Structured decisions by institution, company, or university¶
简体中文 · English
Papers are grouped by the first author's institution. Products and community projects are labeled separately so open-source authorship is not mistaken for paper affiliation.
Independent Researcher¶
- When Multiple Choices Are Correct (2026-10-21, first author Rushab Munot): Distinguish normalized categorical shares from independently queried binary correctness probabilities; the former are not calibrated binary confidence. The paper studies Jev, Luna and Gemini. Bounded component only; full experiments are not reproduced.
Aalto University¶
- Multimodal Reranking Is a System One Decision (2026-10-20): Instead of generating a ranked list or training a separate pointwise reranker, the model scores each candidate by reading the probability of its letter label at the first answer position. The readout follows setwise and FIRST-style methods; the paper tests how this approach works zero-shot on images and visual documents, and whether text-only decision tuning transfers to visual reranking. Bounded component only; full experiments are not reproduced.
Câmara dos Deputados¶
- Jev-as-a-Judge for Brazilian Portuguese Information Retrieval: Relevance Assessment and Reranking (2026-10-16): On Quati it separately uses an independent Noul probability that each document directly answers the query, gathers every batch, and sorts the results. Candidates do not compete in a single Choice, probabilities are not mixed with BM25 scores, and there is no trained calibrator. Evaluation protocol interface; not a model capability result.
TypeSafe AI¶
- Jev: Official hosted System One model with a public typed Choice, Score and Noul decision protocol. Internal architecture, training data and weights are not released.
Knowledgator¶
- GLiClass (2025-08-11): This avoids unstable parsing of generative LLM outputs and separate cross-encoder inference on each text–label pair. The repository retains the defining mechanisms: candidate content affects scores, and one normalization produces a distribution over all candidates.
Nanyang Technological University¶
- Prism-1 (2026-10-15): A softmax combines independent scores. Softmax alone cannot guarantee this without attention isolation and position reset. Bounded component only; full experiments are not reproduced.
Massachusetts Institute of Technology¶
- RLCR (2025-07-22): RLCR predicts both an answer and its confidence and adds a bounded proper scoring rule so honest probabilities maximize expected reward. The local baseline minimizes multiclass Brier over the full candidate distribution.
University of Southern California / AWS AI Labs¶
- Calibration-Aware RL (2026-01-19): The paper adjusts decision-token probabilities directly to reduce ECE without sacrificing accuracy. The repository reports accuracy alongside Brier score and ECE, fits temperature only on validation, and forbids test-set tuning.
Independent community implementations¶
- Bespoke Labs — Nimble: Typed-schema logits model with public Qwen3.5-9B LoRA, data and recipe.
- TianyuCodings — NanoJev: Qwen3-0.6B with Choice, Boolean and Score decision heads.
- ConvAI Innovations — Laya: Public checkpoint, demo, fine-tuning and temperature-calibration path.
- TheoLeeCJ — SemIf: Engineering baseline with shared-state prefill and direct option logits.
This list covers priority integrations. See the full 48-entry snapshot for other community authors and projects. Community attribution comes from public repositories and is not paper affiliation.