arXiv subjects

Adam Block

Publications and source records attributed to Adam Block.

At least 19 recordsLinked to original sources

arXiv2026

MarkTune: Improving the Quality-Detectability Trade-off in Model-Embedded LLM Watermarking

Language model watermarking schemes fall into two broad categories: inference-time methods, which modify the decoding process at generation time, and model-embedded methods, which encode a secret signal directly into the model weights. Inference-time watermarking methods incur additional inference overhead and are not applicable in the increasingly prevalent open-weight model settings. In contrast, model-embedded watermarks introduce no additional inference latency-generation uses the standard sampling pipeline-and are particularly well-suited to open-weight settings. However, existing model-embedded approaches such as GaussMark face a fundamental quality-detectability trade-off: achieving strong detection power typically requires weight perturbations that noticeably degrade generation quality. Our key insight is that regularizing toward a high-quality reference distribution decouples quality preservation from proximity to the base model, enabling a better quality-detectability trade-off. Building on this insight, we introduce MarkTune, a theoretically grounded on-policy fine-tuning framework that uses the GaussMark detection statistic as a reward while preserving generation quality through cross-entropy on high-quality human text or KL divergence to a stronger oracle model. Empirically, MarkTune substantially improves the quality-detectability frontier of GaussMark, approaching the detectability of strong inference-time schemes while preserving generation quality and downstream task performance. Marktune also remains robust under the evaluated text-editing and open-weight attacks and generalizes across datasets, retaining substantial detection power on corpora unseen during training. Code is available at https://github.com/zhyzmath/MarkTune-LLM-Watermarking.

cs.LG↗

arXiv2026

ReAPO: Retry-Aware Policy Optimization for Efficient Exploration in LLM Reasoning

RLVR improves language-model reasoning, but commonly relies on independent stochastic rollouts for exploration, which can repeatedly sample similar failed strategies. Prior work has seen success with approaches that guide exploration by introducing dependence into the sampling process by including incorrect solutions in the context for subsequent attempts. Still, these methods introduce the additional question of how to assign credit to each attempt. We propose to view the credit assignment through the lens of Stochastic Shortest Paths, a framework in classical reinforcement learning that seeks to minimize the expected cumulative cost to reach a terminal state (in our case, a verified successful response). We show that, under suitable simplifying assumptions, this perspective naturally leads to a simple modification to GRPO, which we call Retry-Aware Policy Optimization (ReAPO). ReAPO allocates additional generation to failed trajectories, allowing successful retries to be reinforced without rewarding preceding failures, and assigns credit in a theoretically principled way to each attempt. Across Qwen3 and OLMo-2 models from 0.6B to 4B and eight mathematical and scientific reasoning benchmarks, ReAPO improves first-attempt performance over GRPO as well as natural recent baseline algorithms, while additional ablations demonstrate the utility of both serial sampling and ReAPO's credit assignment. On Qwen3-0.6B, ReAPO also outperforms GRPO with twice the group size or twice the training steps, reaching their peak accuracy with 39% and 31% fewer generated tokens, respectively. Taken together, our results suggest that ReAPO is a simple yet effective approach for improving LM reasoning performance through improved exploration and retry-aware credit assignment.

cs.LG↗

arXiv2026

Understanding the Benefits of Scaffolds through Computational Separations

Inference-time harnesses, i.e., external procedures that call a model many times and keep state across calls, can substantially improve model capabilities. In this paper, we investigate the benefits of using such harnesses and the extent to which these benefits can be internalized through training. More precisely, we conceptualize a harness $\mathcal{A}$ as an arbitrary wrapper around a base model $M$ and attempt to distill a student model from the harness's traces. Empirically we find that, even with simple scaffolds and small models, the inference-time harness significantly outperforms a distilled student. We then prove that such a gap is due to a fundamental representational limitation: inference-time harnesses can compute a closure of their per-call primitive that is $\NL$-complete even when the primitive is a trivial $\AC^0$ lookup, and that no fixed-depth forward pass can represent under the widely believed conjecture that $\NL\not\subseteq\TC^0$. Our theory suggests that expanding the representational capacity of the student relative to the task at hand, e.g., by enlarging the model, making the task easier, or lengthening its chain-of-thought, is necessary to close the gap, which we verify empirically. Finally, we scale to an agentic harness on competition math, where the same gap reappears. Overall, our results show that a harness confers on the base model a profound gain in representational power, one that training a same-class student on its traces cannot transfer back into the weights.

