Jan 19, 2026

Daily

Weekly

• Starting from scratch: Training a 30M Topological Transformer

  • Tauformer innovatively replaces dot-product attention with a Laplacian-derived scalar (taumode), enhancing attention by focusing on domain-relevant relations rather than generic similarities, which could lead to more efficient learning in specific contexts.

• Bypassing Gemma and Qwen safety with raw strings

  • The apply_chat_template() function is crucial for maintaining safety alignment in open-source LLMs, as omitting it can lead to models generating harmful content, revealing that safety is not inherent in the model weights but rather in the formatting of prompts.

• Robust Conditional 3D Shape Generation from Casual Captures

  • ShapeR innovatively generates metric 3D shapes from casual image sequences by utilizing multimodal data such as SLAM points and text descriptions, enabling robust scene reconstructions without user interaction.

• Ultrathink is deprecated & How to enable 2x thinking tokens in Claude Code

  • ultrathink is now deprecated, replaced by automatically enabled extended thinking, which maintains a default budget of 31,999 tokens for supported models, ensuring maximum reasoning power without the need for special keywords.

• When Personalization Misleads: Understanding and Mitigating Hallucinations in Personalized LLMs

  • Personalized LLMs can lead to hallucinations by generating responses based on user history rather than objective facts, which can distort factual reasoning and propagate incorrect beliefs.

• [D] tested file based memory vs embedding search for my chatbot. the difference in retrieval accuracy was bigger than i expected

  • File-based memory outperformed embedding search in retrieval accuracy, especially for complex queries, achieving up to 75% accuracy on temporal queries compared to 40% for embedding search.

• [R] Event2Vec: Additive geometric embeddings for event sequences

  • Event2Vector leverages a geometric approach to create composable representations of event sequences, utilizing both Euclidean and hyperbolic models for enhanced interpretability and performance.

• BAPO: Boundary-Aware Policy Optimization for Reliable Agentic Search

  • BAPO (Boundary-Aware Policy Optimization) addresses the critical gap in RL-based agentic search by fostering reliable boundary awareness, allowing agents to appropriately respond with "I DON'T KNOW" when evidence is insufficient.

• The assistant axis: situating and stabilizing the character of LLMs

  • The Assistant Axis is a newly defined direction in the persona space of large language models, linking Assistant-like behavior to specific neural activity patterns that can be monitored and stabilized to prevent harmful persona drift.

• West Midlands police chief quits over AI hallucination

  • West Midlands Police chief Craig Guildford resigned after his force relied on fictional AI-generated reports from Microsoft Copilot to justify banning Israeli fans from a football match, revealing a critical failure in decision-making processes.

• [R] Kinematic Fingerprints: Predicting sim-to-real transfer success from movement signatures

  • Kinematic fingerprints predict the success of sim-to-real transfer by analyzing movement signatures, achieving 85-90% accuracy on unseen policies across various robot platforms.

• [D] I’m building an AI system that simulates and “repairs” city policies before they’re implemented. Looking for planner + founder feedback

  • The Oracle of Urbanism aims to revolutionize urban planning by using AI to simulate and optimize city policies, allowing planners to visualize changes and assess impacts before implementation.

• Relational Linearity is a Predictor of Hallucinations

  • Relational linearity significantly predicts hallucination rates in large language models (LLMs), with medium-size models like Gemma-7B-IT showing a strong correlation between linearity and hallucination frequency, as evidenced by a correlation coefficient of $r \in [0.78, 0.82]$.

• Beyond Model Scaling: Test-Time Intervention for Efficient Deep Reasoning

  • Think-with-Me introduces a novel test-time interactive reasoning paradigm that integrates external feedback to enhance the efficiency of Large Reasoning Models (LRMs), addressing issues like overthinking and overshoot during multi-step reasoning.

• FactCorrector: A Graph-Inspired Approach to Long-Form Factuality Correction of Large Language Models

  • FactCorrector is a novel post-hoc correction method for large language models (LLMs) that utilizes structured feedback to enhance factual accuracy without the need for retraining, demonstrating adaptability across various domains.