Next-token Preference: Scaling Unsupervised Reward Modeling

arXiv:2603.02225 · cs.LG · Submitted 2026-02-11 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "Next-token Preference: Scaling Unsupervised Reward Modeling".

Jane: The gist: Reward models trained without human supervision can yield steady gains on RewardBench v1 and v2 by exploiting next-token continuation structure from large-scale web corpora.

Tom: First, who's behind it and why it matters.

Title and authors: Tom: So we're talking about this paper today, "Next-token Preference: Scaling Unsupervised Reward Modeling." It's about how to train these reward models without needing tons of human feedback.

Jane: Exactly. The authors are proposing a new way to scale them using things they find in massive web data instead of relying on those expensive human labelings.

Tom: They operationalize reward-based scaling, or RBS, as preference learning over document prefixes and suffixes drawn from large-scale web corpora. That means they use the structure of how text continues on the internet to create implicit supervision for the model.

Jane: And they show that this approach gets steady gains on RewardBench v1 and v2 just by training on eleven million tokens of math-focused web data.

Lu: It's interesting because it moves the process away from needing explicit human annotations, which is a big hurdle for scaling these models up.

Meng: So, what does this mean practically for someone building these reward models? Is it just a nice theoretical exercise?

Tom: Well, the results show that training on those eleven million tokens of math web text consistently improves over initialized checkpoints. They see up to plus seven point seven average points on RewardBench v2 <ref:2603.02225#pg1>.

Jane: And these improvements aren't just isolated; they transfer across diverse initialization backbones spanning different model families and scales. That’s a pretty strong sign for generalizability.

Lu: The methodology involves sampling a random breakpoint in a text sequence to make a prefix prompt p and a suffix continuation r, then treating that original continuation r as the chosen response against other continuations in the batch as negatives.

Meng: So, it’s essentially creating an online set of preference pairs just by looking at how text naturally flows on the web. That seems like a very scalable data source.

Tom: Right. They use a Bradley–Terry objective over these online preference pairs to minimize the average negative log-likelihood over all such comparisons, treating r as the chosen continuation against B minus one rejected continuations, which is formula one of their training loss <ref:2603.02225#pg1>.

Jane: And they don't just stop there. They augment that loss function with a score-centering regularizer called Lcenter to keep the reward outputs well-behaved across all prompts.

Lu: That regularization part is key because it tackles that issue where weak labels can cause the reward model to inflate its scores way too much, leading to overconfident margins.

Tom: So they define the final training loss as L equals LBT plus a constant times Lcenter, and that's how they keep the learning stable under noisy supervision from web data.

Jane: The experimental setup involved training reward models from scratch on large-scale math web text, like FineMath or InfiMM-WebMath. They used eleven million tokens for that initial training budget.

Meng: So, what are the real numbers we're looking at when they talk about performance gains? Are these gains just theoretical improvements on a benchmark?

Title and authors: Tom: The results show that this training consistently improves over initialized checkpoints, achieving up to plus seven point seven average points on RewardBench v2 <ref:2603.02225#pg1>. Plus, they see gains of up to plus sixteen point one on in-domain math subsets and consistent improvements on out-of-domain tasks too <ref:2603.02225#pg1,16.1 on in-domain math subsets and consistent improvements on out>.

Jane: That performance boost translates into downstream utility, which is where the real value for applications comes in because it shows substantial gains on GSM8K and MATH accuracy.

Lu: The authors did some ablation studies, and they found that increasing the batch size in that online preference pairing scheme consistently improves reward model performance, scaling cleanly with B up to sixteen <ref:2603.02225#pg1>.

Meng: And what about the data? Did using a specific dataset make a difference? For instance, training on FineMath-4plus delivered larger peak improvements across both in-domain and out-of-domain subsets compared to InfiwebMath <ref:2603.02225#pg1>.

Tom: They also found that allowing sentence breaks in the prefix–suffix splitting yields much larger gains over the initialized seed on RewardBench v2 across all ID and OOD subsets.

Jane: That makes sense because sentence breaks allow many in-batch sequences to remain contiguous chunks from the same underlying web text context, which makes cross-pair negatives more contextually confusable.

