AutoIndex Optimizes Representation Programs Through Iterative Code Updates
Reactions from ranked influencers
6 posts
๐ AutoIndex makes the Representation Program itself the optimization target. It: 1. executes the current program 2. builds the resulting index 3. analyzes retrieval failures 4. synthesizes candidate updates 5. evaluates them 6. retains verified improvements Then it repeats. This is not prompt tuning. The output is persistent executable software.
๐ฌ One learned program found that repeated LaTeX markup was acting as retrieval noise. AutoIndex synthesized a narrow, threshold-gated transformation that activates only on heavily affected documents. The resulting representation shifted the top results toward relevant evidence and improved both Recall@100 and nDCG@10.
๐ ๏ธ AutoIndex does not converge on one universal preprocessing recipe. It learns corpus-specific programs that can slice, normalize, enrich, reweight, and reorganize. The goal is not to find the best chunk size. The goal is to learn the ๐ฅ๐ฒ๐ฝ๐ฟ๐ฒ๐๐ฒ๐ป๐๐ฎ๐๐ถ๐ผ๐ป ๐ฃ๐ฟ๐ผ๐ด๐ฟ๐ฎ๐บ that exposes the most useful evidence to the retriever.
โณ Scaling the number of iterations matters. A single-iteration variant improves only 3 of 8 tasks. The full procedure produces more consistent gains through repeated analysis, synthesis, execution, and selection. Effective Representation Programs often emerge over multiple rounds of search, not from one-shot code generation.
๐ To isolate the effect of learned Representation Programs, we keep the rest of the retrieval system fixed, including the BM25 retriever, the ranking function, and the indexing backend. Only the program mapping documents to indexed units changes. Across CRUMB: โข ๐ฅ๐ฒ๐ฐ๐ฎ๐น๐น@๐ญ๐ฌ๐ฌ: +8.4% โข ๐ป๐๐๐@๐ญ๐ฌ: +8.3% This is a big improvement given no retriever fine-tuning, embedding updates, or online feedback!
๐ Representation Programs are only one layer. The same paradigm could extend to: โข ๐ฅ๐ฎ๐ป๐ธ๐ถ๐ป๐ด ๐ฃ๐ฟ๐ผ๐ด๐ฟ๐ฎ๐บ๐, โข ๐ฆ๐ฒ๐ฎ๐ฟ๐ฐ๐ต ๐๐ป๐ณ๐ฒ๐ฟ๐ฒ๐ป๐ฐ๐ฒ ๐ฃ๐ฟ๐ผ๐ด๐ฟ๐ฎ๐บ๐, โข ๐๐ด๐ฒ๐ป๐ ๐๐ฎ๐ฟ๐ป๐ฒ๐๐ ๐ฃ๐ฟ๐ผ๐ด๐ฟ๐ฎ๐บ๐, โข ๐ ๐ฒ๐บ๐ผ๐ฟ๐ ๐ฃ๐ฟ๐ผ๐ด๐ฟ๐ฎ๐บ๐, โข and even ๐ง๐ฟ๐ฎ๐ถ๐ป๐ถ๐ป๐ด ๐ฃ๐ฟ๐ผ๐ด๐ฟ๐ฎ๐บ๐ We have spent years optimizing model weights, and more recently prompts. The next frontier may be optimizing the executable programs that shape how AI systems represent information, retrieve, reason, and act.