pith:HH6MRAPL
Memory-Efficient Looped Transformer: Decoupling Compute from Memory in Looped Language Models
MELT shares one KV cache per layer across all reasoning loops and updates it with a learnable gate to keep memory constant.
arxiv:2605.07721 v2 · 2026-05-08 · cs.CL · cs.AI · cs.LG
Add to your LaTeX paper
\usepackage{pith}
\pithnumber{HH6MRAPLHICKZBGPUDT3AWFQEE}
Prints a linked badge after your title and injects PDF metadata. Compiles on arXiv. Learn more · Embed verified badge
Record completeness
Claims
MELT achieves constant-memory iterative reasoning without sacrificing LoopLM performance, using only a lightweight post-training procedure.
The learnable gating mechanism and the two-phase chunk-wise training (interpolated transition followed by attention-aligned distillation) are sufficient to preserve the original reasoning capabilities of the LoopLM starting model without degradation.
MELT decouples reasoning depth from memory in looped LLMs by sharing a single gated KV cache per layer and using two-phase chunk-wise distillation from Ouro, delivering constant memory use while matching or beating standard LLM performance.
Formal links
Receipt and verification
| First computed | 2026-05-20T01:05:15.801090Z |
|---|---|
| Builder | pith-number-builder-2026-05-17-v1 |
| Signature | Pith Ed25519
(pith-v1-2026-05) · public key |
| Schema | pith-number/v1.0 |
Canonical hash
39fcc881eb3a04ac84cfa0e7b058b0211bfd53c6382cb41edc465cc19e5705ba
Aliases
· · · · ·Agent API
Verify this Pith Number yourself
curl -sH 'Accept: application/ld+json' https://pith.science/pith/HH6MRAPLHICKZBGPUDT3AWFQEE \
| jq -c '.canonical_record' \
| python3 -c "import sys,json,hashlib; b=json.dumps(json.loads(sys.stdin.read()), sort_keys=True, separators=(',',':'), ensure_ascii=False).encode(); print(hashlib.sha256(b).hexdigest())"
# expect: 39fcc881eb3a04ac84cfa0e7b058b0211bfd53c6382cb41edc465cc19e5705ba
Canonical record JSON
{
"metadata": {
"abstract_canon_sha256": "7d4a91d69446c6fc830e95b77703f53d4f713f6c9db7e29ac824898a7289f052",
"cross_cats_sorted": [
"cs.AI",
"cs.LG"
],
"license": "http://creativecommons.org/licenses/by-nc-sa/4.0/",
"primary_cat": "cs.CL",
"submitted_at": "2026-05-08T13:25:27Z",
"title_canon_sha256": "96809772eaa38e3a400b81c44488c05ef57c87596756afeaa488b347b1a97941"
},
"schema_version": "1.0",
"source": {
"id": "2605.07721",
"kind": "arxiv",
"version": 2
}
}