CWM: Controllable White-Box Meta-Prompting for Adaptive Retrieval-Augmented Generation and Reasoning Ability
Keuntae Kim, Eunhye Jeong, Yong Suk Choi
EMNLP 2026 - findings
cs.AI
2026-09-14
Hanyang's CWM biases Self-Discover's retrieval module with query Self-PPL, no classifier or multi-sampling; GPT-oss-20b HotpotQA EM rises from 52 to 71.
Adaptive RAG has to decide whether a query needs retrieval. Existing systems hit three walls: a pretrained classifier that is glued to one task and one dataset; methods such as Rowen and SEAKR that estimate uncertainty with multi-sampling or hidden-state probes and therefore spend more inference; and evaluations that almost never ask what happens when a pure reasoning question is sent through a retriever it does not need.
Black-box pipelines cannot see internals, so they add modules or run the same input many times. CWM's bet is narrower: the model's Self-Perplexity on the query is already a pre-generation signal, and it can drive the retrieval decision inside a meta-prompt.
The skeleton is Self-Discover. The model picks from a library of atomic reasoning modules, writes a task-specific meta-prompt, then answers. CWM adds one Retrieval module that asks whether external knowledge is required.
Three stages.
The pipeline is three forwards: module selection, meta-prompt, answer. No external classifier, no multi-sampling. The magnitude of α is a hyperparameter, so retrieval can be tightened or loosened per model and task.
Evaluation uses GPT-oss-20b, Qwen3-14b, and Llama3.1-8b, 100 random items per dataset. Adaptive RAG is HotpotQA, MuSiQue, StrategyQA; reasoning is three BBH subtasks, T4D, and MATH500. When a method decides to retrieve, every system including CWM is given the same gold context. The comparison is routing plus generation, not retriever quality.
On adaptive RAG, CWM's Best setting leads Self-Discover almost everywhere and often beats Adaptive-RAG, Rowen-CL, and SEAKR.
| Model | Dataset | Metric | CWM Best | Self-Discover | Adaptive-RAG |
| GPT-oss-20b | HotpotQA | EM | 71 | 52 | 69 |
| GPT-oss-20b | StrategyQA | Acc | 85 | 65 | 81 |
| Qwen3-14b | HotpotQA | EM | 78 | 77 | 68 |
| Llama3.1-8b | HotpotQA | EM | 68 | 55 | 61 |
| Llama3.1-8b | MuSiQue | EM | 53 | 37 | 33 |
Qwen3-14b's 78 EM on HotpotQA beats the gold-context ceiling of 70; Llama ties that ceiling on MuSiQue at 53. The meta-prompt is helping generation, not only routing. Qwen3 retrieves on 100% of all three RAG sets, so for that model CWM is barely adaptive. GPT-oss retrieves on 54% of StrategyQA, which looks more like actual selection.
Rowen-CL and SEAKR, which rely on multi-sampling, retrieve too rarely on the easier StrategyQA set and sometimes fall below the no-context floor. On the subset each method chooses not to retrieve, CWM is best for every model and dataset, e.g. 54.3 vs Adaptive-RAG's 50.7 on GPT-oss HotpotQA.
On reasoning, CWM usually beats Self-Discover. GPT-oss BBH accuracy is 78.67 vs 65.67 (direct 72.33); Llama MATH500 is 64 vs 56 (direct 21). GPT-oss MATH500 is a retreat: CWM 96, direct already 99, Self-Discover 87. Biasing the retrieval module also changes which other modules get picked, and that side effect is not uniformly helpful.
Cost is essentially Self-Discover's cost. On HotpotQA, GPT-oss averages 121.79 s and 3 calls per item; Rowen-CL is 833.58 s and 39.69 calls; SEAKR 753.88 s. On Qwen3, CWM is 36.23 s against Self-Discover's 37.34 s.
For open-weight deployments that can read embeddings and token likelihoods, this is cheap adaptive RAG: three forwards, no classifier, no sampling. Anyone already running Self-Discover adds one module and an embedding bias.
It is not a retriever paper. Gold context means the numbers answer "when to open the knowledge door", not "what was retrieved". Qwen3 gains while retrieving always, which suggests meta-prompt quality may be worth more than the router, and that "adaptive" has to be read per model.
One hundred items per dataset is thin, and Best is a hyperparameter peak; AVG is lower, 67.83 EM on GPT-oss HotpotQA. The library adds a single Retrieval prompt on top of 39 untouched modules; whether Self-PPL can steer the others is untested.
White-box access rules out API-only models. Self-PPL misreads some queries: a fact newer than the training cutoff can look easy yet need retrieval, and awkward phrasing can look hard yet not. Retrieval in the experiments is gold context. A real retriever would turn routing mistakes into wrong documents; that setting is missing.