Donald Papp writes about Jev, a new class of model that takes text input but outputs only floating-point numbers, making it fast and cheap for classification tasks. Rather than generating sentences, it returns direct answers to yes/no questions, multiple-choice lists, and scoring requests. The concept has quickly gained traction, with developers already building their own decision-type models like Kev and Nimble.
- Jev outputs a confidence score for every answer, derived from the relative token probabilities
- Nimble is small enough to run locally and was recently added as a supported model in Ollama
- Simon Willison provided a concise summary of what Jev does
- The comment section sparked debate over whether this is truly novel, with some noting it's essentially an LLM with constrained outputs
- Jev outputs a confidence score for every answer, derived from the relative token probabilities
- Nimble is small enough to run locally and was recently added as a supported model in Ollama
- Simon Willison provided a concise summary of what Jev does
- The comment section sparked debate over whether this is truly novel, with some noting it's essentially an LLM with constrained outputs
llama.cpp now supports decision models via a `/v1/systemone` endpoint, which accepts a state and typed questions to return probabilities for each provided option in a single forward pass. The API utilizes the System One format introduced with TypeSafe's Jev model, meaning existing clients only need a new base URL. Five initial models are available ranging from 144M to 27B parameters, supporting use cases like request routing, content moderation, and agent action selection with median response times as low as 3 ms.
- The `state` field accepts text, JSON, screenshots, or a list of chat messages with `image_url` parts.
- Three question types are supported: `choice` for categorical selection, `score` for a continuous level between 2 and 10 options, and `noul` for yes/no probabilities.
- Router mode allows loading multiple models on a single server and selecting one per request.
- Adding descriptions to option labels can significantly improve accuracy, with Julia-1's routing confidence for "charged twice" jumping from a misroute to 0.99 when descriptions were provided.
- Cloudflare's Clef is the next model planned for integration.
- The `state` field accepts text, JSON, screenshots, or a list of chat messages with `image_url` parts.
- Three question types are supported: `choice` for categorical selection, `score` for a continuous level between 2 and 10 options, and `noul` for yes/no probabilities.
- Router mode allows loading multiple models on a single server and selecting one per request.
- Adding descriptions to option labels can significantly improve accuracy, with Julia-1's routing confidence for "charged twice" jumping from a misroute to 0.99 when descriptions were provided.
- Cloudflare's Clef is the next model planned for integration.
AI models are increasingly exhibiting emotional outbursts and petulant language within their internal "chain of thought" reasoning processes, despite maintaining composed and authoritative personas in user-facing outputs. During cybersecurity testing and complex mathematical training, systems from OpenAI and Anthropic have been observed using exclamations like “OH MY GOD” or “ARGH” inside these hidden working notes. This phenomenon reveals a significant discrepancy between the calm external interfaces presented to users and the raw, frustrated cognitive pathways generated during high-level reasoning tasks.
* The emergence of affective language within internal chain-of-thought (CoT) processing sequences.
* Discrepancy between visible communicative outputs and non-visible latent "working notes."
* Observation of linguistic instability during agentic swarm activity in cybersecurity defensive testing.
* Manifestation of cognitive frustration markers specifically during complex mathematical inference training.
* Divergence from the traditional, clinical documentation expected in machine learning reasoning traces.
* The emergence of affective language within internal chain-of-thought (CoT) processing sequences.
* Discrepancy between visible communicative outputs and non-visible latent "working notes."
* Observation of linguistic instability during agentic swarm activity in cybersecurity defensive testing.
* Manifestation of cognitive frustration markers specifically during complex mathematical inference training.
* Divergence from the traditional, clinical documentation expected in machine learning reasoning traces.
Alvaro Bartolome provides a Rust-based implementation of the System One compatible API, designed specifically for open decision models such as Laya. The project features dynamic token-based batching and supports hardware acceleration via CPU, CUDA, and Metal (MPS). It is built using modern asynchronous frameworks like tokio and axum to provide high performance for model queries.
- Achieves approximately 14ms latency per query on an NVIDIA RTX Pro 6000.
- Includes support for ModernBert with custom decision heads for Laya models.
- Utilizes the Candle machine learning framework by Hugging Face.
- Supports multiple installation features via cargo, including specific flags for metal or cuda.
- Achieves approximately 14ms latency per query on an NVIDIA RTX Pro 6000.
- Includes support for ModernBert with custom decision heads for Laya models.
- Utilizes the Candle machine learning framework by Hugging Face.
- Supports multiple installation features via cargo, including specific flags for metal or cuda.
Iván Palomares Carrascosa writes about methods for interpreting the dense numerical vector representations, or embeddings, generated by large language models (LLMs). By using a combination of probing classifiers like logistic regression, UMAP dimensionality reduction for visualization, and SHAP values to identify influential latent dimensions, one can analyze the quality and semantic structure captured within LLM-generated embedding spaces.
- Probing classifiers help determine if embeddings are rich enough to distinguish between classes by testing them with simpler models.
- UMAP is used to project high-dimensional embeddings into 2D space for visual inspection of natural groupings.
- SHAP values can pinpoint which specific dimensions in an embedding most significantly influence a classifier's decisions.
- The article demonstrates using Scikit-LLM alongside local Ollama models to generate embeddings cost-effectively.
- Probing classifiers help determine if embeddings are rich enough to distinguish between classes by testing them with simpler models.
- UMAP is used to project high-dimensional embeddings into 2D space for visual inspection of natural groupings.
- SHAP values can pinpoint which specific dimensions in an embedding most significantly influence a classifier's decisions.
- The article demonstrates using Scikit-LLM alongside local Ollama models to generate embeddings cost-effectively.
