acceptodds
Under review as a conference paper at ICLR 2027

DHKR: Data-Demanded Capacity Curricula for Hierarchical Representation Routing in Frozen Language Models

Abstract

While adapting language models often relies on rigid capacities and static routing structures, we introduce Dynamic Hierarchical Knowledge Routing (DHKR), a general framework for representation adaptation on frozen backbones. DHKR treats structural intervention availability as a control variable distinct from input-conditioned allocation and parameter optimization. Using a Knowledge Steering Layer (KSL) as an external portfolio of nonlinear representation interventions, DHKR uses a rule-based controller to retain or release paths within a configured intervention portfolio. Counterfactual analyses show that replacing learned weights with uniform or averaged variants increases validation loss, while restricted availability yields lower validation loss than full-from-start baselines under a matched regularized protocol. These findings show that allocation, structural availability, and Controller execution produce distinct outcomes in the tested setting. Separate trajectories verify that training continues after portfolio release.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
representation adaptation, structural availability, conditional routing, dynamic capacity, frozen language models

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_bR3AOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces DHKR, a framework for adapting frozen language models by controlling the availability of external representation-intervention paths. A two-level Knowledge Steering Layer combines input-dependent routing with a rule-based controller that monitors training and routing statistics. The framework distinguishes learning intervention parameters, allocating inputs across paths, and deciding which paths participate in optimization. Experiments compare Controller, Static, Scheduled, and Full configurations under a shared regularized objective. The Controller achieves lower validation loss across three seeds, while counterfactual probes examine the contribution of learned allocation. Separate expanding trajectories document routing redistribution and optimization after path release. Additional experiments assess generation quality, calibration, and computational cost.

Critical strengths
  1. The paper distinguishes parameter learning, input-conditioned allocation, and structural availability, then examines them through useful controls.
  2. Counterfactual routing probes support the contribution of learned allocation. The three-seed comparison also shows consistent validation-loss improvements over static and full-from-start configurations.
Critical weaknesses
  1. The main comparison leaves feedback-dependent timing entangled with execution details. Controller and Scheduled both expand at step 5 and retain the same capacity afterward. Their difference therefore requires an explanation beyond the release schedule.
  2. The strongest Controller result concerns loss and token accuracy. In the separate generation experiment, fixed capacity performs better on all six metrics. Generation results for the main Controller would clarify its practical benefit.
  3. Key implementation details remain implicit, including the complete release rule, thresholds, regularizer definitions, and coefficients. These details matter because transition handling appears central to the gain.
Questions for authors
  1. What exact execution differences separate Controller from Scheduled, and what happens when their transition operations are matched?
  2. How does the main Controller compare on generation quality and cumulative training compute across the three seeds?
Reviewer_WaidOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper introduces DHKR a framework for adapting a frozen quantized language model using an external KSL layer. Rather than exposing a static pool of sub-networks from the start, as in KADA, DHKR utilizes a rule-based feedback controller that monitors optimization and routing statistics to “release” new paths into the portfolio during training.

Critical strengths
  • Decoupling structural availability from input allocation addresses a gap in dynamic routing and modular fine-tuning.
  • While DHKR requires more trainable parameter, the gain in training time is substantial.
  • Its 2-level sub-network expansion allows the model to grow dynamically.
Critical weaknesses
  • I personally found the paper difficult to follow. The method is complicate and to get important insights about the performance of DHKR the reader is forced to go to the appendix (e.g., for a comparison with existing baselines).
  • The paper can benefit from a clear explanation of the main contributions of DHKR with respect already existing approaches. The authors highlight that DHKR has several similarities/differences with existing frameworks but this information is scattered in the text.
  • Figure 1 could be improved by making it more self-explanatory, e.g., describing what f_k,1/2 is.
  • The related work section is very limited and more references should be added.
  • The paper notes that there are two evaluation losses that can be used. It is unclear how much of the performance gain is driven by the dynamic controller versus the tuning of the regularizers.
  • The paper lacks sufficient sensitivity analysis on hand-tuned hyperparameters to demonstrate that they do not overfit to a specific dataset.
  • More insights and/or visual representation of the selected paths for specific inputs could strengthen the paper.
  • The empirical evaluation is constrained to a single model and just two datasets. There is no exploration of scaling behaviors, extensions to different model families, and different datasets.
Questions for authors
  • DHKR doesn’t focus on minimizing the number of trainable parameters. In table 8 the authors report a comparison of the time taken for training compared to two baselines. What about the memory/computational cost required?
  • Which GPU was used for the experiments?
  • Would it be possible to compare DHKR with a standard MoE model?
Reviewer_ejSUOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes Dynamic Hierarchical Knowledge Routing (DHKR). It adds a hierarchical Knowledge Steering Layer on top of a frozen language model, and uses a controller based on training feedback to release subnetworks over time, aiming to achieve adaptive capacity allocation. The experiments use counterfactual interventions on routing weights, comparisons between different capacity strategies, and dynamic expansion trajectories to study how structural availability affects model optimization.

Critical strengths
  1. The paper separates structural availability from input-dependent routing and parameter optimization. This gives a useful perspective for studying dynamic capacity allocation.
  2. The experiments are targeted. The authors use weight substitution, path removal with renormalization, and several capacity configurations to analyze how different mechanisms affect performance.
Critical weaknesses
  1. No direct evidence for data-demanded expansion. In Table 2, Controller and Scheduled both expand from K=1 to K=2 at the same preset step 5, so their capacity trajectories are identical. The performance gap therefore cannot be attributed to feedback-driven decisions. The run with feedback-triggered expansions is a single trajectory without a control.

  2. The benefits of hierarchy and dynamic expansion are not shown. All three-seed experiments use L=1. Static K=2 beats Full K=5, and fixed capacity beats capacity-enabled DHKR in generation quality. This suggests that limiting capacity helps, but not that dynamic expansion beats a well-chosen fixed capacity. A budget-matched flat routing baseline is also missing.

  3. Limited scope and baselines. The results rely on one 8B backbone and Alpaca/SFT. The routing analysis uses one seed and 100 samples. Baselines such as LoRA are not tuned or compute-matched.

Questions for authors
  1. What exactly differs between Controller and Scheduled in Table 2? Can you ablate the feedback trigger alone and report feedback-driven expansion over multiple seeds?
  2. Can you add Static K=1 and a budget-matched flat routing baseline?
  3. Does DHKR consistently beat fixed-capacity configurations on other models or tasks?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.