Modern language-model fine-tuning typically pairs each prompt with a single response, even though many prompts admit multiple valid completions. This effectively reduces a multi-modal conditional distribution to a one-sample view, a phenomenon we call the "mode lottery," where training emphasizes a subset of plausible modes while leaving others underrepresented. We study multi-response training (MRT), which retains multiple responses per prompt, and develop a principled account of when and why it helps. Our key insight is that prompts and responses are distinct statistical resources: additional prompts reduce uncertainty about the input distribution, while additional responses reduce uncertainty about the conditional output distribution. This yields a variance-budget tradeoff that predicts when retaining multiple responses is worthwhile, shows diminishing returns as prompt-level uncertainty dominates, and explains why large redundant corpora can exhibit an implicit multi-response effect. We further analyze response selection, and show that Random-K-of-N is the unbiased default for distributional fine-tuning, reward-based selection can induce mode collapse, and a submodular quality-diversity objective provides an efficient alternative with theoretical guarantees. Controlled simulations validate the predicted variance and selection effects, including a striking failure mode where reward-only selection produces gradients misaligned with the true objective. Across structured and real-world datasets, including a new multi-prompt, multi-response benchmark, MRT consistently improves distributional generalization, with the largest gains in high response-diversity, low prompt-redundancy regimes. MRT reframes response multiplicity as a data-allocation problem with clear guidance: when responses are cheap and diverse, keeping more than one is not a heuristic, but a statistically grounded choice.
Post-training of language models is commonly framed as a sample-score-update loop implemented by gradient descent. A recent line of work, exemplified by RandOpt, relocates this loop to weight space, sampling Gaussian perturbations around a pretrained model and ensembling the top-K rewarded specialists at inference. While competitive with PPO and GRPO under matched training compute, this prediction-level ensemble incurs K forward passes per test example and does not extend cleanly to free-form generation. We ask whether the rewarded population can instead be folded into a single deployable model, replacing the inference-time ensemble with one consolidated update. A split-half analysis over 25 model-task pairs reveals reproducible low-rank structure in every case. We turn this geometry into CoRP (Consolidating Rewarded Perturbations), a gradient-free operator that combines reward-weighted aggregation, compatibility-aware reweighting, and a held-out validation gate, with no gradient flowing through the language model. Across five language models from 0.5B to 8B and five tasks covering math, code, and creative writing, CoRP improves the base model by 8.1 points on average. Using one tenth of RandOpt's perturbation budget, CoRP exceeds single-inference RandOpt by 6.5 points and recovers more than half of the gain of the 50-pass majority-vote ensemble, at one forward pass per test example.
Training prompts in online reinforcement learning (RL) differ substantially in how informative they are for the current policy: some are already saturated while others are too difficult to yield reliable learning signals, yet both receive equal rollout budget under standard training. We propose an exploration-guided prompt scaffolding framework that adapts the training prompt distribution dynamically throughout RL post-training of multimodal large language models (MLLMs). Central to our approach is the Exploration Potential Score(EPS), a lightweight rollout-based proxy for prompt utility derived from KL-regularized policy improvement theory, computable directly from on-policy rollout statistics without additional overhead. Rather than discarding low-utility prompts, we use a teacher model to generate scaffolded rewrites that preserve the original task intent while making subsequent training more informative, reframing teacher supervision as training-data refinement rather than output imitation. Integrated with GRPO on Geo3K and MMK12, our method consistently outperforms the baseline on both in-domain and out-of-distribution benchmarks, achieving up to 9.7% relative improvement in-domain and gains of 11.5% on MathVision and 11.1% on MMMU-Pro.
Silicon sampling uses language models as proxies for human survey respondents, treating each model call as an independent draw from the persona's response distribution. We show this draw does not exist: instruction-tuned models do not sample from distributions, they collapse to a single output. The same persona on the same question returns the same answer on more than half of items in a public-opinion benchmark, and the model's internal probabilities concentrate on a single option. The failure is associated with, and amplified by, instruction-targeted post-training: instruction-tuned models are worse than their own bases in every family we can compare, the gap widens at each successive post-training stage and with the size of the tuning update, and continued pretraining on non-instruction tokens leaves it unchanged. Yet the knowledge survives: the same model that cannot sample from a distribution can describe it accurately in a single call. We call this gap the KNOWS/DOES split. Exploiting the split, a single call that asks the model to describe the response distribution more than halves the error against human survey data compared to persona aggregation. When per-persona outputs are required, we propose Prompt-Perturbed Argyle (PPA), which reduces the same error by 21%, spreading each persona's answers to mirror real population differences at no added cost.