Diffusion and foundation models are gaining steering and transfer across domains

Tilting diffusion at test time

Diffusion and foundation models are gaining steering and transfer, starting with how a diffusion model is guided after training. A common recipe to improve diffusion models at test time is to introduce the gradient of the reward into the dynamics of the diffusion, but standard reward-tilting methods are often ill-posed because user-specified rewards are typically only well-defined at the end of generation. That gap between a reward defined at the end and guidance applied during generation is the problem the new method addresses. m:…6445

NVIDIA researcher kkreis proposes Flow Map Trajectory Tilting, or FMTT, to solve issues with standard reward-tilting methods. The company says FMTT performs better ascent on the reward than standard test-time methods involving the gradient of the reward, and FMTT can be used for exact sampling via importance weighting or for identifying local maximizers of the reward-tilted distribution. The company also says FMTT enables new forms of image editing by interfacing with vision language models. The claim is about steering quality at test time, not about training a larger model. m:…6445

Distilled diffusion for language

The second thread is diffusion applied to language itself. Diffusion language models intrinsically fail to capture correlations between decoded tokens, and that failure leads to a trade-off between sampling quality and throughput. DiLaDiff is a variant of masked diffusion language models that uses a continuous latent space with semantic capabilities, learned by an auto-encoder fine-tuned from an existing masked diffusion language model. m:…4158

DiLaDiff uses a latent diffusion model to learn the prior over the encoder distribution, and uses a consistency model to distill the learned prior into a few-step latent generative model. The company says consistency distillation lowers the computational overhead of continuous diffusion, and DiLaDiff outperforms the masked diffusion baseline in sampling quality even without distillation. The company also says DiLaDiff accelerates inference compared with the masked diffusion baseline, and latent generation in DiLaDiff takes negligible time compared with discrete decoding. The practical point is a shorter path from latent representation to text, with quality held up by the latent space rather than by more decoding steps. m:…4158

Sequence models that cross domains

The third thread moves from text and images to biological sequences, and here the evidence comes from preprints, so the results should be read as reported by the authors rather than as settled findings. In a preprint, PSICHIC+ predicts probability distributions over binding affinities from protein sequence and ligand SMILES data, and outperforms regression- and classification-based approaches in zero-shot virtual-screening benchmarks. Combining PSICHIC+ with Boltz-2 improves hit-to-lead affinity prediction compared with Boltz-2 alone, without requiring high-resolution protein structures. That last detail matters for the intended use, because it removes a data requirement that would otherwise limit where the model can be applied. m:…8067

In a preprint, single-cell foundation models can recover developmental order and cell identities in organisms separated by 1.2 billion years of evolution without retraining. The UCE model achieves 97 percent gene coverage in cross-species tasks compared with under 20 percent for vocabulary-based models. The contrast is the point: a vocabulary-based approach covers only a small share of genes when it crosses species, while the reported model covers nearly all of them in the same tasks. The authors report transfer without retraining, which, if it holds up, changes how much new data a new organism would require. m:…0809

Across all four items, the shared direction is control and reuse after the main training run: tilt a diffusion trajectory toward a reward, distill a latent prior into fewer steps, or carry a sequence model into a new species or a new screening task. This suggests developers should test these models on the transfer they actually need, since the reported gains are tied to specific benchmarks and settings rather than to general performance. A model that steers well on one reward or transfers well on one species pair has not yet shown the same result everywhere. m:…6445 m:…8067 m:…0809