UQO Working Paper No. 30· v1.1· November 2026· 97 pages
Beyond Text: Cache Compression and Knowledge Transfer in Multi-Agent Systems
A Reproducible Study of Fidelity, Communication Cost, and Repeated Handoffs with Same-Family Caches on Two 3B Checkpoints
Lossless handoffs between agents are equivalent within one point and differ only in cost, while every lossy choice depends on the checkpoint and the regime: packed 4-bit quantization moves Llama-3.2-3B by −1.7 points and Qwen2.5-3B by −83.2.
Twenty-two text, evidence and KV-cache representations are scored on the same questions by the same receivers over 13,280 trials, with bytes measured on serialized payloads and wall-clock latencies. The conclusions hold for one checkpoint per family, same-family caches, one sender budget and one eviction ratio; cross-family cache transfer was audited, not executed, and every effect is contextual rather than learned.
Simon-Pierre Boucher — contact@spboucher.ai
within 1 point
Quality of the lossless representations (full text, its zstd encoding, the native cache, the reloaded cache and the 8-bit packed cache) on the pooled test set, an equivalence with a pre-stated margin; they differ only in cost
−1.7 vs −83.2 points
Effect of packed per-token 4-bit quantization on Llama-3.2-3B and on Qwen2.5-3B; the collapse is consistent with, but not isolated to, key rounding along the head dimension
−27.0 / −16.3 points
Structured evidence minus prose summary at an equal 512-token sender budget, and once the structured sender may deliver at least the summary's bytes; a values-only schema is worth +9.5 points at equal tokens
3.49 vs 0.93
Ratio of receiver prefill time to link time per token on an M3 Ultra receiver, where moving an 8-bit cache over 1 GbE beats re-reading the text at every length, and on an M5 Max receiver, where it loses at 2k–8k tokens
0 in 960
Relay messages containing an unsupported entity; relays lose facts at every hop where the budget binds, mis-attribute 0.0–3.3 % of the facts and are nearly lossless once messages fall below the cap
89.6 % → −17.6 points
The same-family ridge KV mapper reconstructs 89.6 % of the receiver's value energy yet transfers no content, scoring 17.6 points below the receiver's own summary
−0.6 points, −46 %
Router quality relative to full text and amortized latency saving when questions are unknown and the document is reused; the single static action "always 8-bit cache" does the same
Abstract
Agents built on language models hand information to one another as text, summaries, structured evidence, shared-memory references or—when they share a checkpoint—attention state (KV cache). We ask under which task, model, cache, network and reuse conditions each representation preserves task-relevant information at the lowest measured end-to-end cost. Twenty-two conditions are measured on one checkpoint per family (Qwen2.5-3B-Instruct, Llama-3.2-3B-Instruct, bf16, MLX; same-family caches only, cross-family transfer audited, not executed) over private-fact ledgers, LongBench QA and executable tasks (13,280 trials, 160 test documents, frozen protocol, six Holm-corrected registered tests), plus two-node transfer over 1 GbE, relays of 1–8 handoffs, fan-in, a same-family learned KV mapper, hosted receivers and a development-fitted router. Lossless representations are equivalent within one point and differ only in cost; on the measured link an 8-bit cache beats re-reading the text at every length on an M3 Ultra receiver and loses at 2k–8k tokens on an M5 Max receiver. Packed per-token 4-bit quantization moves Llama by -1.7 points and Qwen by -83.2, a collapse consistent with key rounding along the head dimension. At an equal 512-token sender budget, structured evidence minus prose summary is -27.0 points; once the structured sender may deliver at least the summary's bytes it is -16.3. Relays lose facts at every hop where the budget binds and are nearly lossless once messages fall below the cap, with no unsupported entity in 960 messages. The mapper transfers no content. The router scores -0.6 points relative to full text at 46 % lower amortized latency, as does the static action "always 8-bit cache". Code, records, protocol and a minimal reproduction are released.