cs.LG↗

arXiv2026

LLM Evaluation is a Forecasting Problem

Standard procedure in machine learning is to randomly split train and test data, emulating IID draws from one underlying data distribution. In practice, models are trained on past data and deployed in the future, which may come from distinct distributions. In this paper, we find that web text is indeed highly non-stationary, as text distributions shift over time. We further observe that model performance changes continuously over time across diverse benchmarks. Consequently, future model performance is better predicted by measurements made on recent evaluation data, and is best predicted by time series forecasters applied to historic performance rather than by historic performance itself. Choosing which model to deploy based on forecasted accuracy enables more reliable selection of performant models, thus improving downstream accuracy in the regime practitioners care about: deployment. Taken together, our experiments suggest that time should be treated as a fundamental variable in LLM evaluation. We release our code and data.

cs.LG↗

arXiv2026

Oracle-Efficient and Parameter-Free Agnostic Smoothed Online Learning

Online learning is an attractive framework in many domains because it permits well-defined learning even when data are dependent or chosen adversarially. This generality, however, comes at a steep price, introducing significant statistical and computational barriers. Recently, smoothed online learning has emerged as a promising framework that interpolates between the fully adversarial and fully stochastic settings by assuming that the conditional law of each covariate has density at most $1/σ$ with respect to some fixed base measure $μ$, and it is known to match the statistical and computational guarantees of classical learning while still allowing for much of the flexibility of online learning. However, existing oracle-efficient algorithms require either (i) sampling access to the base measure $μ$ or (ii) labels that are perfectly predicted by a fixed hypothesis. Both assumptions limit the applicability of these algorithms, in contrast to statistical learning, where empirical risk minimization (ERM) learns efficiently in the agnostic setting without any knowledge of the data distribution. We show that neither assumption is necessary, giving the first oracle-efficient algorithm that achieves sublinear regret in the agnostic setting without knowledge of $μ$. Our algorithm, based on Gaussian Follow-The-Perturbed-Leader, is parameter-free: it requires no knowledge of $μ$, the smoothing parameter $σ$, or the horizon $T$, and it achieves regret $\widetilde O(d\sqrt{T/σ})$ for binary classes of VC dimension $d$ with a single call to an ERM oracle per round, which is optimal up to a $\sqrt{d}$ factor. En route to establishing the regret bound, we introduce several new techniques that may be of independent interest.

cs.LG↗

arXiv2026

Training for the Model You Return: Improving Optimization for Iterate-Averaged Language Models

Many modern Language Model (LM) pipelines return an averaged model, such as an exponential moving average of the training iterates, rather than the final iterate itself. This raises a fundamental question: given that we will return an iterate average, how should we change training to improve the performance of this average? We study this question by formulating optimizer design for the iterate-average estimator as an optimal-control problem. In a continuous-time stochastic quadratic model, we solve for the control strategy that minimizes the error of the returned average subject to a penalty on the size of the intervention. A practical approximation to this controller yields PACE, a lightweight wrapper around AdamW that pulls the live weights toward their exponential moving average with a clipped, per-coordinate control strength. We prove that a stylized version of PACE converges at the standard stochastic convex optimization rate, up to a factor depending on the averaging rule, while in the quadratic setting it can strictly improve the limiting squared error of the iterate-average estimator and can do so by an arbitrarily large factor on some instances. Empirically, our results suggest that PACE improves over AdamW and EMA-evaluated AdamW in supervised fine-tuning of 1-2B parameter LMs and in GPT-2 pretraining on FineWeb for a wide range of learning rates, decay schedules, and other hyperparameters.

