analiza aiqbench
This research investigates latent communication protocols between independently trained language model cells, revealing that strict zero-shot transfer fails due to incompatible "packet languages" and that globally visible models often induce negative transfer rather than aiding adaptation. The study provides critical empirical evidence on interoperability limits in modular AI architectures, demonstrating that reinitializing specific interface components significantly improves performance over relying on inherited priors from high-performing global models.
Portable Semantics, Private Dialects: Reuse and Negative Transfer in Latent Communication Between Language-Model Cells
arXiv:2609.11365v1 Announce Type: cross Abstract: In shared-genome language-model societies, restricted evidence visibility favors reusable, value-indexed latent packet interfaces, whereas the sole high-performing globally visible model in the parent study learned an episode-entangled code. This companion study asks whether independently trained societies share one packet language, where strict zero-shot transfer fails, and whether inherited interface state helps or harms later learning. First, a leakage-controlled causal interoperability audit over all 30 ordered pairs of six independently trained restricted societies -- under sealed held-out structure and a preregistered raw/orthogonal/linear/nonlinear alignment ladder -- shows the six semantically similar interfaces do not form one raw language: one same-initialization pair is exactly interoperable in both directions, a second shows asymmetric partial compatibility, and all 26 cross-initialization directions fail every frozen alignment rung. Second, within the tested decomposition and a single sealed source formulation, a source-span control localizes strict zero-shot failure to interpretation and execution of the new operator instructions. Third, in a matched adaptation factorial, the globally trained communication interface acts as a severe negative-transfer prior: reinitializing only the packet reader, writer, and mouth raises final depth-three accuracy from 0.169 to 0.857. Fourth, across two restricted checkpoints and two independently frozen target streams each, inherited interfaces never exceeded fresh-interface controls by the preregistered 0.10 margin. All primary conclusions are bounded to a near-transfer 17-state setting; the negative-transfer factorial concerns one globally visible parent-cohort checkpoint, while an appendix adds a post hoc tagged-global twin case study.
Abstract:In shared-genome language-model societies, restricted evidence visibility favors reusable, value-indexed latent packet interfaces, whereas the sole high-performing globally visible model in the parent study learned an episode-entangled code. This companion study asks whether independently trained societies share one packet language, where strict zero-shot transfer fails, and whether inherited interface state helps or harms later learning. First, a leakage-controlled causal interoperability audit over all 30 ordered pairs of six independently trained restricted societies -- under sealed held-out structure and a preregistered raw/orthogonal/linear/nonlinear alignment ladder -- shows the six semantically similar interfaces do not form one raw language: one same-initialization pair is exactly interoperable in both directions, a second shows asymmetric partial compatibility, and all 26 cross-initialization directions fail every frozen alignment rung. Second, within the tested decomposition and a single sealed source formulation, a source-span control localizes strict zero-shot failure to interpretation and execution of the new operator instructions. Third, in a matched adaptation factorial, the globally trained communication interface acts as a severe negative-transfer prior: reinitializing only the packet reader, writer, and mouth raises final depth-three accuracy from 0.169 to 0.857. Fourth, across two restricted checkpoints and two independently frozen target streams each, inherited interfaces never exceeded fresh-interface controls by the preregistered 0.10 margin. All primary conclusions are bounded to a near-transfer 17-state setting; the negative-transfer factorial concerns one globally visible parent-cohort checkpoint, while an appendix adds a post hoc tagged-global twin case study.
17 pages, 1 figure, 5 tables. Companion to arXiv:2608.20054. Code and evaluation records: this https URL ; checkpoints and fitted alignment maps: this https URL
Artificial Intelligence (cs.AI)
arXiv:2609.11365 [cs.AI]
(or arXiv:2609.11365v1 [cs.AI] for this version)
https://doi.org/10.48550/arXiv.2609.11365 arXiv-issued DOI via DataCite (pending registration)
Text integral sub licența deschisă a publicației · 277 cuvinte
Subiecte conexe
[AINews] Reality Checks on AI News (Yegge shuts down Gas Town, Databricks’ +60% Astra cost)
Latent Space · 17 sept. 2026
[AINews] OpenAI shuts off Cursor
Latent Space · 29 aug. 2026
[AINews] The Future of Latent Space
Latent Space · 25 sept. 2026
Enabling DeepSeek-V4-Flash Training on AMD Instinct MI355X GPUs with Primus
AMD · 3 sept. 2026