Oct 31, 2025

• Reasoning Models Reason Well, Until They Don't
  Large reasoning models (LRMs) demonstrate impressive performance on reasoning tasks but fail catastrophically when faced with complex problems, revealing a critical limitation in their capabilities.

• Kimi Linear: An Expressive, Efficient Attention Architecture
  Kimi Linear introduces a hybrid linear attention architecture that significantly enhances performance and efficiency, achieving 51.0 on MMLU-Pro and 84.3 on RULER with a 3.98x speedup over traditional methods.

• Sustainable memristors from shiitake mycelium for high-frequency bioelectronics
  Shiitake mycelium serves as a sustainable substrate for memristors, achieving operational frequencies up to 5.85 kHz with an accuracy of 90 ± 1%, showcasing its potential in bioelectronics and neuromorphic computing.

• FastJAM: a Fast Joint Alignment Model for Images (NeurIPS 2025)
  FastJAM introduces a lightweight graph-based framework for joint image alignment, achieving results in seconds compared to the hours required by previous methods, utilizing sparse keypoints and graph neural networks (GNNs) for enhanced efficiency.

• Exceptional Measurement of Chirality
  Breakthrough research enhances the measurement of flexible left- and right-handed molecules, known as chiral molecules, using a novel approach to Vibrational Circular Dichromism (VCD) spectroscopy, enabling real-time monitoring of biochemical processes and high-throughput pharmaceutical screening.

• We found LRMs look great…until the problems get harder (AACL 2025)
  Large Reasoning Models (LRMs) excel in simpler tasks but exhibit a sharp decline in performance as reasoning complexity increases, revealing a critical threshold beyond which they struggle.

• triton_bwd: Enabling Backpropagation for the OpenAI Triton language
  Triton language enables fast GPU kernel coding in Python, but lacks straightforward backpropagation for custom operations; a new library, triton_bwd, introduces automatic differentiation for Triton kernels.

• Update: Added Full Drift Benchmark Report (PKBoost vs LightGBM vs XGBoost — 16 Scenarios)
  PKBoost outperforms LightGBM and XGBoost by +50-60% in PR AUC gains, demonstrating superior performance across 16 drift scenarios, with an average PR-AUC of 0.8509 and minimal degradation of only 2.82%.

• Layer-0 heads that pre-bias hedging over facts in GPT-2 (replicated in Mistral-7B) — code + DOI
  Layer-0 heads in GPT-2 significantly downweight factual continuations while boosting hedging tokens, with zeroing specific heads improving logit-difference by +0.40–0.85 and enhancing calibration metrics (ECE 0.122→0.091, Brier 0.033→0.024).

• How to benchmark open-ended, real-world goal achievement by computer-using LLMs?
  Benchmarking open-ended, real-world goal achievement by LLMs reveals that while agents excel in verbal tasks, they struggle with real-world actions, often leading to inflated self-assessments of their performance.

• Korea Joins AI Industrial Revolution: NVIDIA CEO Jensen Huang Unveils Historic Partnership at APEC Summit
  South Korea's AI initiative involves deploying 260,000 NVIDIA GPUs across various sectors, marking a significant leap towards a sovereign AI infrastructure that aims to transform industries like manufacturing and robotics.

• A New Species of Artificial Intelligence: KMS-Stabilized Reasoning with Harmonic Algebra
  KMS-stabilized reasoning with harmonic algebra proposes a theoretical framework for AI that surpasses classical limits, enabling continuous processing and formal stability guarantees, which could lead to exponential speedups in specific problem classes.

• ExpertFlow: Adaptive Expert Scheduling and Memory Coordination for Efficient MoE Inference
  ExpertFlow enhances Mixture-of-Experts (MoE) inference by integrating adaptive expert prefetching and cache-aware routing, significantly reducing memory demand and computational overhead during model execution.

• Stop Wasting Your Tokens: Towards Efficient Runtime Multi-Agent Systems
  SupervisorAgent introduces a lightweight framework for runtime supervision in Multi-Agent Systems (MAS), enhancing efficiency by reducing token consumption by an average of 29.45% while maintaining success rates.