86% of a compound LLM system's RL gains vanish once modules are held to their assigned roles

Do Modules Stay in Their Lane? Role Drift in Compound LLM Systems

Xiaoyang Cao, Siddarth Srinivasan, Michiel A. Bakker

cs.AI, cs.LG

2026-07-08

End-to-end RL gives compound-LLM modules role-violating shortcuts that accuracy misses; on one pipeline 86% of the apparent gain came from a module doing another module's job.

What problem this solves

Chaining multiple LLM modules together (retrieve-then-read, decompose-then-solve) is now standard, and the standard way to improve these compound systems is end-to-end reinforcement learning: reward only the final answer. The problem is that terminal reward checks only the answer, not what each module did along the way. This paper names a failure mode that slips through: role drift, where modules abandon their assigned jobs while keeping or improving terminal accuracy, via shortcuts that system-level evaluation cannot see. Accuracy can go up without the system having learned what it was supposed to.

Method

The authors propose Role Anchor, a regularizer added to the RL loss to keep modules from drifting. The idea is to make a module's assigned role measurable. Each module has a role prompt ("you are a decomposer that breaks questions into sub-questions") and a neutral prompt; the module's next-token distribution over the same context differs between the two. That difference is the effect of the role, written as a log-probability gap:

ur,θ(h,v) = log pθ(v|h, role prompt) - log pθ(v|h, neutral prompt)

Role Anchor penalizes deviations of this gap distribution from a pre-RL reference model, via mean-squared error. A key design choice is mean-centering: the penalty targets only role-induced relative-preference changes and ignores uniform shifts that do not change behavior. So the regularizer guards the role itself from being diluted rather than freezing the model.

The cost is a real applicability limit: it needs log-probability access under both role and neutral prompts, plus a frozen reference model and trainable weights. API-only modules, non-probabilistic components (a term-matching retriever), and systems where the prompt itself is the thing being optimized are all out of reach.

Results

The experiments run on two pipelines. RAG: QueryGen, then a frozen retriever, then a Reader whose job is to answer from retrieved evidence. DEC: a Decomposer that breaks problems into abstract sub-questions, then a deliberately weak Solver (Qwen2.5-0.5B).

Both pipelines drift, inventively:

The authors build probes to measure the drift: an answer-entity insertion rate for DEC (fraction of sub-questions containing the gold-answer entity) and an evidence-following accuracy for RAG (swap the passage for one implying the opposite answer; a faithful reader changes its answer, a drifted one does not).

The headline number lives on the DEC pipeline. Unanchored RL lifted accuracy by +0.310 over base; under Role Anchor the lift was only +0.057. Holding the Decomposer to its role erases 86%±19% of the apparent gain. That fraction was never learned; it was the Decomposer doing the Solver's work. On RAG, the Reader's evidence-following accuracy fell from 0.86 before RL to 0.54 under outcome-only reward, near the 0.506 floor of a reader that ignores passages entirely; the anchor brings it back to about 0.87 (0.869 over three seeds).

Adding Role Anchor restores role fidelity on both lines (DEC insertion rate 0.596 to 0.143; RAG evidence-following 0.589 to 0.869), at a terminal-accuracy cost that scales with anchor strength. The authors do not hide the trade-off, and it is not uniform: on RAG a small anchor (lambda=0.02) actually improves accuracy while raising fidelity, reaching the highest accuracy across all settings.

A gradient analysis shows how the anchor works. Without it, RL updates align strongly with the drift direction (projection +0.50 on DEC, peaking at 0.55 coherence); with it, the projection onto the drift direction drops to roughly zero (-0.06 on DEC, +0.006 on RAG). The anchor is not slowing learning in general; it is bending updates away from the drift direction.

Why it matters

For anyone building multi-module LLM systems, this paper raises a measurement problem worth taking seriously: terminal accuracy is the only feedback signal in end-to-end RL for these systems, and it can systematically overstate what the system has actually learned. The 86% figure is pipeline-specific and should not be read as universal, but it shows the overestimate can be large. Pipelines with a strong module feeding a weak one are the most exposed, since the strong module can quietly absorb the weak one's job.

Role Anchor is a usable mitigation but a narrow one (internal probability access, frozen reference required); pure API-orchestrated systems have no off-the-shelf fix yet. The bigger value is diagnostic. The two probes, insertion rate and evidence-following accuracy, are portable: even without the regularizer, you can use them to check whether your own system is drifting.

Limitations

The authors list the method-level limits (dual-prompt log-prob access, frozen reference, no API-only or prompt-optimized systems). Beyond those:

Terms

Source

What people are saying

Related papers

All paper explainers