AIAny
Icon for item

Scaling Laws for Hypernetwork-Based Knowledge Injection in Large Language Models

Studies train-time knowledge injection via hypernetworks that generate fixed LoRA adapters from large fact corpora, empirically characterizing power-law scaling across hypernetwork depth, width, and target model size and reporting improved OOD generalization.

Introduction

Injecting factual knowledge into LLMs at scale is still an unanswered engineering problem: fine-tuning every model or keeping enormous retrieval stores can be costly or brittle. This paper shows that using a hypernetwork to generate a fixed LoRA adapter at train time cleanly decouples injection capacity from the base model and enables a controlled study of how injection performance scales with architecture and model size.

Key Findings
  • Empirical power-law scaling: loss and reasoning accuracy follow broadly predictive power-law behavior when varying hypernetwork depth, width, and target model size — so you can predict returns from scaling decisions instead of relying on ad-hoc tuning.
  • Better OOD generalization at scale: as hypernetworks grow, they generalize more reliably to out-of-distribution queries, indicating that train-time hypernetwork injection can transfer beyond the exact facts seen during training.
  • Competitive against other train-time methods: hypernetwork-generated LoRA adapters exhibit steeper scaling exponents in OOD evaluations compared to conventional LoRA fine-tuning and full fine-tuning, implying more favorable returns with scale.
  • Large evaluation corpus: the authors construct MegaWikiQA (tens of millions of multi-hop QA examples across 39 domains) to benchmark factual reasoning and out-of-distribution performance — so claims are validated on a wide, multi-domain dataset.
Who it's for & tradeoffs

Great fit if you need to inject large corpora of static factual knowledge into many models or model sizes and want predictable scaling behavior and improved OOD robustness. The approach reduces per-model manual fine-tuning by generating adapters centrally.

Look elsewhere if you require instant, per-query updates to facts (online updates are not the focus here), or if minimal infrastructure and very low-latency per-request edits are your primary constraints. Hypernetworks add training complexity and an extra artifact (generated adapters) to manage, and they still require substantial compute and data to reach the best scaling regimes.

Methods and dataset

The method trains a hypernetwork to map sets of facts (from a large corpus) to weights of a fixed LoRA adapter; that adapter is then inserted into a target LLM to enable answering fact-based questions. The paper evaluates depth/width/target-size axes and measures loss, reasoning accuracy, and OOD generalization. MegaWikiQA — a constructed dataset containing tens of millions of multi-hop QA examples across 39 domains derived from Wikidata5M — is used for large-scale evaluation and to measure out-of-distribution transfer.

Information

  • Websitearxiv.org
  • AuthorsNischay Dhankhar, Dos Baha, Abulhair Saparov
  • Published date2026/07/21

More Items

Proposes Riemannian Isometric Policy Optimization (RIPO) to fix exploration collapse in PPO-style RL for LLMs by aligning policy updates with the policy manifold's Riemannian geometry, improving exploration–exploitation balance and optimization stability across competition benchmarks.

Quantifies active visual observation in multimodal LLMs with ActiveVision, a 17-task benchmark that forces repeated perception rather than one-shot description. Finds frontier MLLMs fail badly (top model 10.6% vs humans 96.1%) and that model-generated vision code does not close the gap.

Performs full-parameter post-training of trillion-parameter MoE DeepSeek-V4 models on an Ascend NPU SuperPOD, using a hierarchical optimization of model parallelism, communication orchestration, and kernel execution to increase Model FLOPs Utilization. Also builds CPT/SFT pipelines with solver-verified synthetic data for Operations Research, reporting strong zero-shot Pass@1 results.