concrete demonstrations of autonomous algorithm discovery at production scale.
Delay, Plateau, or Collapse: Evaluating the Impact of Systematic Verification Error on RLVR
This paper shows that systematic verifier errors in RLVR pipelines—not just random noise—can cause models to learn consistently wrong behaviors, directly challenging the reliability of code and math reasoning training workflows that depend on automated reward signals.
Advancing voice intelligence with new models in the API
OpenAI released new Realtime API voice models with integrated reasoning, speech translation, and transcription in a low-latency pipeline, significantly expanding the capabilities available to developers building production voice applications.