Knowledge Refresh, Agent Memory, and Unified 3D

Featured

01 Training KGD Decouples Knowledge From Geometry

Continuous pretraining must track changing user behavior without damaging representations already tuned for downstream tasks. KGD separates these jobs. A refreshable encoder captures behavioral knowledge, while task models access it through read-only cross-attention and calibrate their own representation geometry.

The training objective also filters behavioral supervision. Multi-token prediction keeps only future items with collaborative or semantic relevance. This avoids treating unrelated clicks from adjacent sessions as meaningful dependencies.

KGD reportedly beats strong baselines by 4%–12% across eight public benchmarks. It maintains its advantage over a 90-day production stream, while comparison methods show no gain. Shopee's online A/B test reports 1.75% higher GMV per user in homepage search and 1.53% higher advertising revenue. The full paper still needs to clarify refresh frequency, training and storage costs, and each module's contribution.

Key takeaways:

• Separate knowledge updates from downstream representation adaptation to protect existing tasks during each refresh.
• Filter spurious cross-session associations instead of treating every adjacent click as a dependency.
• Judge refresh systems by long-term data streams and online results, not one-time offline gains.

Source: Knowledge-Geometry Decoupling: Refreshable Pretrained Transfer for Streaming Recommendation

02 Evaluation More Memory Does Not Make Agents Better

Persistent memory is often treated as the foundation for personal agents that improve over time. Saving preferences, history, and skills only proves that the agent retained information. It does not prove that future behavior improved.

PAST-Bench tests that distinction across 26 scenarios and 204 task episodes. It compares matched conditions with and without experience retention. The benchmark also checks whether gains follow the intended save, retrieve, and update path.

Across seven base models and four agent frameworks, similar surface gains can have very different supporting evidence. The researchers added five targeted interventions to Hermes, with the largest improvements on tasks requiring outdated state replacement. Results still depend on the model and specific capability.

Key takeaways:

• Do not equate persistent memory with self-improvement.
• Longitudinal evaluations should test both later gains and the save, retrieve, and update chain behind them.
• Validate memory systems by model and capability because average scores hide important differences.

Source: PAST-Bench: Benchmarking the Foundations of Recursive Self-Improvement in Personal Agents

03 Multimodal One Model for 3D Workflows

Generating a 3D object is only the first step. Production workflows also need models to understand parts, edit selected regions, and leave everything else unchanged.

Hunyuan3D-Buffalo 1.0 combines 3D understanding, text-to-3D generation, instruction-based editing, and specified-part generation in one architecture. Its understanding module supplies semantic and spatial conditions. A diffusion model then generates or modifies the asset.

The team built an 87-million-sample multimodal corpus, including 12 million editing examples. The abstract says generation and understanding also improve editing performance. The main advance is workflow integration, not another isolated benchmark lead. Geometry quality, complex-instruction reliability, and production control still require full-paper review and testing on real projects.

Key takeaways:

• Evaluate 3D models across understanding, generation, and local editing.
• Geometrically consistent editing data may become a major barrier to building unified 3D models.
• Focus on workflow integration without treating benchmark leadership as proof that 3D generation is solved.

Source: Hunyuan3D-Buffalo 1.0: A Unified Multimodal Model for Scalable 3D Generation, Understanding, and Editing