RecGPT-V2: Hierarchical Multi-Agent User Intent Reasoning¶

Core mechanism evidence. Paper and local evaluations are separate.

Paper information¶

Field	Value
Paper	arXiv 2512.14503
Affiliations	Alibaba / Taobao
First public release	2025-12-16 (arXiv v1)
Authors' code	No official author code linked as of 2026-08-09.
Adapter / method	`recgpt-v2`
Local code	`src/auto_research/reproductions/recgpt_v2/`

Original paper¶

Background and contribution¶

Fixed V1 chains duplicate work and generalize poorly. V2 uses hierarchical intent agents, compressed mixed behavior representations, dynamic meta-prompts, multi-reward constrained RL, and Agent-as-a-Judge.

flowchart LR
  H[Behavior history] --> A[Short, long, and category experts] --> M[Meta-router and mixed representation] --> U[KL-constrained preference update]

Key figure from the paper¶

Original Figure 2: RecGPT-V2. From the paper; copyright belongs to the authors. Not a local result.

Core mechanism¶

\[ \pi(a\mid h)=\sum_k g_k(h)\pi_k(a\mid h),\qquad \max_\pi\ \mathbb E[R]-\beta D_{\mathrm{KL}}(\pi\Vert\pi_0). \]

Paper results¶

Paper exclusive recall 9.39%→10.99%, GPU use −60%, tag prediction +24.1%, explanation acceptance +13.0%. Taobao live CTR +2.98%, IPV +3.71%, TV +2.19%, NER +11.46%.

Local reproduction¶

Vs. single-route V1 proxy: Hit@10 0.14048→0.14762, NDCG@10 0.07514→0.08905 (+18.52%), head share 0.21762→0.18167. Metrics.

auto-research reproduce --paper recgpt-v2 --dataset-dir data --seed 42

Reproduction limits¶

Agents are small trainable public-data experts, not private LLMs. No human judge labels, explanations, or online resource measurements.