Lu: The paper suggests that this unsupervised reward modeling can serve as a promising foundation for reducing the cost of human feedback and expanding the reach of alignment methods.

Meng: If we look at downstream tasks like best-of-N selection, these trained reward models substantially boost task accuracy, especially when you use larger actor sizes like Qwen two point five 7B.

Tom: And when compared against strong off-the-shelf reward model baselines, their web-trained models remain competitive and in several cases exceed them on both the ID task MATH500 and the OOD task Toxigen.

Jane: There’s also a point about mid-training; this method shows promise for using it as a beneficial initialization for subsequent curated preference training, recovering much of that instruct-level benefit when continuing from an RM checkpoint.

Lu: The implication here is that you might get a good starting point without paying for the initial human annotation effort.

Meng: From an engineering standpoint, this suggests we can build more scalable pipelines by substituting expensive human data with structured web text patterns for preference learning.

Tom: So, to wrap up on "Next-token Preference: Scaling Unsupervised Reward Modeling," the main point is that scaling reward models without costly and potentially unreliable human annotations is feasible.

Jane: They demonstrated that exploiting next-token continuation structure from large web corpora yields steady gains on RewardBench v1 and v2.

Lu: The core idea is converting raw text sequences into implicit preference supervision by sampling random breakpoints for prefix prompts and suffix continuations, which then feed into a Bradley–Terry objective.

Meng: And they stabilize this with the score-centering regularizer to keep the reward scores comparable across different prompts.

Title and authors: Tom: The results show training on eleven million tokens of math web text yields up to plus seven point seven average points on RewardBench v2, which helps boost accuracy in downstream tasks like GSM8K and MATH.

Jane: The authors confirm that this method shows promise for mid-training by providing a beneficial initialization for subsequent curated preference training, recovering much of the instruct-level benefit when continuing from an RM checkpoint.

Lu: To add to that, they suggest improving performance by increasing batch size in the online preference pairing scheme and allowing sentence breaks to create more contextually confusing negatives.

Meng: So, it's not just about one method; there are several pipeline choices they tested that showed where performance scales best, like the ID improvements scaling cleanly with B.

Tom: Overall, this paper suggests that a substantial fraction of the supervision traditionally attributed to curated human preferences is already latent in large text corpora.

Jane: This opens up a path toward more scalable and reliable reward modeling pipelines by using unsupervised reward models as a foundation.

Lu: The implication is that we can reduce the cost and expand the reach of human feedback-based alignment by leveraging these massive web text sources.

Meng: It's about making the process of getting good reward models less reliant on expensive, subjective human data collection.

Tom: That’s where we are with "Next-token Preference: Scaling Unsupervised Reward Modeling." We’ve seen how exploiting next-token continuation structure can yield steady gains without explicit human supervision.

Jane: It points toward a future where unsupervised reward models can serve as a promising foundation for alignment, making the process much more scalable.

Lu: This work shows that we don't need to rely solely on curated preference pairs to train effective reward models for large language models.

Meng: So, this means we might see more robust alignment methods that start from less expensive initial stages based on general web data patterns.

Tom: That’s the gist of "Next-token Preference: Scaling Unsupervised Reward Modeling." We've explored how to scale reward models using preference learning over document prefixes and suffixes drawn from large-scale web corpora.

Jane: It shows that we can achieve steady gains on RewardBench v1 and v2 by exploiting the next-token continuation structure, which is a key mechanism in language generation.

Lu: The authors show that training on eleven million tokens of math focused web data yields up to plus seven point seven average points on RewardBench v2 across various model sizes.

Meng: The paper also highlights that they can stabilize training under noisy supervision by incorporating a score-centering regularizer, which helps keep the reward scales consistent.

Tom: So, this approach moves away from relying only on explicit human labels and instead leverages the structure of large text corpora for implicit supervision.

Jane: It suggests that we can build more scalable and reliable reward modeling pipelines by using unsupervised reward models as a foundation for alignment.

The paper's summary: Tom: So we're talking about how you can train reward models without needing tons of human feedback by just looking at how text flows on the internet.

Jane: Exactly, this paper is about taking that idea and making it work for real scaling by using the structure of next-token continuation from massive web corpora.

