How much does retrieval budget improve when triggered by dual signals instead of fixed intervals?
This explores whether retrieval works better when the model decides *when* to fetch new information based on live signals (like its own uncertainty) instead of grabbing chunks on a fixed schedule — and what the corpus actually says about the size of that gain.
This explores whether retrieval works better when the model decides *when* to fetch information based on live signals rather than on a fixed schedule. A quick honesty note first: the corpus doesn't frame the trigger as 'dual signals' versus 'fixed intervals' — what it consistently studies is *adaptive* triggering (one or more live signals) against *fixed* triggering (retrieve every N tokens, or once up front). So I'll answer the real question underneath: how much does letting the model choose its retrieval moments actually buy you?
The sharpest result is FLARE's: triggering retrieval on low token-probability — the model's own confidence dropping — improves both accuracy *and* efficiency over one-shot or continuous retrieval When should retrieval happen during model generation?. The key insight is that wasted retrievals aren't free: fixed-interval fetching spends budget on tokens where the model already knows the answer, and dilutes the context with noise. So the gain isn't only 'better answers' — it's that the *same or smaller* budget goes to the moments that genuinely need it. A complementary note shows you don't even need an expensive adaptive controller to capture most of this: calibrated token-probability uncertainty beats multi-call adaptive retrieval at a fraction of the LM and retriever calls Can simple uncertainty estimates beat complex adaptive retrieval?. The cheap signal — the model's self-knowledge — turns out to be more reliable than elaborate external heuristics.
If you want a concrete number for the value of *targeting* retrieval rather than spraying it, DeepRAG is the place to look: framing each reasoning step as a decision to retrieve-or-rely-on-memory yields a ~22% accuracy improvement, and the note is explicit that much of that comes from *eliminating noise* from unnecessary retrievals — not just from fetching more When should language models retrieve external knowledge versus use internal knowledge?. That reframes 'budget' nicely: adaptive triggering improves the budget twice over — fewer calls, and cleaner context per call.
Where does 'dual signals' actually live in this collection? It's latent rather than named. The failure-modes overview treats fixed-interval triggering as one of three *structural* defects in RAG — 'fixed intervals waste context' — implying the fix is signal-driven triggering, but it stops short of prescribing how many signals Where do retrieval systems fail and why?. Meanwhile, the budgeting conversation has moved beyond *when to fetch* toward *how much to think between fetches*: limiting per-turn reasoning (not just total time) preserves the context an agent needs to absorb the next retrieval round Does limiting reasoning per turn improve multi-turn search quality?. That's effectively a second budget knob layered on the trigger — the closest the corpus comes to a dual-signal regime.
The thing you might not have known you wanted: the bottleneck this corpus keeps circling isn't *how* you trigger, it's *what you do with each turn's compute*. CoRAG turns retrieval into a scalable test-time dial — greedy for speed, tree-search for accuracy — so retrieval budget becomes a tunable axis just like reasoning tokens Can retrieval be extended into multi-step chains like reasoning?. So the honest answer to 'how much does a dual trigger improve budget' is: the corpus shows adaptive triggering is a clear win (efficiency *and* accuracy, sometimes ~20%+), but the bigger lever it points at is treating retrieval as a controllable compute budget end-to-end, not just a smarter on/off switch.
Sources 6 notes
Active retrieval triggered by low token probability improves both accuracy and efficiency compared to one-shot or continuous retrieval. FLARE demonstrates that models signal genuine knowledge gaps through low confidence, enabling dynamic budget allocation to actual information needs.
Calibrated token-probability uncertainty consistently beats multi-call adaptive retrieval on single-hop tasks and matches performance on multi-hop, using a fraction of the LM and retriever calls. The model's self-knowledge proves more reliable than external heuristics for deciding when to retrieve.
DeepRAG models each reasoning step as a Markov Decision Process where the model learns when to retrieve versus rely on parametric knowledge. The 21.99% improvement comes from better-targeted retrieval and elimination of noise from unnecessary external knowledge.
RAG systems fail at three structural levels: adaptive triggering (fixed intervals waste context), semantic-task mismatch (embeddings measure association, not relevance), and mathematical limits (embedding dimension constrains representable document sets). These require fundamentally different retrieval approaches, not tuning.
Unrestricted reasoning within single search turns consumes context needed for subsequent retrieval rounds, degrading the agent's ability to incorporate new evidence. Setting per-turn reasoning budgets, not just overall time limits, prevents this context erosion and maintains search quality across iterations.
CoRAG extends chain-of-thought training to retrieval by using rejection sampling to generate intermediate retrieval chains. Test-time compute can scale through chain length and count, creating a compute dial—greedy decoding for speed or tree search for accuracy—just like reasoning-token scaling.