- The paper introduces DALI, an LLM-agent framework that evolves leadership-identification rules and combines symbolic reasoning with neural classification for adaptive group recommendation.
- DALI improves average performance by up to 19.9% on AGREE and 13.0% on GroupIM, while AGREE+DALI raises Mafengwo NDCG@5 from 0.582 to 0.735.
- The results show that DALI is most effective when groups exhibit clear power asymmetries, while gains diminish on strong baselines or datasets dominated by collaborative decision-making.
Motivation and problem statement
Group recommendation (GR) systems aggregate individual member preferences into a collective decision, and existing approaches fall into two families: handcrafted aggregation rules (mean, least misery, weighted sum) and learned neural aggregation (attention-based frameworks such as AGREE). Both families implicitly assume that group output is a uniform function of individual inputs. The Leadership-Dominant Preference Hypothesis, introduced by LARGE, challenged this assumption by showing that some groups are dominated by a single influential member whose preferences effectively determine group choices. However, LARGE relies on static or predefined identification rules for distinguishing leader-dominated from collaborative groups, which cannot adapt to evolving power structures.
DALI addresses this gap by employing LLM-based agents to autonomously generate, validate, and evolve leadership identification rules during training, combined with a neuro-symbolic dual-stream discriminator that fuses symbolic rule reasoning with a neural classifier. The authors claim this is the first framework to use LLM agents for dynamic leadership identification in group recommendation.
Framework architecture
DALI is organized as a four-component hierarchical system built on top of an arbitrary base GR model whose attention module produces per-member weights αu.
Agent architecture. The LLM agent comprises four modules:
- Role Module: establishes two agent personas — a Rule Governance Expert that classifies groups as leadership-oriented or collaborative by matching weight-distribution features against the current rule repository, and a Rule Evolution Engine that detects obsolete rules and proposes new quantifiable IF-THEN candidates.
- Memory Module: a three-layer store consisting of a versioned rule repository with semantic versioning and rollback support, a case feature bank archiving exemplar leadership/collaborative weight patterns, and an append-only performance evolution log correlating metric fluctuations with rule versions.
- Planning Module: implements a Performance-Rule Feedback Loop (PRFL) that monitors training metrics in real time; upon detecting performance drift it triggers rule evaluation, generation, sandboxed validation, versioned deployment, or rollback.
- Action Module: executes classification by parsing attention weight tensors, matching rules, and invoking LLM behavioral analysis.
Neuro-symbolic hybrid aggregation (NSHA). Two channels process the same weight tensor W: a symbolic channel Pgsymbolic=Γ(RV(W)) driven by the agent's rules, and a neural channel — a three-layer gated MLP with ReLU/tanh activations producing Pgneural via softmax. A triple-threshold decision logic governs fusion: when both channels agree, the result is adopted directly; otherwise outputs are blended as Pg=γPgsymbolic+(1−γ)Pgneural, where γ decays linearly to zero over training epochs. This schedule deliberately trusts symbolic inductive bias early (when the rule base is immature but the neural channel is untrained) and shifts reliance to the neural model later. Identified leader groups collapse aggregation to the argmax-weight member's embedding (pg=pu0), while collaborative groups retain attention-weighted summation.
Training protocol. Phase 1 pretrains user representations with all aggregation modules frozen, following the base recommender. Phase 2 jointly optimizes L=Lgroup+Lweight, where Lgroup inherits group/user-level cross-entropy plus member-group mutual information loss, and Lweight enforces weight concentration in identified leader groups via a dominance score W0 compared against a dynamically sampled collaborative benchmark, penalizing leader groups whose relative dominance falls below margin W1. The dynamic benchmark avoids fixed-threshold distribution bias.
Experimental results
Experiments use Mafengwo (5,275 users, 995 groups, 1,513 items) and CAMRa2011 (602 users, 290 groups, 7,710 items), with a 60/20/20 group-level split. DALI is plugged into five base models (GroupIM, AGREE, HCR, CubeRec, DisRec) and compared against the same bases augmented with LARGE.
