HARNESS-LM: Three-stage retrieval with a strong document tower and lightweight query tower¶

All three local stages execute, without rebuilding the document index.

Paper information¶

Field	Details
Paper	Paper or original source
Institution	Microsoft AI / Bing Ads.
First public release	2026-05-22 (arXiv v1)
Author code	No official author code linked by the paper as checked on 2026-08-09.
Local adapter or method key	`harness-lm`
Local reproduction code	`src/auto_research/reproductions/harness_lm/`

Paper method and evidence¶

Sponsored Search can encode documents offline but needs a cheap online query encoder. Train a strong teacher, align a small query encoder to its embedding space via L2, then refine with supervised contrastive learning against the frozen teacher index.

flowchart LR
  A[Train teacher] --> B[L2 query alignment]
  B[L2 query alignment] --> C[Freeze teacher document index]
  C[Freeze teacher document index] --> D[Refine query via InfoNCE]

Original paper figure¶

Figure shows the three-stage training pipeline. From the original paper; copyright remains with the authors.

Core equation¶

\[ s(q,d)=\langle f_Q(q),f_D(d)\rangle,\qquad \mathcal L_{align}=\sum_i\|f_Q^S(q_i)-f_Q^T(q_i)\|_2^2. \]

For refinement, freeze \(f_D^T\) and update only \(f_Q^S\) with InfoNCE.

Reported results¶

Bing Ads deployment reports revenue +1.0%, impressions +0.6%, clicks +0.4%, serving about 190 million ads.

Local reproduction¶

auto-research reproduce --paper harness-lm --seed 42

Versus a small symmetric content retriever, Hit@10 -23.08%, NDCG@10 -28.05%. Alignment MSE 0.02375→0.02115, contrastive loss 7.1900→5.8921; optimization converges without test transfer. Stable metrics. Collaborative SVD+content proxies Qwen3-Embedding. No 4B/0.6B training, GPT expansion or CPU pruning frontier.