FrontiersMindAI released DecisionFormer, a Python library and CLI for fine-tuning, loading, and serving decision models on top of any Hugging Face model. Given a state and typed questions, it returns calibrated probabilities for every option in a single forward pass without text generation. It targets practitioners building routing and decision systems rather than chat applications.
A new open-source project, LLM2Jev, adapts local Hugging Face-compatible causal language models into Jev-style structured decision engines, emitting typed Choice, Score, and Noul answers with probabilities. It uses prefill-only binary inference — scoring candidates with yes/no logit judgments without generating answer tokens — and leverages SGLang's Radix Cache for shared-prefix efficiency. The project is independent and unaffiliated with Jev or TypeSafe.
A new open-source project, LLM2Jev, adapts local Hugging Face-compatible causal language models into Jev-style structured decision engines, emitting typed Choice, Score, and Noul answers with probabilities. It uses prefill-only binary inference — scoring candidates with yes/no logit judgments without generating answer tokens — and leverages SGLang's Radix Cache for shared-prefix efficiency. The project is independent and unaffiliated with Jev or TypeSafe.
InternLM released Intern-Decision-4B, a multimodal structured decision model fine-tuned from Qwen3.5-4B, alongside a 0.8B variant, with model weights, a demo, and a GitHub repo. The models take a shared state, a question schema, and optional images, and return answer distributions for all questions in a single forward pass without free-form generation. This gives practitioners a deterministic structured-candidate-scoring API for decision tasks rather than autoregressive text sampling.
InternLM released Intern-Decision-4B, a multimodal structured decision model fine-tuned from Qwen3.5-4B, alongside a 0.8B variant, with model weights, a demo, and a GitHub repo. The models take a shared state, a question schema, and optional images, and return answer distributions for all questions in a single forward pass without free-form generation. This gives practitioners a deterministic structured-candidate-scoring API for decision tasks rather than autoregressive text sampling.
The model's developer released DecisionTune 1.0, an Apache-2.0 395M encoder (ModernBERT-large plus a 4 KB scoring head) that scores given options rather than generating text, aimed at routine agent decisions like tool routing and yes/no checks. It runs fully locally via PyTorch, MLX, or ONNX backends and ships with an MCP server for local assistants. The developer is seeking feedback on where it picks wrong.
The model's developer released DecisionTune 1.0, an Apache-2.0 395M encoder (ModernBERT-large plus a 4 KB scoring head) that scores given options rather than generating text, aimed at routine agent decisions like tool routing and yes/no checks. It runs fully locally via PyTorch, MLX, or ONNX backends and ships with an MCP server for local assistants. The developer is seeking feedback on where it picks wrong.
Perplexity released Decider v1.3 and Decider Flash v1.3, open-weight decision models that return probability distributions over answer options instead of generating text. Both support text and images, and v1.3 adds a new `locate` type that extracts verbatim spans with character offsets for NER, PII detection, and extractive QA. The models are available on Hugging Face, OpenRouter, and Perplexity's API.
Perplexity released Decider v1.3 and Decider Flash v1.3, open-weight decision models that return probability distributions over answer options instead of generating text. Both support text and images, and v1.3 adds a new `locate` type that extracts verbatim spans with character offsets for NER, PII detection, and extractive QA. The models are available on Hugging Face, OpenRouter, and Perplexity's API.
A new open-source tool called brier lets any open model on Hugging Face produce Jev-style typed decisions with calibrated probabilities in one pass, no generation or fine-tuning. It reads answer distributions from next-token log-probabilities and corrects position bias and miscalibration at three levels (zero, 50+, or 60+ labels). On banking20, zero labels lift Falcon3-1B-Base from 16% to 60% accuracy, and 300 labels bring six model families to 80–86%.
A new open-source tool called brier lets any open model on Hugging Face produce Jev-style typed decisions with calibrated probabilities in one pass, no generation or fine-tuning. It reads answer distributions from next-token log-probabilities and corrects position bias and miscalibration at three levels (zero, 50+, or 60+ labels). On banking20, zero labels lift Falcon3-1B-Base from 16% to 60% accuracy, and 300 labels bring six model families to 80–86%.