JEL: C63, C88, D83, L86, O33
Key Results
A taxonomy with measured coverage
Five executed representation classes (text compression, KV compression, cache reuse, same-checkpoint and same-family cache transfer) and the audited but unexecuted ones, each with its question-access regime, compatibility requirement and composability across sources.
A common harness with a validated cache layer
Text and cache representations of the same document are scored on the same questions by the same receiver, with identity round-trip, prefix-equivalence and manifest-checked loading tests, deterministic scorers, evaluator-isolated answer keys and document-level cluster-bootstrap inference with one registered Holm family.
Quality and resource trade-off maps
Twenty-two conditions on two checkpoints and three task families under question-blind and question-aware regimes, with measured bytes, latency decompositions and peak memory, and a comparison of prose summaries with structured evidence that is byte-matched as well as token-matched.
Repeated handoffs and fan-in
Relays of 1 to 8 transfers followed with an evaluator-only fact ledger at the level of (entity, attribute, value) triples under a non-binding and a binding per-hop budget, and a fan-in study with conflicting evidence.
A break-even rule for two-node cache transfer
Measurements on a 1 GbE LAN against remote text prefill on two receivers of different prefill speed turn the break-even into a stated rule: the ratio of receiver prefill time to link time per token.
Negative results on the mapper and the router
A learned KV mapper from Qwen2.5-1.5B to Qwen2.5-3B is evaluated against an identity-transfer negative control and transfers no content; a development-fitted router is not better than the best static action.
Quality control
This paper went through a multi-stage quality control: 7 quality-control reports and 114 points checked. The version presented on this page is the revised manuscript. Scores are on a 0–100 scale and are a diagnostic of the manuscript, not an acceptance probability.
Score after revision
73.4/100
64.9 before quality control · +8.5
Quality control dated November 9, 2026
- Quality-control decision
- major revision
- Scientific integrity
- PASS
- Reproducibility
- R1reproduced from documented analysis data
- Maturity
- Bone revision cycle from submission
- Quality-control reports
- 7
- Points checked
- 114
Scores by dimension
Before quality controlAfter revision
- Originality & contribution68.0 → 70.0
- Scientific rigor62.4 → 74.0
- Identification & methodology60.7 → 68.0
- Data & reproducibility79.3 → 87.0
- Robustness & validation52.1 → 66.0
- Writing69.3 → 75.0
- Figures & tables60.7 → 76.0
- Literature positioning60.7 → 78.0
Revised manuscript97 pages12 figures36 tables6 equations62 references43,370 words
What quality control changed
Quality control left the thesis in place and narrowed four of its sentences. The summary-versus-structured comparison is now reported both token-matched and byte-matched, the two-node break-even is stated receiver by receiver, the router's advantage is shown to equal a static choice, and the relay result is restated as loss at every hop where the budget binds. The scope is now in the subtitle: same-family caches on two 3B checkpoints.
- "Moving an 8-bit cache beats re-reading the text at every length" → true on an M3 Ultra receiver (prefill-to-link ratio 3.49), false at 2k–8k tokens on an M5 Max receiver (ratio 0.93)
- "Prose summaries beat structured evidence" at a 512-token budget → −27.0 points at equal tokens and −16.3 points once the structured sender may deliver at least the summary's bytes
- "Relays are nearly lossless after the first hop" → relays lose facts at every hop where the budget binds (−8.5 points at hop 2 under a schedule that tightens there) and are nearly lossless once messages fall below the cap
- "The router halves amortized latency at equal quality" → a refitted router scores −0.6 points relative to full text at 46 % lower latency, the same as the static action "always 8-bit cache"; the claim for the question-aware regime is withdrawn
- "4-bit quantization is free for Llama and fatal for Qwen" → −1.7 and −83.2 points on the named checkpoints, with the collapse described as consistent with key rounding along the head dimension, not isolated to it
- Holm correction applied per results file → one registered family of six tests, with a ledger of the ten hypotheses stating which are rejected on their own terms
Limitations stated in the paper
- The lossy conclusions are statements about two named 3B checkpoints under one quantization scheme, one eviction ratio and one 512-token sender budget; reading the 4-bit result as a property of a model family would over-read it.
