← Back to feed

Standard pretraining losses have a structural problem baked into the objective: when training data is ambiguous, cross-entropy and diffusion MSE force models to hedge by averaging across all plausible outputs. The result is mode blurring — outputs that are statistically plausible but commit to nothing, and image models that require VAEs or VQGANs to compress away the ambiguity before generation can work at all.

A paper from Alexi Gladstone and colleagues proposes a different training loop: Explorative Modeling. At each training step, the model generates K candidate outputs. The loop evaluates which candidate best matches the real data target, then runs backpropagation only on the winning candidate. The other K-1 candidates are discarded.

In its simplest form, this is literally a for-loop wrapped around a standard training step.

Why the Standard Objective Fails

The core failure mode is well understood in theory but rarely addressed at pretraining scale. When a model sees a sequence that could plausibly continue in multiple ways — two equally valid next tokens, two valid image completions from the same caption — cross-entropy loss forces it to distribute probability mass across both continuities. This is mathematically correct for maximum likelihood estimation, but it means the model learns to be vague about ambiguous regions.

The diffusion variant of the same problem causes image generators to produce blurry samples in regions where the training data is genuinely ambiguous. The standard solution has been a two-stage pipeline: train a VAE to compress images into a continuous latent space where ambiguity is reduced, then run diffusion inside that latent space. The VAE is trained separately and adds engineering complexity.

Explorative Modeling sidesteps the averaging problem entirely by never averaging. If the model generates K candidates and commits only to the best, the loss signal always comes from a definite output, not a weighted compromise.

The Three Pretraining Axes

The paper frames exploration as a third axis for pretraining investment, alongside the established two:

Scale — more data, more parameters, more compute. The dominant axis since 2020, subject to diminishing returns that are actively debated.

Architecture — attention variants, mixture-of-experts, hybrid state-space models. A slower-moving axis with discrete innovations.

Exploration — the K parameter. Larger K means the model explores more candidates per training step, increasing the chance of finding a high-quality match to train on, at the cost of K forward passes per update.

This tradeoff has a different compute profile than scale. Scaling K from 1 to 4 costs 4x more compute per training step, but may produce better mode-committing behavior than spending the same compute on additional data or parameters. Whether the frontier returns on K exploration exceed those on scale is an open empirical question — the paper establishes the paradigm, not the scaling curves.

End-to-End Generation

The most practical immediate consequence is for multimodal generation. Explorative Modeling makes end-to-end generation tractable: instead of training on a VAE-compressed latent space, a model can train directly on raw pixels or raw audio, with exploration replacing the compression step.

HiDream O1-Image demonstrated this recently — an 8B pixel-space model trained end-to-end without a VAE, reaching state-of-the-art on open-weight text-to-image by scrapping the residual stream design that had been standard since VQGAN. Explorative Modeling provides a principled explanation for why end-to-end generation is now viable: exploration handles the ambiguity that previously required a compression bottleneck.

Frontier Implications

Current frontier models — Claude Fable 5, GPT-5.6 Sol, Kimi K3 — are trained with cross-entropy and RLHF objectives at scales the Explorative Modeling paper does not approach. Whether exploration transfers to trillion-parameter multi-modal runs is unverified.

If it does, the compute economics are interesting. Frontier labs are already running training clusters at the gigawatt scale. Exploration is parallelizable: K candidates can be generated simultaneously across replicas, not sequentially. At sufficiently large batch sizes, K exploration adds a modest multiplier to compute cost, not a sequential penalty.

The HN front-page placement suggests practitioner attention. The question for 2026 is whether any frontier lab includes exploration in their next pretraining run and publishes scaling results. If Explorative Modeling reaches the same diminishing-returns inflection point as scale, it becomes a commodity technique. If it yields orthogonal improvements, it opens a new research frontier alongside RLHF and chain-of-thought.

Key Numbers

• Paper: Explorative Modeling — Unlocking a Third Pretraining Axis and End-to-End Generation (arxiv)
• Training loop change: K forward passes per step, backprop on best match only
• Application: autoregressive text, diffusion image, end-to-end pixel-space generation
• Demonstrated: HiDream O1-Image (pixel-space, no VAE, open-weight SOTA)
• Frontier validation: pending