Lev-3B

Thinking fast and deep with looped decision models

Project · Paper on alphaXiv · PDF · GitHub · Models & data

One checkpoint. Multiple depths. A probability for every option. Lev gives text decisions an adjustable compute budget: reuse the same recurrent weights, then read the shared decision head at the chosen depth. It returns typed option probabilities without decoding a chain of reasoning.

Native depths	JevBench · depth 1	JevBench · depth 2	Package
1–2	69.3%	72.7%	LoRA adapter + decision head

Paper measurements on 231 public JevBench items. See the full table and measurement conditions below.

Results · Quick start · Model details · Data · Limitations · Citation

The Lev family

Model	Backbone	Native depths
Lev-1.4B	Ouro-1.4B	1–4
Lev-2.6B	Ouro-2.6B	1–4
Lev-3B	Nanbeige4.2-3B-Base	1–2

Each repository contains its seed-0 adapter, decision head, tokenizer assets, and per-depth temperature calibration. Base model weights are downloaded separately.

Reported results

Depth	Public JevBench accuracy	Hard-tier accuracy	H200 bf16 median latency (ms)
1	69.3%	46.8%	23.0
2	72.7%	49.5%	44.7

These are the paper's existing measurements on 231 public JevBench items, not new evaluations of this package. Accuracy uses the study's evaluation path; timing is a separate, single-H200, bf16, one-request-at-a-time measurement without fused kernels or graph capture. Hardware and implementation changes may change latency. The checkpoint's weights_dtype is preserved from training; the actual evaluation kernel and backbone dtype must be matched for parity.

Lev-3B exceeded its fixed-depth frontier on locked test, but the result is workload-specific. Most controls and the 1.4B/3B headline models have one evaluated seed in that study.

Quick start

Use the accompanying Lev code, whose import name remains kev. A standard Transformers text-generation pipeline or PEFT loader alone does not implement the decision head and loop readouts.

After installing the companion code, download this repository:

from huggingface_hub import snapshot_download
checkpoint = snapshot_download("songdj/Lev-3B")

Then run the companion example with the downloaded local directory:

python examples/fixed_depth.py --checkpoint /path/to/downloaded-checkpoint \
  --depth 2 --temperatures /path/to/downloaded-checkpoint/temperatures.json

For a reproducible experiment, pin snapshot_download to the verified Hub commit hash shown in the repository history. Checkpoint.load uses the base identifier and immutable revision in head.pt to load the base tokenizer. The included tokenizer assets are preserved and cleaned for inspection and compatible local use.

Calibration

head.pt retains its original scalar temperature of 1.0. Per-depth calibration is a separate file, temperatures.json, and must be applied explicitly. It was fitted on 1,148 questions from the decision-v7 calibration partition. The fixed-depth example applies softmax(logits / T_depth); passing --all-depths computes every loop and returns all readouts.

Model and training

• Base: Nanbeige/Nanbeige4.2-3B-Base, pinned revision 4a38e817c14b9f2c4b69b1e05cbda90395872f6b.
• Native loop depths: 1–2; shared weights and pointer head at each depth.
• LoRA rank 16, alpha 32, dropout 0.05; one epoch, learning rate 1e-4, effective batch 64, uniform loss across depths, and seed 0.
• This seed-0 model accepted 31,334 records after context filtering; the constructed source mixture contains 18,764 records plus 12,576 replay records before context admission. Max state length is 5,120, with none-pair augmentation 0.25.
• Source training arguments and local provenance strings are removed from head.pt; all inference metadata and every head tensor are preserved. Adapter tensors are copied byte for byte. adapter_config.json explicitly pins the same base revision as head.pt.

Adaptive depth: offline analysis only

The paper's stopper predicts agreement with the deepest answer and is fitted on unlabelled requests from the target workload. Two-fold cross-fitting produces workload-specific classifiers and thresholds; there is no universal trained stopper in this model package. The companion scripts/stopper_config.py supports offline fitting and application. Live per-question stopping and continuation are not implemented in the current release code. Reporting fewer simulated loop steps is not a measured serving speedup.

Data availability

The Lev collection contains three models and three dataset repositories. Dataset packages are versioned separately from the original frozen inputs:

Resource	Purpose	Notes
LevBench	Development/test source configurations	Rights-filtered subset
Lev-train-v2	S31/S60/S120, replay and order recipes	Rights-filtered subset
Lev-results	Existing numerical results, calibration and latency	Question text removed

Each dataset card reports hosted counts, missing sources, pinned revisions, applicable licenses, and file hashes. These subsets do not reproduce the complete frozen corpus; unresolved sources require lawful local acquisition. No new evaluation was run for packaging.

See the companion data documentation and dataset cards for reconstruction and distribution requirements.

Limitations

Extra depth helps mostly on harder items and can also overturn a correct answer. Agreement with the deepest answer does not guarantee correctness. Temperature calibration and stopping behavior depend on the workload. The low-latency advantage of small Lev models partly reflects gaps in the dense baselines; at matched FLOPs on development data, Lev is at or below those alternatives. The study covers 1.4B–3B backbones and the public JevBench items.

validation.json records the package checks: decision-head metadata and tensors, and byte-identical adapter and tokenizer files. manifest.json gives SHA-256 hashes of the packaged artifacts. Packaging did not read test examples or training-question text and did not rerun evaluations.

License and upstream use statement

The pinned upstream model card declares Apache-2.0.

The base card at the pinned revision is the provenance for these statements; no separate LICENSE or NOTICE file was listed in that revision when this package was prepared. The Apache-2.0 text and attribution notices accompany this derived adapter/head package. Base weights and third-party data retain their own terms; no dataset text is included. See provenance.json for the pinned card digest and source artifact hashes.

Paper

Lev: Thinking Fast and Deep with Looped Decision Models — Dingjie Song, Zhiling Yan, Yuxuan Zhang, and Lichao Sun.

Read the paper on alphaXiv or download the PDF; the arXiv identifier will be added when it is announced. Code is on GitHub and the research overview is on the project page.

@misc{song2026lev,
  title        = {Lev: Thinking Fast and Deep with Looped Decision Models},
  author       = {Song, Dingjie and Yan, Zhiling and Zhang, Yuxuan and Sun, Lichao},
  year         = {2026},
  howpublished = {alphaXiv},
  url          = {https://www.alphaxiv.org/abs/2610.lev-looped-decision-models}
}