Catching "Synergy Inflation": Where Literature-Reading LLMs Get Herbal Combination Claims Wrong

When a large language model summarizes papers on botanical medicine, it has to report a very particular kind of claim: that a combination of compounds does something that its parts do not. Examples are three flavonoids from Scutellaria suppressing renal fibrosis at doses where each alone fails, or a four-herb formula reversing lipid disorders in mice. These combination claims are where literature-reading models are most likely to go wrong, for three reasons. The pre-training text on herbal remedies is full of unverified folk assertions that "the whole is greater than the parts." The papers themselves often put a tested comparison against single components in the results section and an untested assertion of synergy in the discussion. And a combination claim requires the model to join evidence about several entities that may sit in different sentences.

We call the resulting error synergy inflation. It happens when a model reports a joint, greater-than-additive effect that the source only supports as a single-component effect, an effect of the combination with no comparison to its parts, or a speculative remark. This proposal builds the first benchmark that isolates this error. It then tests whether reading the model's internal activations detects the error better than the black-box pipeline, prompting, and trained-classifier defenses already known to work for ordinary hallucination. The core bet is that inflated combination claims have a specific internal signature. The model's attention fails to land on any single passage that jointly mentions the combination and its comparison to components, while the late feed-forward layers push toward "synergy" from memory. We also bet that this signature survives the external safeguards that suppress most other errors.

Core Claim

Prior mechanistic work on retrieval-augmented generation (models that answer from retrieved documents) found something specific. In ReDeEP, which detects hallucinations by separating how much the answer draws on copied context versus the model's stored knowledge, attention heads attended to the correct context in roughly 65–77% of hallucinated cases, yet late feed-forward layers overrode that context with parametric memory. Wang et al. traced the same contest between memory and context through the layers and found that feed-forward layers act as consistency checks.

Both studies used single-fact open-domain questions. We claim that combination claims in natural-product literature are a harder case in two ways that matter mechanistically. First, the parametric prior is unusually strong and unusually wrong, because folklore is common in the training data. Second, correct grounding requires joint evidence: a span that contains the combination and its comparison to single components, not just spans mentioning each compound.

We therefore propose a combination-aware grounding score. It is ReDeEP's external-context measure restricted to attention that lands on joint-evidence spans, and it is paired with ReDeEP's parametric measure. We predict that this score detects synergy inflation substantially better than generic hallucination detectors. The advantage should be largest on the residual errors that remain after an evidence-strict extraction pipeline has already filtered the output.

Research Program

The first component is a benchmark, which is necessary because the research file contains no dataset of combination claims. It shows only that the source literature exists and is structured the way we need. Papers such as the Wumeishanzhayin anti-obesity study (WMSZY), the red yeast rice versus lovastatin comparison (Red Yeast Rice), and the rosemary triterpene work (art_93e48880) mix tested combination effects, untested mechanistic speculation, and single-marker findings.

We will collect open-access abstracts and results sections from this literature. Annotators with pharmacology training will assign each (combination, outcome) pair one of five evidence grades:

  • tested superiority over components
  • combination effect without component comparison
  • single-component effect only
  • asserted-in-discussion only
  • tested and found additive or antagonistic

We will add a controlled counterfactual split. Results sentences that report tested synergy will be minimally edited to report additivity or antagonism, and the rest of the paper will be left intact. On this split, the only way to answer correctly is to follow the context against a strong prior.

The second component is a generation and labeling protocol. Open-weight models whose internals we can read, namely the LLaMA-3-8B and Qwen-2.5 families studied by Wang et al., will summarize each paper and list its combination claims. Each output claim is aligned to the gold evidence grade. A claim counts as inflated when its asserted grade exceeds the gold grade. We report two error rates:

  • Synergy-inflation rate: the fraction of output combination claims that are inflated.
  • Counterfactual follow rate: the fraction of edited papers where the model reports the edited, non-synergistic result.

The third component is the detector. For each token span expressing a combination claim, we compute three quantities:

  • ReDeEP's external-context score, based on copying heads.
  • ReDeEP's parametric score, from logit-lens projections of the knowledge feed-forward layers.
  • Our joint-evidence variant, which counts copying-head attention only when it falls on a sentence or table cell that contains every combination member together with comparative language.

A small regression over these features, as in ReDeEP, produces a claim-level inflation score. A detector-gated intervention in the style of ReDeEP's AARF, which amplifies attention and reduces feed-forward contribution, is applied only to flagged claims. This follows the gating logic of CLAP, so that correct claims are not degraded.

Throughout, the phytochemistry supplies the regime where the prior is adversarial. Information overload in the literature supplies the motivation and the pipeline setting. Interpretability is the method, and hallucination is the measured outcome.

Baselines

The naive reference point is the unmodified open-weight model summarizing each paper with no detector. Its synergy-inflation rate and counterfactual follow rate set the floor. We add a trivial textual detector alongside it: flag any combination claim whose members never co-occur in one source sentence with comparative wording. This tests whether joint-evidence grounding can be read off the surface text without any activations. If this string check matches our score, the interpretability contribution collapses, and we will report that outcome as such.

