All Articles
16413 articles total
Post-Hoc Reasoning in Chain of Thought: Decoding and Steering Pre-Committed Answers
Mechanistic evidence shows that instruction-tuned models often pre-determine their answers before generating reasoning steps, allowing for prediction via linear probes on residual stream activations.
Evaluating Risks in Weak-to-Strong Alignment: A Bias-Variance Perspective
This research analyzes weak-to-strong alignment risks using a bias-variance-covariance lens, identifying strong-model variance as a key indicator of 'blind-spot deception'.
From Noise to Diversity: Random Embedding Injection in LLM Reasoning
The study introduces Random Soft Prompts (RSPs), a training-free method to improve LLM reasoning by injecting random embedding vectors, which increases early-stage token diversity.
Knowledge Graph Re-engineering Along the Ontological Continuum (extended version)
The paper proposes the 'ontological continuum' framework to better navigate and integrate diverse Knowledge Graph modeling practices in neuro-symbolic AI.
ARCO: Adaptive Rubrics with Co-Evolution for Multi-Step LLM-Based Agents
ARCO introduces an adaptive, co-evolving rubric-based reward system for multi-step LLM agents to improve interpretability and performance via step-level rewards.
Teaching LLMs String Matching, Backtracking, and Error Recovery to Deduce Bases and Truth Tables for the Combinatorially Exploding Bit Manipulation Puzzles
The authors present a novel algorithmic approach using string similarity and backtracking to help LLMs solve complex bit manipulation puzzles, winning awards in the NVIDIA Nemotron Challenge.
Generative AI and Agency in Education: A Critical Scoping Review and Thematic Analysis
A scoping review examines how Generative AI affects learner and teacher agency in educational settings, highlighting risks of inequality and diminished autonomy.
Loss-Complexity Landscape and Model Structure Functions
The research establishes a mathematical framework connecting model complexity to statistical mechanics, using loss-complexity landscapes to interpret phase transitions in generalization.
Show HN: Lucen a Python compiler that parallelizes for-loops via comment pragmas
A Python compiler named Lucen that parallelizes for-loops using comment pragmas.
Barzilai-Borwein Fails Superlinear Convergence on an Open Set of Quadratics for Every Dimension $n\geq 4$
Research proving that the Barzilai-Borwein method fails to achieve superlinear convergence on certain quadratic problems for dimensions n>=4.
GraphVid: Interactive Graph-Controllable Video Generation
Introduction of GraphVid, an interactive, graph-conditioned model for precise multi-object control in video generation.
3D-Aware VLMs with Implicit and Explicit Geometries
Presentation of VLM-IE3D, a framework that enhances 3D spatial awareness in Vision-Language Models using implicit and explicit geometric tokens.
A Counterfactual Cause in Situation Calculus
A theoretical study proposing a counterfactual analysis-based notion of cause within the situation calculus framework.
Fragile Preferences: A Deep Dive Into Order Effects in Large Language Models
Investigation into position biases in LLMs, revealing that model preferences can be influenced by the order of options, leading to sub-optimal decisions.
From Checklists to Clusters: A Homeostatic Account of AGI Evaluation
A proposed framework for evaluating AGI based on homeostatic property clusters rather than simple snapshot scores.
WebCoach: Self-Evolving Web Agents with Cross-Session Memory Guidance
Introduction of WebCoach, a self-evolving framework that provides web agents with cross-session memory to improve long-term planning and robustness.
Interpretable Embeddings with Sparse Autoencoders: A Data Analysis Toolkit
Proposing the use of Sparse Autoencoders (SAEs) as an efficient and controllable toolkit for analyzing and interpreting large-scale text corpora.
Understanding Critical Thinking in Generative Artificial Intelligence Use: Development, Validation, and Correlates of the Critical Thinking in AI Use Scale
A study on the development and validation of a scale to measure critical thinking in users of Generative AI.
Error Certificates for KV-Cache Eviction via Randomized Design
Researchers propose a randomized KV-cache eviction method to provide error certificates, allowing systems to distinguish between cache-induced and inherent model failures.
Compact Latent Coordination for Autonomous Vehicles at Unsignalized Intersections
The MAPS architecture uses a centralized 'proto-plan' embedding to coordinate autonomous vehicles at unsignalized intersections via hierarchical reinforcement learning.