LatentPort: Beyond KV Cache - Cross-Model Transfer of Recurrent Memory in Hybrid Language Models: A 4B-to-9B Hybrid-State Handoff Without Target Prefix Replay
cs.CL · cs.AI
Quick verdict
Claims first cross-model handoff of recurrent state between hybrid LLMs; rigorous but single-pair, teacher-forced, near-native gate failed, generation untested.
Summary
The paper reports the first, to the authors' knowledge, cross-model handoff of persistent recurrent inference state between differently sized hybrid language models: Qwen3.5-4B-Base exports its Gated DeltaNet recurrent matrices and convolution history, alongside translated attention KV, into Qwen3.5-9B-Base, which continues from the 4K-token prefix without replaying it. On 64 PG19 documents, translated KV alone leaves NLL at 3.115 versus native 9B's 2.161; adding the GDN package lowers NLL by 0.7473 nats/token (95% CI [0.6921, 0.8047]), improving all 64 documents. An eight-cell factorial selects translating KV while directly reusing GDN state, and a 434,176-parameter rank-4 behavioral correction reduces excess NLL to 0.076 above native 9B, beating continued 4B inference by 0.0521 nats/token (CI [−0.0843, −0.0185]) with zero historical prefix tokens. The stricter near-native gate failed (top-1 agreement 0.864; NCR 0.918), and the 16K branch was not run. Evidence covers one direction, one geometry-matched Base pair, and teacher-forced 4K continuation; free generation remains untested. Strengths include preregistered gates, wrong-donor shuffle controls, and unusually candid limitation reporting; the main weaknesses are single-pair generality, untested generation, and confounds in the learned-versus-direct GDN map comparison.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 3
● low 1
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Persistent GDN recurrent and convolution state, plus translated attention KV, transfers usefully from Qwen3.5-4B-Base to Qwen3.5-9B-Base without target prefix replay — claimed as the first such cross-model handoff between differently sized hybrid LLMs.
- Adding the GDN persistent-state package to fixed translated KV lowers teacher-forced NLL by 0.7473 nats/token (95% CI [0.6921, 0.8047]), improving all 64 PG19 documents.
- Direct reuse of GDN recurrent and convolution state outperforms the tested learned translation maps, consistent with partial functional compatibility of persistent-state coordinates.
- A 434,176-parameter behavioral correction brings the handoff to 0.076 nats/token above native 9B and beats continued 4B inference by 0.0521 nats/token (CI [−0.0843, −0.0185]) while processing zero historical prefix tokens.
- Wrong-donor shuffle controls show the transferred state carries document-specific information, since shuffled state performs worse than an empty 9B.
Strengths
Preregistered protocol with integrity hashes
Splits frozen via salted SHA-256 ordering (App. B), seeds/revisions pinned (App. A), FULL_STATE vs NEAR_NATIVE gates defined before verdicts (App. D), raw-evidence hashes published (App. I); the failing gate is honestly recorded rather than relabeled.
Clean primary intervention with strong controls
Holding translated KV fixed, the GDN package lowers NLL by 0.7473 nats/token (CI [0.6921, 0.8047]) with 64/64 PG19 documents improving (Fig. 3); wrong-donor shuffled state (3.3162) is worse than empty 9B (3.2457, Table 3), demonstrating document-specific transfer.
Candid treatment of negative and confounded results
Subtle runtime confounds caught and controlled
Factorial discipline with preregistered materiality
Weaknesses
Single-pair, single-direction external validity
Teacher-forced endpoint only
Handoff is not equivalent to native processing
Package effect cannot be decomposed
Headline correction size understates system scale
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
LatentPort asks whether a hybrid LLM's live inference state — attention KV plus Gated DeltaNet (GDN) recurrent matrices and convolution history — can be handed from a 4B model to its 9B sibling without the receiver re-reading the prefix. The authors claim the first such demonstration across differently sized hybrid models without target prefix replay (Sections 1, 5.6), qualified with 'to our knowledge' and a structured related-work boundary covering KV transfer [2,5,14], hidden-state messaging [12], external memory [4], activation injection [13], and same-model reuse [7,8,10,20]. The scope is deliberately narrow and repeatedly restated: one direction (4B→9B), one geometry-matched Base sibling pair, 4K prefixes, 64 teacher-forced targets per document. The abstract itself concedes the near-native gate failed, the 16K branch was not run, and free-generation equivalence is unproven — an unusual degree of self-limitation the body sustains.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What does the LatentPort paper demonstrate?
It shows that a 4B hybrid language model (Qwen3.5-4B-Base) can hand its internal inference state — attention KV cache plus Gated DeltaNet recurrent matrices and convolution history — to its 9B sibling, which then continues the text without re-reading the original 4K-token context. The 9B's predictions after the handoff beat continued 4B inference and come within 0.076 nats/token of native 9B processing, but never fully match it.
Did the transferred state match native 9B performance?
No. Even after a 434,176-parameter correction, continuation loss stayed 0.076 nats/token above native 9B, top-1 agreement with native predictions was 0.864, and the paper's stricter 'near-native' quality gate failed on all three thresholds. Consequently the conditional 16K-context branch of the experiment was never run.
Does this work for any pair of models?
No. Evidence covers exactly one direction (4B→9B) of one geometry-matched Qwen3.5 Base sibling pair whose recurrent-state shapes coincide by design, with 4K prefixes and teacher-forced continuations. Other families, mismatched geometry, Instruct models, and free-form generation are all untested, and the authors note shared pretraining may explain the compatibility.
Is code or data available for this paper?
The text describes extensive frozen evidence archives, SHA-256 hashes of raw results, deterministic verification scripts, and pinned model revisions and dataset shards bundled with the project, but it names no public code repository URL, so independent re-execution currently depends on artifacts supplied by the authors.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figure panels appear only as extracted axis-label and caption fragments; the rendered plots could not be visually verified, though all plotted values are cross-checkable against Tables 3, 4, 7, and 8.
- The reference list consists largely of 2026 arXiv preprints whose content could not be verified from the provided text, so the novelty-boundary assessment relies on the paper's own related-work characterization.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.