Lev-results

The evidence behind thinking fast and deep.

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

Depth-wise outputs, calibration and latency evidence — with question text removed.

Native question observations	JevBench depth observations	Study models
472,902	19,173	31

Coverage: This package contains existing study artifacts; coverage and omissions are documented below.

Overview · Coverage · Quick start · Format · Provenance & limits · Explore Lev

---

Overview

Inspect the stored evidence behind Lev's depth–quality tradeoffs. This repository brings together per-question outputs, existing aggregates, fitted calibration and latency material from the completed study.

Material	What you can inspect
native-depth-rows	Raw logits, depth and option indices, model, seed, group and numeric targets
jevbench-probabilities	Stored probabilities, confidence, validity and correctness at each depth
Results	Existing aggregate and latency material
Calibration	Existing fitted calibration artifacts
Figures	Archived study figures

Question, option and state text are removed. Stored numerical outputs are preserved; this package introduces no new evaluation scores.

Coverage

Export	Hosted observations	Files	Unit
Native depth rows	472,902	158	One question with its depth × option logits
JevBench probabilities	19,173	83	One question at one depth

The two row counts use different units and should not be summed as a unique-question total. Outputs cover 31 models; available identities, splits and calibration coverage are listed in coverage.json.

Explicit gap: 3 supplementary files (3,222 observations) are withheld because of duplicate question identities. No new evaluation was used to fill gaps. See the coverage record for skipped rows and missing material.

Quick start

Save an immutable revision before loading:

from huggingface_hub import HfApi
from datasets import load_dataset

repo = "songdj/Lev-results"
revision = HfApi().dataset_info(repo).sha  # save for reproducibility

native = load_dataset(
    repo, "native-depth-rows", split="development",
    revision=revision,
)
jevbench = load_dataset(
    repo, "jevbench-probabilities", split="test",
    revision=revision,
)

Native splits are calibration, development, test and supplementary; the JevBench config contains test. These examples load existing outputs and do not run an evaluation. For exact artifact downloads and file hashes, use the manifest at the same revision.

Format

Field family	Preservation rules
Native identity	(panel, record_id, question_id); group_id retains bootstrap grouping across models
Depth and options	Native logits use [depth][option]; depth 1 is index 0; option_indices retains encoded order
Targets	Numeric target_scalar / target_vector are preserved; textual targets have an opaque target_identity only
Categories	Source, task, type, variant, tier and difficulty retain compact category values
JevBench	Stored probability order, predicted index, confidence, validity and correctness are preserved
Missing values	Optional fields remain null; invalid observations retain null or empty probabilities

Identity and numerical fidelity

SHA-256 identities are deterministic and domain-separated by identity kind and public panel. These are pseudonymous identifiers, not a claim of irreversible anonymization. Native token counts are model-specific. Textual native option labels were not stored in these inputs and cannot be independently reconstructed here.

JevBench supplies one stored identifier per task and no separate group metadata, so group_id equals record_id. Option identities hash original option keys scoped to a task. Vector order and predicted index are checked against stored key order across depths and models. Stored correct flags are copied, not recomputed. No logarithm, clipping, renormalization or inferred gold label is introduced.

Included native rows were originally read in fp32 with TF32 disabled and temperature 1. Selected JevBench reports declare fp32 and no additional per-depth temperature transform; original checkpoint temperature behavior is separate. Calibration files are existing artifacts. Latency fits were measured separately in bf16, batch one, on an H200; the code release describes that measurement context.

Provenance and limits

Coverage & input/output hashes · File manifest

Only pinned, already stored material is included. Clean aggregate, calibration and latency artifacts are copied byte-for-byte. Packaging performs structural checks and text removal; it does not run models, fit temperatures, tune stoppers or calculate scores.

Locked test outputs are included under a narrow release authorization. This does not permit tuning on them or reopening frozen evaluation records. Ordinary evaluation and export guards remain in effect.

This package does not establish fresh-environment GPU parity, complete source-data reconstruction or full row-based figure reproduction. Only the three headline seed-0 adapters are separately staged; results also cover other study checkpoints. Source rights remain a separate gate; no blanket source-data license is asserted.

Explore Lev

Resource	Start here
LevBench	Development and test data across five tiers
Lev-train-v2	S31 / S60 / S120 mixtures, replay and composition recipes
Lev-results	Stored outputs, calibration and latency evidence
Lev-1.4B · Lev-2.6B · Lev-3B	Headline seed-0 adapters and decision heads

Lev: Thinking Fast and Deep with Looped Decision Models · Paper on alphaXiv · PDF · Project page · Code

Citation

@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}
}