m-polignano/decision4rec-kev-4b

Decision4Rec Kev-4B (pool reader)

The trained decision model of Decision4Rec-Full, the full architecture of Decision4Rec: it reads the 30 best candidates of a multi-agent pool, with item text, statements of the knowledge graph of the scenario and the agents that found each candidate, and returns one probability per candidate in a single forward pass (System One protocol, choice question).

Private preview. This model will be made public with the paper.

Model details

Base model	Qwen/Qwen3.5-4B-Base (revision 1001bb4d826a52d1f399e183466143f4da7b741b), frozen
Starting point	the released jaredpalmer/kev-4b adapter and pointer head
Trained parameters	rank-16 LoRA adapters on all linear layers, pointer head
Training data	34,667 real next choices of users of Amazon Video Games, Amazon Musical Instruments and Steam, over the 30 best candidates of the pool (480 longer records dropped)
Objective	likelihood of the item the user really chose next; no outputs of any commercial model were used
Recipe	one epoch, learning rate 5e-5, one-cycle schedule, micro-batch 2 x 4 accumulation (8 records per step), bf16 autocast, gradient checkpointing, states up to 1,792 tokens
Trainer	Kev kev.train, commit 84847f0

Results

NDCG@10 of Decision4Rec-Full on 2,000 development test users per domain (full-catalogue ranking, leave-one-out). Industrial and Scientific was never seen in training.

System	Video Games	Musical Instr.	Steam	Industrial & Sci.
best trained baseline	0.0622	0.0370	0.0790	0.0310
pool and residual ranker, no decision model	0.0730	0.0468	0.0786	0.0386
with this model (fused)	0.0780	0.0504	0.0801	0.0407

Use

Serve it with Kev and point Decision4Rec at the server:

cd ~/kev && uv run --extra serve python -m kev.serve --run m-polignano/decision4rec-kev-4b --port 8009
decision4rec run --config configs/small/vg_rtm.yaml --backend configs/backends/kev-4b-rtm.yaml

The model expects the requests built by Decision4Rec (state with the user history and statements; one choice question with the candidates and their "found by" lines). Other requests work, but were not part of training.

Limitations

• Trained on English product and game catalogues; dense domains with long histories (MovieLens) gain less.
• Development results only; a confirmatory test on fresh users is planned.
• It ranks candidates it is given; it does not retrieve.

Licenses

The adapter and head are released under Apache-2.0. The base model keeps its own license (Qwen3.5). The training data come from public research datasets that are not redistributed here.

Citation

@article{polignano2026decision4rec,
  title  = {Decision4Rec: A Jev-Inspired Framework of Decision Models on Knowledge and Active Graphs for Multi-Agent Recommender Systems},
  author = {Polignano, Marco and Mehta, Het Darshan and Lops, Pasquale and De Luca, Ernesto William and Semeraro, Giovanni},
  year   = {2026}
}