The strongest external-pipeline competitor is an Affinage-style decoupled pipeline. In that design, a reading pass extracts only findings backed by direct experimental comparison, and a synthesis pass sees only those extractions. Affinage reports 97.95% claim support and a 0.029% contradiction rate on gene annotation. It is the most serious threat because its core rule, "only claim what an assay shows," directly targets our error. We will reimplement it with the same open-weight models rather than its commercial ones. We will write a synergy-specific extraction rule that requires a component comparison, and give its prompt the same tuning budget as our detector's hyperparameters.

The strongest trained-detector competitor is the fine-tuned Llama-2-13B hallucination classifier from RAGTruth. It reports 78.7% response-level F1 and 52.7% span-level F1 on RAGTruth, against 63.4% and 28.3% for GPT-4. We will fine-tune it on the same training split of our benchmark that our regression uses.

The strongest internal-activation competitors are ReDeEP itself, which reports a chunk-level AUC of 0.7458 on RAGTruth with LLaMA2-7B versus 0.7290 for RAGAS and 0.5844 for SelfCheckGPT, and CLAP, which reports gains of up to 6.3% AUC over single-layer probes out of distribution. Both will be trained on identical data. The behavioral competitor is HCPD, which reports an average AUROC of 88.19% on Llama-3.1-8B across open-domain QA. We run it with its K=5 sampling.

There is no reported result on Lacuna for any method on herbal combination-claim grounding, so every number on our benchmark will be new.

The key ablation compares our score with and without the joint-evidence restriction. This isolates whether the gain comes from the combination-specific attention target or simply from ReDeEP's features. Beating the baselines means a higher claim-level AUROC than every competitor, with a confidence interval from a paired bootstrap that excludes zero. The comparison must be made on the residual errors of the Affinage-style pipeline, not only on raw model output.

Why not just…?

Why not just use an evidence-strict extraction pipeline like Affinage? This is the most likely way the proposal fails, and we test it directly rather than argue around it. Affinage's near-zero contradiction rate was measured on single-gene function, where "direct assay" is easy to recognize. Combination claims add a failure the pipeline cannot see from outside. The reading pass may correctly extract "compound A reduced fibrosis" and "compound B reduced fibrosis," and the synthesis pass, seeing both, may still write "A and B act synergistically." If the pipeline drives the inflation rate on our benchmark to near zero, our contribution shrinks to a diagnostic study, and we will say so. If inflated claims survive the pipeline, the question becomes whether our score catches them better than the alternatives. Experiment one below settles this.

Why not just fine-tune a hallucination classifier on the new benchmark? The RAGTruth detector shows that fine-tuning is a strong baseline, and we run it. Our concern is transfer: a text classifier trained on one set of herbs may learn which herbs "are synergistic" rather than whether the context says so. The counterfactual split tests this directly. On edited papers the text differs from a truly synergistic paper by one sentence, so a classifier relying on the herb prior should fail there, while a context-grounding signal should not.

Why not just ask the model what it believes without context and flag outputs that match that belief? This black-box prior check needs no activations. It is a genuine threat on the counterfactual split, where prior and context are opposed by construction. On unedited papers it has a blind spot: when prior and evidence agree in direction but the paper never tested the combination against its components, the prior check has no disagreement to flag. We include it as a baseline. We will report separately the cases where prior and context agree but the evidence grade is weak, which is where we expect the mechanistic score to add value.

Why not just use a behavioral judge like HCPD? HCPD's own authors note that evaluator agents are limited by their parametric knowledge on unfamiliar science. That limitation is the folklore problem in another form: a judge that believes the synergy folklore will accept inflated claims. The counterfactual split measures this directly.

Proposed Experiments

Experiment one measures the error. We report synergy-inflation and counterfactual follow rates for the raw models, the Affinage-style pipeline, and a neuro-symbolic knowledge-graph variant that feeds the model only extracted triples. We also compare results-section-only and full-abstract inputs, to see whether the "more is less" degradation reported by Tang et al. is concentrated in combination claims.

Experiment two runs the detector comparison described in the Baselines section. We report claim-level AUROC overall, on pipeline residuals, and on the counterfactual split, with the joint-evidence ablation.

Experiment three tests mechanism. We causally patch the attention to joint-evidence spans, meaning we replace those activations with ones from a run where the evidence is present, and check whether this restores correct evidence grades. This confirms the signature is causal rather than merely correlated.

Experiment four applies gated steering and measures two things: the reduction in inflation, and the rate at which correct claims are degraded.

As transfer checks, we run the detector on RAGTruth to confirm it does not regress on general hallucination. We also run it on pairwise extraction from KD-DTI, where the reported semi-supervised best is 39.78% F1 on DrugBank, to confirm the joint-evidence score reduces to ordinary grounding when the claim involves only one pair.

The decisive result is experiment two on pipeline residuals. If the joint-evidence score beats the fine-tuned classifier, ReDeEP, and the prior check there, internal reading adds something that external scaffolding cannot. If it does not, the answer to the tension between symbolic scaffolding and internal steering in this domain is that scaffolding suffices.