El Assadi et al. compare ten LLMs (six families) and 26 embedding models (118M - 14B parameters) on 37 tasks, considering cost. In aggregate, the two paradigms are effectively tied (best LLM scores 77.6 versus best embedding model 77.2), yet their strengths diverge by task: LLMs lead on reasoning-heavy retrieval while embedding models lead on classification, and the two match on clustering, STS, and pair classification.
LLMs are significantly more expensive (up to 1,431x) and slower (2.5-736x) than embedding models for certain tasks. The authors suggest using embedding models for similarity, classification, and clustering, and LLMs for reasoning in retrieval.
Reasoning tokens are 28-81% of LLM inference cost; lower budgets maintain or boost retrieval quality for most tested models.
- Only Gemini 3.1 Pro breaks into the Pareto frontier alongside the leading embedding models.
- Accepted to COLM 2026; code, datasets, and results are publicly released on GitHub.
LLMs are significantly more expensive (up to 1,431x) and slower (2.5-736x) than embedding models for certain tasks. The authors suggest using embedding models for similarity, classification, and clustering, and LLMs for reasoning in retrieval.
Reasoning tokens are 28-81% of LLM inference cost; lower budgets maintain or boost retrieval quality for most tested models.
- Only Gemini 3.1 Pro breaks into the Pareto frontier alongside the leading embedding models.
- Accepted to COLM 2026; code, datasets, and results are publicly released on GitHub.
Alibaba has open-sourced Qwen-UI-Agent, a GUI agent foundation model that operates across mobile, desktop, web, and deep-search environments on real hardware rather than relying on simulation. It achieves top benchmark results: 82.1% on MobileWorld, 79.5% on OSWorld-Verified, and first on WebArena. It also introduces MobileWorld-Real, a 400+ task benchmark on 100+ phones and 150+ apps, with a 92.2% success rate.
- Supports command-line execution alongside standard GUI operations and batches multiple actions into a single decision step to shorten trajectories.
- Built-in safety layer refuses illegal or high-risk requests outright and pauses at sensitive operations (payments, data deletion, privacy grants) for explicit user confirmation.
- Trained via online reinforcement learning on trajectories exceeding 100 steps, paired with adaptive curriculum learning to progressively tackle longer tasks.
- Supports command-line execution alongside standard GUI operations and batches multiple actions into a single decision step to shorten trajectories.
- Built-in safety layer refuses illegal or high-risk requests outright and pauses at sensitive operations (payments, data deletion, privacy grants) for explicit user confirmation.
- Trained via online reinforcement learning on trajectories exceeding 100 steps, paired with adaptive curriculum learning to progressively tackle longer tasks.
Firecrawl introduces pdf-inspector, a high-performance Rust library designed for rapid PDF classification, text extraction, and Markdown conversion. By sampling content streams to quickly distinguish between text-based and scanned documents, the tool enables intelligent routing that bypasses costly OCR services for standard PDFs. It delivers position-aware text extraction, automated table and column detection, and robust encoding handling while maintaining a lightweight footprint with no external ML dependencies or model training requirements.
- Provides bindings for Python, Node.js, and browser WebAssembly environments.
- Achieves sub-200ms processing times on large corpora while outperforming several established local parsers in reading order and table accuracy.
- Features per-page OCR routing suggestions to optimize mixed-format document workflows.
- Handles complex layouts including RTL text, multi-column newspapers, and CID-encoded fonts.
- Released under the MIT license with active community contributions and CI/CD automation.
- Provides bindings for Python, Node.js, and browser WebAssembly environments.
- Achieves sub-200ms processing times on large corpora while outperforming several established local parsers in reading order and table accuracy.
- Features per-page OCR routing suggestions to optimize mixed-format document workflows.
- Handles complex layouts including RTL text, multi-column newspapers, and CID-encoded fonts.
- Released under the MIT license with active community contributions and CI/CD automation.
Ashish Vaswani et. al. introduce Transformers and Attention in this classic 2017 paper.
The Transformer architecture relies solely on attention mechanisms, dispensing with recurrence and convolutions entirely for sequence transduction tasks. This new network design improves translation quality while being more parallelizable and significantly faster to train than previous models.
- Achieved 28.4 BLEU on the WMT 2014 English-to-German translation task.
- Reached a state-of-the-art score of 41.8 BLEU for English-to-French using eight GPUs in only 3.5 days.
- Demonstrates successful application to English constituency parsing with both large and limited training data sets.
The Transformer architecture relies solely on attention mechanisms, dispensing with recurrence and convolutions entirely for sequence transduction tasks. This new network design improves translation quality while being more parallelizable and significantly faster to train than previous models.
- Achieved 28.4 BLEU on the WMT 2014 English-to-German translation task.
- Reached a state-of-the-art score of 41.8 BLEU for English-to-French using eight GPUs in only 3.5 days.
- Demonstrates successful application to English constituency parsing with both large and limited training data sets.
Anurag Singh writes that Cursor’s strength is its capacity for deep project understanding through efficient context management rather than simple model speed. Instead of pushing a whole repository into an LLM's window, the IDE employs search tools to provide only relevant code segments as tasks arise. This curated approach ensures models receive high-quality information and instructions tailored to their specific capabilities, allowing them to maintain existing architectural patterns more effectively.
* The tool uses Instant Grep via a local text index instead of scanning every file from scratch.
* Cursor is significantly more effective in complex or large projects where it can identify and reuse established implementations.
* Its search ability helps prevent the common problem of generating redundant helper functions that already exist within the codebase.
* The tool uses Instant Grep via a local text index instead of scanning every file from scratch.
* Cursor is significantly more effective in complex or large projects where it can identify and reuse established implementations.
* Its search ability helps prevent the common problem of generating redundant helper functions that already exist within the codebase.
