The Laws of Context Allocation: Causal Measurement and Closed-Loop Orchestration in Generative Search
Abstract
A causal measurement probe and iterative submodular scheduling overcome evidence-utilization and context-allocation bottlenecks in retrieval-augmented generation, enabling sequential feedback-driven portfolio synthesis.
As Retrieval-Augmented Generation (RAG) shifts toward diverse portfolio generation, it is stymied by two critical bottlenecks: flawed measurement of evidence utilization, and suboptimal context budget allocation. We resolve both sequentially. To resolve measurement, we expose a pervasive ``diagnostic illusion'': standard relevance proxies fail catastrophically on hard negatives. We replace them with an efficient causal leave-one-out probe that accurately isolates generative reliance and formally calibrates the structural dilution of LLM attention. To resolve allocation, we deploy this causal probe in a deconfounded factorial grid. We prove that the prevailing strategy of monolithic context widening is an architectural trap penalized by relevance decay. Instead, allocating compute iteratively across multiple sequential generations drives transformative portfolio recall gains of 16.7--20.5 absolute percentage points, scaling robustly up to 32B models. Finally, we unify these solutions into a deployable closed-loop submodular scheduler. Augmented by an attribution-steered contrastive decoder to override LLM attention inertia, our architecture systematically forces fresh evidence integration. By dominating classical open-loop baselines, we establish sequential, feedback-driven orchestration as the definitive paradigm for generative search. Our code, data, and causal measurement instruments are available at https://github.com/PeiYangLiu/ascp.
Community
We found that the metrics the RAG community uses to check "did the model actually
use this document?" are mostly measuring the wrong thing.
On standard evaluation pools — where the negatives are passages retrieved for
unrelated queries — BM25 and query–document cosine look essentially perfect at
identifying the evidence a generator relied on (AUC 0.99). Swap the padding for
same-query hard negatives that are topically dense but entail no answer, and they
fall to chance (0.57). A causal leave-one-out probe barely moves (0.85 → 0.85).
The apparent reliability was an artifact of the distractors being easy.
That matters because these proxies are what people use to decide what goes into
the context. Once we measure utilization causally instead, a clear allocation law
shows up: for a fixed budget of k×T evidence slots, spending it on several narrow
sequential contexts beats one wide context by 16.8–20.5 absolute points of
portfolio recall, and the gap holds up to 32B. Widening the context mostly buys a
slightly better single answer while leaving the rest of the answer space untouched.
We close the loop with a submodular scheduler that reads the probe's feedback each
round to demote saturated evidence and promote under-used evidence (+0.033 to
+0.081 PR over every selection-style baseline, BH q < 0.001).
Code and the causal measurements are released — including an 11.5K-record benchmark
where the necessary evidence set is known by construction, so you can check whether
your own attribution method survives the swap from off-query to same-query negatives.
📄 https://arxiv.org/abs/2608.23252
💻 https://github.com/PeiYangLiu/ascp
🤗 https://huggingface.co/datasets/PeiyangLiu/ascp-context-attribution
Get this paper in your agent:
hf papers read 2608.23252 Don't have the latest CLI?
curl -LsSf https://hf.co/cli/install.sh | bash Models citing this paper 0
No model linking this paper
Datasets citing this paper 1
Spaces citing this paper 0
No Space linking this paper
Collections including this paper 0
No Collection including this paper