Better answers from a mostly transferred context.
Our Qwen3 4B → Ministral 3 14B experiment asks whether a receiver can answer using a translated cache from a different model family. The original HeteroFold implementation already provided that handoff. This round tested a smaller question: can a little receiver-side work repair some of the information lost in translation?
On 128 locked confirmation questions—64 Qasper articles and 64 HotpotQA examples—equal-task macro answer F1 increased from 53.59 to 58.63. The 5.05-point improvement has a paired, task-stratified 95% bootstrap interval of 1.37 to 9.13 points. These were held-out questions from the training splits, separate from the 64 development questions used to choose the repair.
The largest observed improvement came from Qasper: 32.81 → 41.19 F1. HotpotQA moved from 74.36 → 76.08, but its task-specific gain remains uncertain. The aggregate result should not be read as equally strong evidence for every task.
Keep the earlier memory. Reread the most recent passage.
The selected method asks Ministral to reprocess the final 256 context tokens before answering. It still attends to the translated cache for the earlier context. A useful analogy is receiving someone else's notes and rereading the last passage before responding.
This is a hybrid repair. The replayed tail is conditioned on an approximate earlier cache, so even that tail is not equivalent to a fully native pass over the document. Its additional receiver computation is included in the timing measurements.
We compared six fixed candidates on development questions: unchanged transfer, native first-32 or first-128 tokens, replay-last-64 or replay-last-256 tokens, and first-32 plus last-64. Replay-256 won on development macro F1. Its selection was frozen before confirmation; the checkpoint, prompts, sampling rules, and task caps stayed fixed.
The extra reading fit inside a faster handoff.
On a separate 32,768-token timing fixture, the original handoff took 599.63 ms. The optimized path including replay took 505.43 ms, a 15.71% reduction. At 4K and 16K, the reductions were 6.20% and 12.71%.
The savings came from reusing producer token offsets, source-layer gathers, and rotary factors. At 32K, tokenization and alignment fell from 234.84 to 124.42 ms, while mapping and cache construction fell from 280.57 to 235.65 ms. Receiver work increased because of the reread, but the net handoff still saved about 94 ms.
Timing used FlashAttention-2, batch two, and two H100 80 GB GPUs connected by NVLink, with three repetitions per configuration. It measures handoff from an existing source cache through first logits. It excludes creating that source cache and generating the full answer. Quality and latency used different inputs, so this is not evidence of the same answer quality at every plotted context length.
The timed backend gives a more cautious quality result.
We also repeated quality evaluation on FlashAttention-2, the backend used for timing. On the same 128 cases, macro F1 moved from 53.55 to 57.02. Its 3.47-point gain had a 95% interval of −0.49 to 7.66, crossing zero.
That repeat does not yet establish a reliable overall quality gain on the timed backend. It also does not add independent questions to the sample. Backend changes altered generated answers despite fixed case-derived sampling seeds; we retained both outcomes without reselecting the candidate.
The primary SDPA quality improvement and the separate FA2 latency reduction are supported individually. A strong claim of a simultaneous quality-and-speed win on FA2 needs a larger independent confirmation.
A measured improvement, with a native-quality gap still to close.
Native receiver macro F1 was 63.16 on SDPA, leaving a 4.53-point gap after repair. This experiment does not establish native-quality parity, superiority over the published paper, or a production serving speedup. Fresh-request comparisons with native inference remain provisional while native timing is corrected.
What it does establish is useful: a small, selected reread can improve our prior cross-family transfer on held-out questions. The next step is to confirm that improvement on the serving backend and evaluate the complete request on the same realistic workloads.
Measurement record: October 11 transfer-iteration v002; later unmeasured candidates are excluded. Read the measurement summary and provenance.