Back to AI Research

AI Research

Jev-DLC turns explicit corrections into compact state for later decisions

Key Takeaways

  • The method stores action identities and preference relationships above a frozen scorer, with chronological evaluation of how feedback affects future choices.
  • Jev-DLC studies personalization through explicit feedback on decisions.
  • A user may accept one action or revoke it in favor of another.
  • The method uses typed interfaces to identify both the selected action and, when relevant, the decision being replaced, so the correction has a defined meaning.
  • The [paper](https://arxiv.org/abs/2610.31521) treats a correction as evidence about a recurring action identity and as a directed relationship between preferred and rejected choices.

Jev-DLC studies personalization through explicit feedback on decisions. A user may accept one action or revoke it in favor of another. The method uses typed interfaces to identify both the selected action and, when relevant, the decision being replaced, so the correction has a defined meaning.

Preserve two kinds of feedback

The paper treats a correction as evidence about a recurring action identity and as a directed relationship between preferred and rejected choices. It stores these signals in complementary bounded channels above a frozen scorer.
That separates the feedback representation from retraining the underlying model. The system uses the stored signals to shape later scores without retaining raw interaction history, adding a scorer call or performing a runtime language-model backward pass, according to the authors.

Test effects on later requests

The researchers introduce TCPA-Bench, a chronological evaluation that credits a correction only through its effect on later decisions. This prevents a method from appearing personalized merely because it can repeat the correction on the same event.
On TCPA-Repeat, Jev-DLC reaches 44.83 percent user-macro accuracy overall, 2.43 percentage points above Recent-4 history replay. After at least two earlier feedback events, it reaches 56.07 percent, a 7.54-point margin over the strongest reported comparison at that depth.

Feedback quality remains part of the system

Those figures belong to the paper’s benchmark and feedback conditions. They do not establish that every user preference can be represented by one stable action identity, or that compact state provides a complete privacy guarantee.
For personalization researchers, the useful design question is how a correction should persist when a similar decision returns. Explicit typed feedback can make that relationship easier to examine than an undifferentiated chat history. Evaluation should still test conflicting corrections, changing preferences and effects on later tasks, with the time ordering preserved.

Comments