cs.LG↗

arXiv2026

Behavior Cloning is Not All You Need: The Optimality of On-Policy Distillation for Noisy Expert Feedback

Imitation Learning is a natural framework for learning in sequential decision-making systems and has emerged as the dominant paradigm through which we understand language model training. A central puzzle is that, while in theory offline IL can be horizon-free and optimal, in practice online methods such as on-policy distillation often outperform offline methods such as supervised fine-tuning. We propose a noisy expert model to explain this gap, in which the learner only has access to a noisy version of the expert's policy, but wishes to compete against the reward achieved by a clean expert, motivated by the fact that in many applications, e.g. training language models to perform long chains of thought, the expert is often imperfect. In this setting, we show a sharp separation between offline and online IL. Offline learning from noisy trajectories is fundamentally hard: to compete with the clean expert, the sample complexity must grow exponentially, in contradistinction to the clean expert setting where no explicit horizon dependence exists. In contrast, we prove that online interaction with the noisy expert via a novel variant of OPD enables polynomial dependence on the horizon in general. We further show that, under a natural condition on the expert noise distribution, which we show to be necessary for any horizon-free sample complexity, one can obtain such a guarantee, although our proposed algorithm sacrifices statistical efficiency in its dependence on the size of the policy class. Our analysis leads to an alternative loss function that is commonly considered empirically for LM training. We further provide algorithms and lower bounds, and extend our results to the more realistic setting of unknown corruption when the clean expert is deterministic, thereby providing a theoretical foundation for why OPD can outperform SFT when training language models from imperfect teachers.

cs.LG↗

arXiv2026

From Curiosity to Caution: Mitigating Reward Hacking for Best-of-N with Pessimism

Inference-time compute scaling has emerged as a powerful paradigm for improving language model performance on a wide range of tasks, but the question of how best to use the additional compute remains open. A popular approach is BoN sampling, where N candidate responses are generated, scored according to a reward model, and the highest-scoring response is selected. While this approach can improve performance, it is vulnerable to reward hacking, where performance degrades as N increases due to the selection of responses that exploit imperfections in the reward model instead of genuinely improving generation quality. Prior attempts to mitigate reward hacking, via stronger reward models or heavy-handed distributional regularization, either fail to fully address over-optimization or are too conservative to exploit additional compute. In this work, we explore the principle of pessimism in RL, which uses lower confidence bounds on value estimates to avoid OOD actions with uncertain reward estimates. Our approach, termed as caution, can be seen as the reverse of curiosity: where curiosity rewards prediction error as a signal of novelty, caution penalizes prediction error as a signal of distributional uncertainty. Practically, caution trains an error model on typical responses and uses its prediction error to lower reward estimates for atypical ones. Our extensive empirical evaluation demonstrates that caution is a simple, computationally efficient approach that substantially mitigates reward hacking in BoN sampling. We also provide a theoretical analysis in a simplified linear setting, which shows that caution provably improves over the standard BoN approach. Together, our results not only establish caution as a practical solution to reward hacking, but also provide evidence that curiosity-based approaches can be a general OOD detection technique in LLM settings.

cs.LG↗

arXiv2026

Revisiting the (Sub)Optimality of Best-of-N for Inference-Time Alignment

Best-of-N (BoN) sampling is a widely used inference-time alignment method for language models, whereby N candidate responses are sampled from a reference model and the one with the highest predicted reward according to a learned reward model is selected. Despite its widespread practical use, recent theoretical work has suggested that it is statistically suboptimal and vulnerable to reward hacking, the process by which models exploit weaknesses in the learned reward model to achieve high estimated reward without genuinely improving performance. We revisit this question under assumptions that more closely reflect practice than that of prior work. In particular, in contradistinction to earlier analyses that focused on expected true reward, which may not be meaningful in many practical settings, we investigate how inference-time alignment affects the win-rate, a pairwise comparison-based metric more closely aligned with how reward models are trained and evaluated in practice. We demonstrate that, under minimal conditions on the quality of the reference model and learned reward model, properly tuned BoN is both computationally and statistically optimal in achieving high win-rate, partially explaining its widespread practical success. Because BoN remains susceptible to reward-hacking in this setting, we propose a simple and practical variant that provably eliminates reward-hacking while maintaining optimal statistical performance. Finally, we show that prior approaches are provably suboptimal when considering win-rate, highlighting the importance of choosing appropriate objectives when analyzing inference-time alignment methods.