The headline results are substantial for foundational and attention-based models:
| Base model |
Dataset |
Avg. improvement (LARGE) |
Avg. improvement (DALI) |
| GroupIM |
Mafengwo |
+5.9% |
+13.0% |
| AGREE |
Mafengwo |
+5.0% |
+19.9% |
| AGREE |
CAMRa2011 |
−0.5% |
+19.9% |
| DisRec |
Mafengwo |
+2.1% |
+4.3% |
| CubeRec |
Mafengwo |
−0.3% |
+0.3% |
For example, AGREE+DALI reaches NDCG@5 of 0.735 on Mafengwo versus 0.582 for AGREE alone (+26.3%). Notably, LARGE degrades performance when applied to GroupIM and AGREE on CAMRa2011, whereas DALI remains positive there — evidence the authors cite for the robustness of the adaptive rule mechanism over static identification. Gains shrink toward zero for already-strong baselines (HCR, CubeRec), where DALI still delivers non-negative average improvement and generally positive NDCG gains.
The paper attributes cross-dataset variance to group structure: CAMRa2011's dense interactions among few small stable groups (families) favor collaborative decisions with little leadership signal, while Mafengwo's larger ad-hoc travel groups exhibit more pronounced power asymmetries. This is an honest scoping observation: DALI's benefit is positively correlated with the clarity of internal power structure, so it is not a structure-agnostic enhancer.
Ablation, sensitivity, and case study
Ablations on Mafengwo with GroupIM as base isolate each component. Removing the neural discriminator (GroupIM-A) drops HR@5 from 0.652 to 0.626; removing the rule agent (GroupIM-N) drops it to 0.617; using frozen manually designed rules (GroupIM-F) yields 0.613; the bare base model achieves 0.579. The full system outperforms every variant on all four metrics, supporting the claim that dynamic rule evolution — not merely the presence of rules — is necessary, since frozen rules fail to generalize complex power structures.
Batch-size sensitivity peaks at batch size 32 (HR@5 0.652, NDCG@5 0.511), with degradation at 16 (high gradient variance destabilizing rule learning) and at 256 (low gradient bias trapping optimization). The usable range is roughly 32–128.
The case study examines three evolved rules tied to observable metric dynamics: an abrupt decision rule (single-epoch NDCG@10 fluctuation > 0.01 triggers rollback and leader-weight recording), a power transition rule (three consecutive epochs of > 0.005 decline lowers the rollback threshold), and a deadlock resolution rule combining group/user metric divergence with loss signals. Each rule demonstrably captured subsequent analogous events (e.g., post-rule fluctuation narrowing from ~0.006 to ~0.003), illustrating that the agent extracts metric–leadership correlations that would be difficult to specify manually.
Limitations and open questions
Several caveats bear directly on the reported results. First, the gain magnitude is dataset-dependent: improvements on CAMRa2011 are modest for most bases except AGREE, and near-zero for CubeRec, so the framework's value hinges on the presence of genuine leadership dynamics — a property not always verifiable a priori. Second, the method depends on the quality of the base model's attention weights as the sole input signal to both channels; no experiment tests robustness when attention distributions are noisy or poorly calibrated. Third, the symbolic channel's interpretability claim rests on rule-matching paths, but the paper does not evaluate rule correctness against ground-truth leadership labels beyond qualitative case studies, and no human evaluation of the generated rules' plausibility is reported. Fourth, computational overhead of iterative LLM invocation during training is not quantified relative to baseline training cost. Finally, the triple-threshold logic and the linear decay of W2 are design choices validated only empirically; whether alternative schedules or thresholds transfer to other domains remains open, as does extension to multi-type group dynamics beyond the binary leader/collaborative taxonomy and to cross-domain adaptation.
Conclusion
DALI couples an LLM-agent rule engine operating under a practice–reflection–evolution closed loop with a neuro-symbolic dual-stream discriminator, replacing static leadership heuristics with adaptively evolved, interpretable rules. Across two datasets and five base recommenders it consistently matches or exceeds the static-rule LARGE framework, with the largest gains on attention-based and foundational models (+13% to +19.9% average improvement) and diminishing returns on strong baselines. The results position dynamic, agent-driven rule evolution as a viable mechanism for modeling power asymmetries in group decision-making, contingent on the presence of identifiable leadership structure in the target population.