- Cross-family cache transfer was not executed, and the mapper's negative result is not reconciled by experiment with a published positive closed-form result.
- Every effect is contextual: it disappears when the message or cache is removed, and no receiver acquired knowledge.
- Byte, latency and memory figures are specific to one runtime version and four machines; 881 timing trials overlapped another job on the same GPU and are flagged, not excluded.
Data
Synthetic private-fact ledgers
Generated documents whose facts are new by construction, used to test communication dependence without pre-training contamination; answer keys are held by the evaluator only. The templates are narrower than real documents.
Public long-document QA and executable tasks
LongBench subsets QASPER and 2WikiMultihopQA scored with the official F1, and small executable tasks scored by hidden tests, simplified on development data before any test item was run.
Per-attempt trial records
13,280 core trials on 160 test documents, a development campaign of 8,580 trials over 166 development documents for the refitted router, and relay, fan-in, two-node, mapper and hosted runs; every attempt is recorded with measured bytes, latency spans and peak memory.
Methodology
Twenty-two handoff conditions
Truncation, extractive selection, prose summaries, structured evidence, metered retrieval, a lossless codec, LLMLingua-2, native and serialized caches, packed 8-bit and 4-bit quantization and eviction rules, on Qwen2.5-3B-Instruct and Llama-3.2-3B-Instruct in the MLX runtime.
Frozen protocol and one confirmatory family
Protocol, hypotheses and analysis plan were frozen before any test-split item was evaluated; six registered tests are Holm-corrected as one family, and a ledger gives the registered wording, the test executed and the outcome of each of ten hypotheses. Where the executed test departs from the registered one, both are reported.
Receiver-side accounting
Byte counts include manifests and quantization metadata, latencies include serialization, loading and fresh-query processing, and reusable question-blind messages are reported separately from question-aware ones.
Two-node transfer
Cache transfer over 1 GbE links is compared with remote text prefill on an M3 Ultra and an M5 Max receiver; the two receivers' grids are separated by ten days.
Deterministic scoring
Exact, numeric, yes/no, abstention, F1 and hidden-test scoring only; no language-model judge and no human annotation. A strict exact-match re-scoring is reported as a check.
Reproducibility
- The repository contains the harness, a hashed lock file, data manifests and generators, the frozen protocol with its hash and dated amendments, every per-attempt record, the mapper checkpoints, the router policies, and generated tables and figures with a provenance index and a claim ledger.
- A clean clone with a fresh environment regenerates the committed results and tables byte-identically, verified on a machine that had not run any experiment.
- A minimal profile re-runs 64 trials: 64 of 64 outputs were byte-identical on the original machine; on a second machine 55 of 64 outputs were byte-identical and 62 of 64 scores identical.
- Model checkpoints are not redistributed; their identifiers, revisions and weight hashes are recorded. Hosted-API trials require provider credentials and reproduce only up to provider nondeterminism.
- The source code has no licence file yet; the data licences are recorded.
Cite this paper
Boucher, S.-P. (2026). Beyond Text: Cache Compression and Knowledge Transfer in Multi-Agent Systems. UQO Working Paper No. 30, Université du Québec en Outaouais.
@techreport{boucher2026wp30,
author = {Simon-Pierre Boucher},
title = {Beyond Text: Cache Compression and Knowledge Transfer in Multi-Agent Systems},
institution = {Université du Québec en Outaouais, Département des sciences administratives},
number = {30},
type = {UQO Working Paper},
year = {2026},
month = {nov},
url = {https://www.spboucher.ai/research/wp30},
}More in AI Systems