small language model

Token cost by programming language: the same programs in 10 languages on Claude, Gemini and GPT

Two small programs (an inventory class and Dijkstra's shortest path) written in Python, Ruby, JavaScript, TypeScript, PHP, Rust, Java, C++, Go and C, with input tokens counted on Claude Opus/Sonnet 5.5, Claude Haiku 4.5, Gemini 3.x and GPT (o200k_base). Headline: both programs together cost 809 GPT tokens in Python and...

Milli: Continuous Spatiotemporal Manifold Engine for Token-Efficient Visual Intelligence

Frontier Multimodal Large Language Models (MLLMs) process dynamic visual feeds by discretizing time into frame arrays and space into rigid 14x14 pixel patches. To prevent quadratic self-attention O(N^2) from exhausting context windows, production pipelines downsample incoming video to 1–2 frames per second (FPS), disca...

Evidence-Grounded Large Language Model Reasoning for Patient Safety Using Knowledge Graphs and Retrieval-Augmented Generation: Development and Evaluation Study

ZeroFloat: Fully Integer Training of Language Models Matches AdamW up to 1B Parameters

We train transformer language models from scratch in which every training operation is integer arithmetic: the forward pass (matrix products, RMSNorm with integer square root, multi-head attention with an integer base-2 softmax), the backward pass (including the softmax and normalization Jacobians), and the parameter u...

From Dyadic Derivatives to Pareto Circuit Surgery: Resolving the Quantization Wall, Static-Vector Dilution, and Low-Rank Attention Editing in Qwen2.5-7B

Prior work on Qwen2.5-7B-Instruct showed that the raw size of a residual-stream displacement does not predict its effect, while a simple pre-intervention measure of local sensitivity does rank which soft patches will move the model’s decision margin. Three problems remained. A second-order curvature assay failed, in pa...

Web search in Claude Code and Codex: tokens, searches and sources per question

Four kinds of question asked with web search on in Claude Code and Codex, two runs each: input and output tokens, searches and page opens, time, the sites seen in results and the sites cited in the answer, plus the same questions with no tools. Headline: with web search on, a one-line fact question cost 150,414 input t...

Deployment Matrix for Critical Infrastructure and Defence: Kolibri-1, Qwen3.6 and Mistral Small 4 on a Deterministic German Rule-Application Benchmark

English Three language models (Aleph Alpha Kolibri-1, Qwen3.6-35B-A3B, Mistral Small 4) were compared on one H200 with vLLM 0.29.0 on a synthetic German rule-application benchmark. Each question comes with its source passages plus three distractors, so the test measures reading and rule application, not knowledge store...

DIGITAL PSYCHIATRY; Current Evidence, Challenges, and Future Directions — Integrating SDM, D-SDM, DRP, and DRL

This research paper presents an interdisciplinary examination of digital psychiatry and its emerging applications in mental health assessment, treatment, monitoring, and artificial intelligence-assisted psychological care. Drawing on peer-reviewed research involving telepsychiatry, smartphone-based interventions, virtu...

Lightweight Transformer-based Ai Model for Multilingual and Code-mixed Sentiment Analysis

Social media users increasingly communicate across language boundaries, often combining two or more languages within the same sentence and frequently writing Indian languages in Roman script. This creates a difficult setting for sentiment analysis because conventional monolingual models must cope with spelling variatio...

Beyond Better Models: Barriers and a Path Forward in Generalisable TCR Antigen Specificity Prediction

T cells are at the centre of adaptive immunity, protecting the organism against infection and cancer. T cell receptors (TCRs), expressed on the surface of T cells, enable the recognition of target cells through interactions with peptides presented by major histocompatibility complex (pMHC) molecules. Reliable, large-sc...

Abstract C009: Impact of a national community advisory board on clinical trial design, participant recruitment, and implementation: 1 year of experience

Abstract Engaging underrepresented populations in clinical trials remains a persistent challenge. To promote equitable participation in a hybrid decentralized study (i.e., recruitment and enrollment at sites or virtually), we created a national community advisory board (CAB), defined as a group of community members who...

Verbatim evidence quotes and answer correctness in small local language models: evaluation code, model responses and results

Evaluation code, configuration, model responses, item-level scores and result tables for a study of whether a verbatim evidence quote shows that a language model extracted a clinical trial result correctly. Local Qwen3 models (1.7B primary, 0.6B, and 4B-Instruct-2507) returned a direction label (significantly increased...