MARC: modular representation compression¶

简体中文

Fidelity: core mechanism. Runs the defining mechanism verifiable on public data. Private data, production models, and serving remain outside scope.

Paper information¶

Field	Details
Paper	arXiv 2604.18146
Organization	Shanghai Jiao Tong University
First public release	2026-04-20 (arXiv v1)
Authors' code	No public author code was found when checked on 2026-09-05.
Adapter / method	`marc`
Local reproduction code	`src/auto_research/reproductions/marc/`

Original paper¶

Motivation and method¶

After recommendation compression, intermediate LLM layers can outperform the final layer (MRA). MARC separates compression from task adaptation and uses information constraints to decouple the proxy task from recommendation. It retains task-relevant information while reducing dimension.

flowchart LR
  A["Public behavior and candidates"] --> B["MARC core mechanism"]
  B --> C["Matched-budget method output"]
  A --> D["Common direct baseline"]
  C --> E["Full-catalog ranking and structural diagnostics"]
  D --> E

Key figure from the paper¶

Figure 5: Overview of the paper's method and key components. From the original paper; copyright remains with the authors.

Core equations¶

\[ z=P_k\big(h_{mid}-\Pi_{nuisance}h_{mid}\big),\qquad s(i)=\langle z_u,z_i\rangle. \]

Paper results¶

A commercial search-advertising online A/B test reports eCPM +2.82%. This is paper evidence, not a local public-data outcome.

Local reproduction¶

Local comparison: Under the same full-catalog MovieLens protocol, baseline NDCG@10 is 0.05401 and treatment NDCG@10 is 0.04693, -13.11% relative. The local proxy and production task differ, so there is no extrapolation to online lift.

See the three-seed artifact for complete results, means, standard deviations, and 95% confidence intervals.

auto-research reproduce --paper marc --dataset-dir data --seeds 42,43,44

Reproduction boundaries¶

The local run uses a public MovieLens-1M subset and auditable proxies to check the core computation only. It does not reproduce private logs, the production foundation model, online traffic assignment, or serving. There is no claim to reproduce absolute paper metrics or online gains.