small language model

From Aligned Embeddings to Instruction-Following Perception: A Survey of Vision–Language Model Development

Recognizing an object from a text label, describing an unfamiliar scene, and following an instruction about a small image region require different forms of visual-language correspondence. The development of vision-language models can be understood through the effort to support these increasingly varied outputs. This su...

Small Language Model Log

Taking Self-Awareness Apart: A Structure-by-Structure Test of Introspection, Retention and the Self-Model in Language Models

Prompts, model responses, blind-coding files, pre-registrations, code and outputs for fifteen experiments testing introspection, retention and the self-model in language models structure by structure (Claude Opus 5 / 5.5, Sonnet 5.5, Haiku 4.5 / 5.5; Qwen2.5-1.5B/3B with LoRA fine-tuning and hidden-state probes; small...

Deep-Learning-Based Assessment of Breast Arterial Calcification on Mammography: Technical Validation, Cardiovascular Prognostic Value, and Clinical Translation—A Narrative Review

Breast arterial calcification (BAC) is a manifestation of medial arterial calcification frequently detected incidentally on screening mammography. Increasing evidence suggests that BAC may also serve as a marker of cardiovascular risk. This narrative review critically synthesizes the evidence on artificial intelligence...

Evaluation of Algorithmic Carbon Footprint and Artificial Intelligence (AI) Tools in Environmental Education Programs: Systematic Review

The integration of Artificial Intelligence (AI) in educational settings creates pedagogical opportunities while raising environmental concerns regarding data center energy consumption and physical infrastructure footprint. This systematic review synthesizes scientific evidence on the algorithmic carbon footprint of AI...

Readout Scaling Shapes Sparse Feature Recovery

Together, these results show how readout scaling, the data distribution, and training shape sparse learning, and RC-DS provides an information-preserving method for studying and controlling the readout's contribution.

Argument Graph Uncertainty: Quantifying Uncertainty from the Logical Structure of Reasoning Chains

This work introduces Argument Graph Uncertainty, a post-hoc framework for multiple choice question tasks that estimates model confidence directly from the logical structure of a single chain-of-thought reasoning trace, and demonstrates through extensive combination analyses that argument graph uncertainty captures an o...

STP-GUI: Spatially-Aware Token Pruning for Efficient Web-GUI Agents

STP-GUI (Spatially-aware Token Pruning for GUI agents), a query-guided token pruning framework that preserves fine-grained instruction-conditioned token scoring while performing spatially coherent window-level selection under aggressive token reduction is proposed.

VisualWorldBench: A Fine-Grained Multi-Task Benchmark for Evaluating Visual World Knowledge in MLLMs

This work introduces VisualWorldBench, a taxonomy-grounded multi-task benchmark built on a hierarchical taxonomy of 21,291 leaf-node classes across 8 visual domains, with 32,541 evaluation instances whose ground-truth labels cover 15,479 categories across four complementary tasks.

XViewRank: Overlap-Conditioned Score Fusion for LLM-Augmented Sparse Retrieval

XViewRank is presented, a lightweight retrieval-side method for conditionally fusing BM25 results from an original query, a hypothesis view, and a step-back view, and the results support overlap-conditioned score fusion as a compact retrieval-side alternative to fixed fusion, while also delimiting where its advantage i...

The Cost of Context: Profiling the Energy Footprint of Input Tokens in Large Language Models

This work establishes context-scaling laws that differentiate engine-level optimisations from inherent model constraints, providing an empirical framework for sustainable LLM deployment.

DarKnight: Retrieval-Guided Agentic Framework for Breaking Dark Web CAPTCHA Gates

DarKnight, a self-improving agentic framework designed to run on a single workstation GPU inside an investigator's Tor crawler and solve CAPTCHAs with small, locally hosted open-weight VLMs that no provider decision can withdraw, without per-market training or third-party services is presented.