dLLM
A class of language models that apply diffusion model principles to discrete text sequence generation. These models generate text through masked-denoising iterations rather than autoregressive decoding, enabling parallel generation and local infilling operations.
Timeline (last 90 days)
In the sampling-based RareTrap, using dLLM mapping reduced character-level perplexity from 10^3–10^7 to 3–5.
Risk estimates based on random mapping and those based on dLLM mapping may differ by an order of magnitude, indicating that risk assessment based on random mapping severely deviates from the true risk under natural prompts.
The computational overhead of dLLMs may be higher than random mapping, and the actual attack cost needs further evaluation.
A recent paper proposes directly coupling the continuous latent space framework to continuous diffusion language models (dLLMs), replacing random mapping with a learned language distribution to generate more natural adversarial prompts.
In the optimization-based ThinkTrap, using dLLM mapping reduced character-level perplexity from 20–28 to 4.3–5.4.
Since all positions can in principle be processed in parallel simultaneously, dLLMs are often faster than autoregressive models when generating fixed-length sequences.
Plan-and-Patch is a plan-and-act framework whose core is using a diffusion language model (dLLM) to generate and repair plans.
All Facts (7)
In the sampling-based RareTrap, using dLLM mapping reduced character-level perplexity from 10^3–10^7 to 3–5.
50%UnverifiedRisk estimates based on random mapping and those based on dLLM mapping may differ by an order of magnitude, indicating that risk assessment based on random mapping severely deviates from the true risk under natural prompts.
50%UnverifiedThe computational overhead of dLLMs may be higher than random mapping, and the actual attack cost needs further evaluation.
50%UnverifiedA recent paper proposes directly coupling the continuous latent space framework to continuous diffusion language models (dLLMs), replacing random mapping with a learned language distribution to generate more natural adversarial prompts.
50%UnverifiedIn the optimization-based ThinkTrap, using dLLM mapping reduced character-level perplexity from 20–28 to 4.3–5.4.
50%UnverifiedSince all positions can in principle be processed in parallel simultaneously, dLLMs are often faster than autoregressive models when generating fixed-length sequences.
50%UnverifiedPlan-and-Patch is a plan-and-act framework whose core is using a diffusion language model (dLLM) to generate and repair plans.
50%