Tom: They’re operationalizing reward-based scaling, or RBS, as preference learning over document prefixes and suffixes drawn from those huge web datasets. It's like they're turning raw text sequences into implicit supervision just by exploiting that natural flow of language on the web.

Jane: And the core idea is that instead of needing humans to label what’s good or bad, you can sample a random break in a text sequence to make a prefix and a suffix, and then treat the original continuation as the preferred response.

Tom: They use that structure to create online preference pairs. Then they train their reward model with a Bradley–Terry objective over those pairs, contrasting the chosen continuation against all the others as negatives.

Jane: But they don't just use that simple pairing; they add this score-centering regularizer to keep the reward outputs well-behaved across all prompts, which helps stabilize training when you're getting supervision from such a massive and noisy source like the web.

Tom: The experimental results show that training on just eleven million tokens of math-focused web text consistently gets you better rewards on RewardBench v2, up to plus seven point seven average points.

Jane: And those improvements aren't just theoretical; they translate into real utility when you use these reward models for things like best-of-N selection, which can substantially boost accuracy on tasks like GSM8K and MATH.

Tom: They also showed that this method works across different starting points, meaning it gives you a solid foundation even if your initial model isn't perfectly trained.

Jane: It suggests that a lot of the supervision we usually attribute to curated human preferences might actually be hidden in large text corpora already.

Tom: This opens up a path toward building reward modeling pipelines that are way more scalable and less dependent on expensive human feedback.

The paper's improvements: Tom: So we're talking about how they actually improve the training process using this method, like what they suggest for making it even better than just basic web data training.

Jane: Right, so besides getting those initial gains on RewardBench, they show a few specific tweaks to the pipeline that really help push the performance higher.

Tom: They looked at increasing the batch size in that online preference pairing scheme and found that performance scales cleanly with B up to sixteen, but then it plateaus after that point.

Jane: That's an interesting observation, because it tells us exactly how much more data you need to keep getting those rewards from this method.

Tom: Then there was the thing about allowing sentence breaks in the prefix and suffix splitting, and they found that this yields much larger gains over the starting point on RewardBench v2 across all subsets.

Jane: That makes sense because if you break sentences apart, you can make many different pairs from one piece of text context, which makes those negative examples more confusing for the model.

Tom: It's like they're tricking the AI into seeing a bigger pattern by creating more conflicting comparisons in that online training loop.

Jane: And they also talked about using this unsupervised reward model as a starting point for later, more curated training phases, which can recover most of the benefit you get from those expensive human labels later on.

Tom: That's useful because it means you don't have to start completely from scratch with zero prior knowledge when you eventually do need that high-quality instruction tuning.

Jane: So the big idea here is that this isn't just a single training method, it’s a whole pipeline of choices—batch size, how you split the text—to get the best possible result from unsupervised data.

Tom: And it really hammers home that this approach moves away from needing expensive human annotations and points toward more scalable ways to build reliable reward models for LLMs.

Conclusion: Tom: So we're wrapping up on "Next-token Preference: Scaling Unsupervised Reward Modeling," which shows that you can scale reward models without relying heavily on expensive human labeling by just using the structure of web text.

Jane: Exactly, it’s about finding that supervision in the way language naturally flows across massive datasets, which is a really clever way to build things up.

Tom: The big implication is that we don't need those huge teams dedicated to manual preference collection to get a solid foundation for reward models.

Jane: It suggests that we can have more scalable and reliable alignment pipelines without the cost of constant human feedback loops.

Tom: And they did show how this foundation helps downstream tasks like best-of-N selection, boosting accuracy on things like GSM8K and MATH when you use larger AI models.

Jane: Plus, it gives us a useful starting point for other training stages, meaning we can save time by not having to start from a completely blank slate.

Tom: Lu, what’s your take on the creative possibilities here?

Lu: I think this opens up a whole new way to think about preference signals; instead of just looking at human choices, we're looking at the statistical patterns in how text is constructed online.

Meng: From an engineering standpoint, it means we can build more robust reward models faster, using readily available web data instead of waiting for massive annotation efforts to finish.

Lalam: For me, this means the AI can start learning better alignment principles from a much broader context right away, which could make the entire culture around how we fine-tune models feel much more intuitive and less fragile.

