Pith. sign in

REVIEW 5 cited by

Avoiding mode collapse in diffusion models fine-tuned with reinforcement learning

Not yet reviewed by Pith; the record is open.

This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.

SPECIMEN: schema-true, not a live event

T0 review · schema-true

One-sentence machine reading of the paper's core claim.

pith:XXXXXXXX · record.json · timestamp

arxiv 2410.08315 v1 pith:KK5NMMBE submitted 2024-10-10 stat.ML cs.LG

Avoiding mode collapse in diffusion models fine-tuned with reinforcement learning

classification stat.ML cs.LG
keywords modelsdiffusiondownstreamfine-tuninglearningalignmentapproachcollapse
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved
0 comments
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Fine-tuning foundation models via reinforcement learning (RL) has proven promising for aligning to downstream objectives. In the case of diffusion models (DMs), though RL training improves alignment from early timesteps, critical issues such as training instability and mode collapse arise. We address these drawbacks by exploiting the hierarchical nature of DMs: we train them dynamically at each epoch with a tailored RL method, allowing for continual evaluation and step-by-step refinement of the model performance (or alignment). Furthermore, we find that not every denoising step needs to be fine-tuned to align DMs to downstream tasks. Consequently, in addition to clipping, we regularise model parameters at distinct learning phases via a sliding-window approach. Our approach, termed Hierarchical Reward Fine-tuning (HRF), is validated on the Denoising Diffusion Policy Optimisation method, where we show that models trained with HRF achieve better preservation of diversity in downstream tasks, thus enhancing the fine-tuning robustness and at uncompromising mean rewards.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 5 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. An Analysis of Posterior Collapse, Parameterization and Initialization in Variational Deep Gaussian Processes

    cs.LG 2026-06 unverdicted novelty 7.0

    Analysis of posterior collapse in variational DGPs shows the linear prior mean benefit comes from better optimization conditioning at initialization; a new zero-mean initialization prevents collapse and matches linear...

  2. Primitive-based Truncated Diffusion for Efficient Trajectory Generation of Differential Drive Mobile Manipulators

    cs.RO 2026-04 unverdicted novelty 6.0

    A primitive-based truncated diffusion model with keypoint attention encoding generates more efficient and diverse trajectories for mobile manipulators than vanilla diffusion in cluttered 3D simulations.

  3. When Models Learn to Ask Why: Adaptive Causal Reasoning for Trustworthy Medical Vision-Language Models

    cs.AI 2026-03 conditional novelty 5.0

    MedCausalX adds adaptive causal reflection tokens and trajectory-level RL on a new CRMed dataset to cut medical VLM hallucinations and raise diagnostic consistency.

  4. Statistical Properties of Training & Generalization

    stat.ML 2026-06 unverdicted novelty 2.0

    Neural scaling laws in deep learning interact with physics constraints and inductive biases beyond classical statistics.

  5. Statistical Properties of Training & Generalization

    stat.ML 2026-06 unverdicted novelty 1.0

    Review of neural scaling laws and their relation to constraints and inductive biases when applying machine learning to physics problems.