SIMGUIDE beats RAG on personalized AI agent planning tasks

Personalized AI agents typically represent a user as one flat profile crammed into a prompt. Researchers behind this paper argue that format breaks down when the same person carries different priorities across separate parts of their life, and breaks down badly when those priorities conflict with each other. Their core claim is that the shortage is not information about the user; it is that the shape of the representation decides whether an agent can act on that information at all.

The researchers introduce SIMGUIDE, a method that splits a user's context into typed, domain-specific blocks called Sims, each grounded with procedural examples drawn from the user's own past decisions rather than stated as flat declarative rules. To test whether this actually helps, they built SIMBENCH, a diagnostic suite of 47 preference-conditioned planning tasks in which the correct plan depends on which of the user's contexts is active, a property the authors say no existing benchmark checks for.

The results split sharply by how the Sims are built. Declarative Sim constraints, stated without grounding, do not outperform retrieval-augmented generation (RAG) used as a personalization baseline. But procedurally grounded Sims, the ones anchored to worked examples from past decisions, outperform RAG on GPT-4o by 7.9 Preference Adherence points (p = 0.013). The same advantage shows up again on 100 tasks drawn from the tau-bench suite, tested on both GPT-4o and Claude Sonnet 4.5 (p is at most 0.023).

The paper reports a parallel finding at the level of model fine-tuning: whether parametric adaptation works at all comes down to the training distribution. Task-matched LoRA fine-tuning lifts generation quality by 12.8 ROUGE-L points over an unadapted base model. Routing separate adapters by Sim type, rather than by user identity, adds a further 7.3 points on top of that gain, and the authors say this holds even with a 28% routing error rate.

The authors' overall conclusion: representation format, not representation content, is the first-order variable in getting a personalized agent to act on what it already knows about a user.

Key facts

SIMGUIDE structures a user's context into typed Sim blocks grounded with procedural examples from past decisions, instead of a flat declarative profile.
SIMBENCH is a diagnostic suite of 47 preference-conditioned planning tasks where the correct plan depends on which user context is active.
Declarative Sim constraints alone do not outperform RAG; grounding in procedural examples is what produces the gain.
Procedurally grounded Sims beat RAG by 7.9 Preference Adherence points on GPT-4o (p = 0.013), replicating on 100 tau-bench tasks across GPT-4o and Claude Sonnet 4.5 (p at most 0.023).
Task-matched LoRA fine-tuning adds 12.8 ROUGE-L points over the base model, and routing adapters by Sim type adds a further 7.3 points, robust to 28% routing error.

Why it matters

Most personalized agent systems reduce a user to a single flat profile pasted into a prompt. The paper's starting claim is that this format itself, not a lack of data about the user, is what stops an agent from acting correctly when a person's priorities differ across contexts or outright conflict. SIMGUIDE's proposal is to split context into typed Sim blocks and ground each one with procedural examples from the user's own past decisions, and SIMBENCH is built specifically to test whether a system can tell which context should govern a given plan, a property the authors say no prior benchmark measures.

Who it affects

The findings target teams building personalized AI agents and agent memory systems, along with researchers comparing RAG-based personalization against structured or fine-tuned alternatives. The parametric-adaptation results, on LoRA fine-tuning and adapter routing, also speak to teams deciding how to adapt a base model to per-user or per-context behavior rather than relying only on prompt-time retrieval.

How to use it

The paper does not state a release, open-sourcing, or deployment plan for SIMGUIDE or SIMBENCH, and it names no price or license terms, so there is nothing to point to on availability. What it does specify is the mechanism: grounding Sim constraints in procedural examples from past decisions, rather than declaring them flatly, is what separates the approach from a plain declarative profile or from RAG.

How solid is it

The headline comparisons carry statistical significance: the 7.9-point GPT-4o result has p = 0.013, and the tau-bench replication across GPT-4o and Claude Sonnet 4.5 holds at p of 0.023 or better. The evaluation spans two separate benchmarks (the 47-task SIMBENCH and 100 tau-bench tasks) and two different underlying models, and the routing-adapter result is reported as holding up under a stated 28% routing error rate. The source text gives no absolute Preference Adherence scores for either RAG or Sims, only the point differential between them, and it names no authors or institutional affiliation, so independent verification of the underlying numbers is not possible from what is available here.

Risks and caveats

Only relative gains are reported (the 7.9-point and 7.3-point differentials), not absolute scores, which limits how the results can be compared against other personalization methods outside this paper. The mechanism behind a Sim block is described only at a high level, as typed, domain-specific blocks grounded with procedural examples, without architectural detail. No author names, affiliations, or publication date are given, and no release or deployment timeline is stated, so the work's provenance and its path to practical use both remain unverified from the text on hand.

“Representation format, not representation content, is the first-order design variable.”

— the paper's authors