AI/ML arXiv cs.AI

Beyond Epistemia: Epistemic Schizologia and Large Language Models as Techno-Semiotic Machines

A philosophical exploration of LLMs as 'techno-semiotic machines' and the concept of 'epistemic schizologia' regarding the gap between linguistic production and responsible understanding.

AI/ML arXiv cs.AI

Quotient Dynamics, Effective Curvature, and Implicit Bias in Positive Quadratic Networks

A mathematical study on the training dynamics, curvature, and implicit bias of positive quadratic networks using quotient Riemannian gradient flow.

AI/ML arXiv cs.AI

Joint Text-Audio Alignment for EEG-to-Text Decoding in Chinese Speech Production and Perception

Presents EEGAlign, a framework that jointly aligns EEG signals with text and audio embeddings to decode Chinese speech production and perception into text.

AI/ML arXiv cs.AI

AIriskEval-edu Demo: Auditing of Pedagogical Risks in Educational Explanations

Introduces AIriskEval-edu Demo, a platform for auditing pedagogical risks in educational explanations using a fine-tuned Llama 3.1 8B model.

AI/ML arXiv cs.AI

Engine-Equal, Human-Unequal: A Reproducible Outcome Skew in Engine-Assessed Equal Chess Positions

An observational study showing that human outcomes in chess positions judged as equal by Stockfish 18 are skewed, suggesting engine evaluation is not a sufficient statistic for human play.

AI/ML arXiv cs.AI

OrchBench: Evaluating Multi-Agent Orchestration Plans in Isolation via Deterministic Simulation

Presents OrchBench, a simulation-based benchmark for evaluating multi-agent orchestration plans in isolation, reducing token and time costs compared to end-to-end execution.

Software Engineering Hacker News

Fixing a bug with byte order marks

A discussion on Hacker News regarding the technical challenges and solutions for fixing bugs related to byte order marks (BOM).

AI/ML arXiv cs.AI

CoTinyVLA: Chain-of-Thought Distillation for a Sub-Billion-Parameter Vision-Language-Action Model

Introduction of CoTinyVLA, a sub-billion parameter vision-language-action model that achieves high robustness in robotic tasks through structured supervision and CoT distillation.

AI/ML arXiv cs.AI

Are the High-weight Neurons the Important Ones in Image Classification Neural Networks?

Research challenging the assumption that high-weight neurons are always the most important in image classification networks, suggesting a non-linear relationship.

AI/ML arXiv cs.AI

Entangled by Design: Spurious Intra-Variable Signal Routing in Tabular In-Context Learners

An analysis of spurious signal routing in tabular in-context learners and the introduction of mitigations like S-swap augmentation to improve causal sensitivity.

AI/ML arXiv cs.AI

From Training to Deployment: Post-Hoc Causal Feature Identification via Sensitivity Ratios

Introduction of the Normalised Sensitivity Ratio (NSR), a post-hoc, model-agnostic diagnostic for identifying causal versus spurious features in trained models.

AI/ML arXiv cs.AI

Distilling Temporal Search and Reasoning: Evolving LLMs for Future Prediction via Harness-Assisted Efficient Data Synthesis

A proposal for a time-truncation harness to synthesize higher-quality data for evolving LLMs' temporal search and reasoning capabilities for future prediction.

AI/ML arXiv cs.AI

Agent Skills Matter: Inferring Proprietary Skills from Execution Trajectories

Introduction of SigLeak, a framework that demonstrates how proprietary agent skills can be reconstructed from execution trajectories through a behavioral side channel.

AI/ML arXiv cs.AI

Matrix-Free Photoacoustic Image Reconstruction via Sensor-Token Self-Attention

The Sensor Attention Network (SAN) uses a Transformer-based architecture to enable real-time photoacoustic image reconstruction without requiring a system matrix at inference.

AI/ML arXiv cs.AI

A Density-Matrix Framework for Electronic-Structure Analysis of Functional-Group and Salt Effects in Lithium-Metal Electrolytes

Presentation of EMolStudio, an AI platform for electronic-structure prediction and analysis of lithium-metal electrolytes to optimize battery chemistry.

AI/ML arXiv cs.AI

COVENANT: Natural-Language Workflow Compilation for Aligned Agent Execution

COVENANT introduces a compiler-and-interpreter architecture that treats natural-language workflow instructions as source programs to prevent LLM agents from skipping steps or taking unsupported branches.

AI/ML arXiv cs.AI

Context Assembly as the Controlled Variable: A Control-Theoretic View of Harness Policies for Frozen LLM Agents

This paper proposes a control-theoretic approach to LLM agents where context assembly (prompting, few-shot examples, etc.) is the controlled variable optimized by an external policy.

AI/ML arXiv cs.AI

A Control System, a Dataset, and a Recipe for Making Frozen LLM Agents Learn a Domain

The authors present a system for making frozen LLM agents learn domains by optimizing a human-legible 'harness' using reinforcement learning and DSPy.

AI/ML arXiv cs.AI

Salient Knowledge Pathways: Sparse Cross-Modal Routing for Efficient Knowledge-Intensive Multimodal Question Answering

SKIP is a new inference architecture for multimodal QA that uses sparse routing and token pruning to significantly reduce FLOPs and latency without losing accuracy.

Cybersecurity arXiv cs.AI

The Disruptive Impact of Large Language Models on Capture the Flag Competitions and the Path Toward Fair Play

A study on how LLMs disrupt Capture the Flag (CTF) competitions, proposing a safeguard framework to maintain the educational value and fairness of cybersecurity contests.