Tom: That's a huge vision there, Lalam. Jane, what’s your final thought on this whole idea?

Jane: I just think this paper shows that the latent supervision in big text is really substantial, proving that we can build reliable tools without always needing those costly human label efforts upfront.

Tom: So to sum up "Next-token Preference: Scaling Unsupervised Reward Modeling," it proves you can get steady gains on RewardBench v1 and v2 by exploiting next-token continuation structure from large web corpora.

Jane: It’s a solid piece of work showing a practical path toward making AI alignment methods much more scalable.

Tom: We've seen how exploring these massive text patterns can yield steady rewards, and now we can look at what’s next in scaling these reward models.

Harvard University · Cornell University · Microsoft Research

cs.LG

Submitted: 2026-02-11

Updated: 2026-10-12

License: http://creativecommons.org/licenses/by-nc-nd/4.0/

Importance score: 90/100

The gist: The gist: Reward models trained without human supervision can yield steady gains on RewardBench v1 and v2 by exploiting next-token continuation structure from large-scale web corpora.

Key concepts

Next-Token Preference (NTP)

An unsupervised reward modeling technique that creates preference pairs directly from the continuation structure of raw web text. It treats the original continuation following a prefix as the preferred response and other nearby continuations as negative examples, requiring no human or model annotations.

Preference Learning over Document Prefixes

The core idea is to operationalize reward-based scaling by sampling random breakpoints in large text sequences to create prompts (prefixes) and continuations (suffixes). This allows the model to implicitly learn which continuation is 'better' based on its context within the web data.

Bradley–Terry Objective with Score-Centering Regularizer

The training loss function used to teach the reward model. It minimizes negative log-likelihood by treating the chosen continuation as superior. A regularization term is added to penalize large reward outputs, ensuring a stable and comparable score scale across different prompts.

Transferable Preference Signals

The finding that web text contains implicit preference signals that can be used for training reward models without explicit human labels. This suggests that much of the supervision traditionally requiring costly human annotations is already present in large, uncurated text corpora.

Terminology

• Accuracy — 1× in this paper · explained in 1 episode

Summary

Improvements for AI systems

Abstract

Sources

• A General Language Assistant as a Laboratory for Alignment
• Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback
• Constitutional AI: Harmlessness from AI Feedback
• Open Problems and Fundamental Limitations of Reinforcement Learning from Human Feedback
• Limits to scalable evaluation at the frontier: LLM as Judge won't beat twice the data
• Impact of Preference Noise on the Alignment Performance of Generative Language Models
• Improving alignment of dialogue agents via targeted human judgements
• Scaling Laws for Reward Model Overoptimization
• Scalable agent alignment via reward modeling: a research direction
• Natural Emergent Misalignment from Reward Hacking in Production RL
• The Effects of Reward Misspecification: Mapping and Mitigating Misaligned Models
• Discovering Language Model Behaviors with Model-Written Evaluations
• EvoLM: In Search of Lost Language Model Training Dynamics
• Towards Data-Centric RLHF: Simple Metrics for Preference Dataset Comparison
• HybridFlow: A Flexible and Efficient RLHF Framework
• Secrets of RLHF in Large Language Models Part II: Reward Modeling
• Investigating Non-Transitivity in LLM-as-a-Judge
• Justice or Prejudice? Quantifying Biases in LLM-as-a-Judge
• Secrets of RLHF in Large Language Models Part I: PPO
• Unmasking and Improving Data Credibility: A Study with Datasets for Training Harmless Language Models

Related papers

• Polynomial-Augmented Neural Networks (PANNs) with Weak Orthogonality Constraints for Enhanced Function and PDE Approximation
• AIRL-S: Unifying Reinforcement Learning and Search-Based Test-Time Scaling via Adversarial Inverse Reinforcement Learning
• Transformers as Bayesian In-Context Experimenters: Smoothness-Adaptive Efficient ATE Estimation
• Convergence issues in Relational Concept Analysis based on AOC-posets
• Beliefs Beyond Posteriors: The Predictive Cost of ELBO Binding
• Understanding Diffusion Models via Ratio-Based Function Approximation with SignReLU Networks