Reading 11 papers

Decision models

In PROJev decision model family, System One decision models, Looped Transformer

1. Oct 2026 Munot

   When Multiple Choices Are Correct

   Tests Jev, Liquid d1, and GPT-6 Luna Decisions on questions with several correct options. Because the probabilities must sum to one, the models concentrate mass on one correct option, so a fixed threshold misses the others, even when every correct option is ranked above every wrong one.

   ConnectsSeveral options are right, but the probabilities must sum to one. Commit or Abstain? studies the mirror case, where no option is right. See Commit or Abstain?, Beyond Calibration.

   In PROJev decision model family, LLM evaluation

2. Oct 2026 Demir

   Commit or Abstain?

   Replaces the correct holding in about a third of CaseHOLD legal questions with a similar real holding from another case and adds a "none of these" option. Generative LLMs still pick a wrong holding far more often than the most cautious decision model. It also distills a decision model's probabilities into Qwen3-8B with no labels.

   ConnectsDistils a decision model into an open LLM, the opposite direction from Releasing the Inner Jev, which trains open models into decision models. See Releasing the Inner Jev, When Multiple Choices Are Correct.

   In PROLLM evaluation, System One decision models

3. Oct 2026

   Releasing the Inner Jev

   An open recipe for Jev-style models: label options with letters, start a decision head from the language-model head's rows for those letters, and fine-tune with LoRA. The authors report that open models trained this way outperform Jev on a wide range of decision tasks.

   ConnectsLEV and SanSi change the backbone instead of the head, spending more passes on harder questions. See LEV, SanSi.

4. Sep 2026

   Beyond Calibration

   Checks whether Jev's probabilities obey the basic rules of probability, for example whether the probabilities of mutually exclusive labels add up correctly. Jev's violations cluster where it is least certain.

   ConnectsEvidence for the deployment rule in Calibration and Risk-Budgeted Decisions: Jev's violations cluster where it is least certain. See Calibration and Risk-Budgeted Decisions with System One Models, When Multiple Choices Are Correct.

5. Oct 2026

   What Can't Jev Decide?

   Grades decision items by how much reasoning a reference LLM needs to solve them. Jev handles items that one forward pass of that LLM can solve, but falls behind on items that need thinking.

   ConnectsDraws the reasoning boundary the next three papers attack, by handing Jev a partial reasoning trace or by spending more passes inside the model. See Truncated Reasoning Under a Budget, LEV, SanSi.

   In PROJev decision model family, System One decision models

6. Oct 2026

   Truncated Reasoning Under a Budget

   Cuts a reasoning model's chain of thought at a token budget and lets Jev pick the answer from the partial trace, which beats forcing the reasoner to answer early.

   ConnectsA reasoner in front, Jev as the answer head. The grading in What Can't Jev Decide? says which questions need the reasoner at all. See What Can't Jev Decide?.

   In PROJev decision model family, System One decision models

7. Oct 2026

   LEV

   A looped decision model that reads the options after each pass through a looped backbone, with a learned stopper that spends more passes on harder questions.

   ConnectsSanSi trains one looped model to serve every compute budget, from a model pretrained to loop. See SanSi, What Can't Jev Decide?.

   In PROSystem One decision models

8. Oct 2026 University of Minnesota

   SanSi

   Turns Ouro, a language model pretrained to loop, into a decision model, training every loop with a proper scoring rule so one model serves every compute budget.

   ConnectsLEV is the looped decision design trained with a stopper instead of a per-loop scoring rule. See LEV.

9. Oct 2026

   System X

   Adds a "reply with text" option, so when the model needs to talk to the user it continues generating from the KV cache built for the decision instead of re-reading the prompt.

   ConnectsLAVOIR scores follow-up questions in the same pass instead. See LAVOIR.

   In PROLLM inference optimization, System One decision models

10. Sep 2026

    LAVOIR

    Puts candidate pieces of missing information next to the options, so the same forward pass predicts both the decision and how much asking each follow-up question would help.

    ConnectsSystem X handles the moment the model needs to speak instead of choosing. See System X.

11. Oct 2026

    Calibration and Risk-Budgeted Decisions with System One Models

    Argues that safe deployment needs task-specific validation and explicit risk budgets, not an assumption that a decision model is calibrated everywhere.

    ConnectsThe deployment rule that Beyond Calibration's findings support: validate per task, budget the risk. See Beyond Calibration.

    In PROSystem One decision models