cs.LG↗

arXiv2026

Partition Function Estimation under Bounded f-Divergence

We study the statistical complexity of estimating partition functions given sample access to a proposal distribution and an unnormalized density ratio for a target distribution. While partition function estimation is a classical problem, existing guarantees typically rely on structural assumptions about the domain or model geometry. We instead provide a general, information-theoretic characterization that depends only on the relationship between the proposal and target distributions. Our analysis introduces the integrated coverage profile, a functional that quantifies how much target mass lies in regions where the density ratio is large. We show that integrated coverage tightly characterizes the sample complexity of multiplicative partition function estimation and provide matching lower bounds. We further express these bounds in terms of $f$-divergences, yielding sharp phase transitions depending on the growth rate of f and recovering classical results as a special case while extending to heavy-tailed regimes. Matching lower bounds establish tightness in all regimes. As applications, we derive improved finite-sample guarantees for importance sampling and self-normalized importance sampling, and we show a strict separation between the complexity of approximate sampling and counting under the same divergence constraints. Our results unify and generalize prior analyses of importance sampling, rejection sampling, and heavy-tailed mean estimation, providing a minimal-assumption theory of partition function estimation. Along the way we introduce new technical tools including new connections between coverage and $f$-divergences as well as a generalization of the classical Paley-Zygmund inequality.

stat.ML↗

arXiv2025

The Coverage Principle: How Pre-Training Enables Post-Training

Language models demonstrate remarkable abilities when pre-trained on large text corpora and fine-tuned for specific tasks, but how and why pre-training shapes the success of the final model remains poorly understood. Notably, although pre-training success is often quantified by cross-entropy loss, cross-entropy can be a poor predictor of downstream performance. Instead, we provide a theoretical perspective on this relationship through the lens of \emph{coverage}, which quantifies the probability mass the pre-trained model places on high-quality responses and which is necessary and sufficient for post-training and test-time scaling methods such as Best-of-N to succeed. Our main results develop an understanding of \emph{the coverage principle}, a phenomenon whereby next-token prediction (more generally, maximum likelihood) implicitly optimizes toward a model with good coverage. In particular, we uncover a mechanism that explains the power of coverage in predicting downstream performance: \emph{coverage generalizes faster than cross-entropy}, avoiding spurious dependence on problem-dependent parameters such as the sequence length. We also study practical algorithmic interventions with provable benefits for improving coverage, including (i) model/checkpoint selection procedures, (ii) gradient normalization schemes, and (iii) test-time decoding strategies.

stat.ML↗

arXiv2025

A Theory of Learning with Autoregressive Chain of Thought

For a given base class of sequence-to-next-token generators, we consider learning prompt-to-answer mappings obtained by iterating a fixed, time-invariant generator for multiple steps, thus generating a chain-of-thought, and then taking the final token as the answer. We formalize the learning problems both when the chain-of-thought is observed and when training only on prompt-answer pairs, with the chain-of-thought latent. We analyze the sample and computational complexity both in terms of general properties of the base class (e.g. its VC dimension) and for specific base classes such as linear thresholds. We present a simple base class that allows for universal representability and computationally tractable chain-of-thought learning. Central to our development is that time invariance allows for sample complexity that is independent of the length of the chain-of-thought. Attention arises naturally in our construction.

stat.ML↗

arXiv2025

EMA Without the Lag: Bias-Corrected Iterate Averaging Schemes

Stochasticity in language model fine-tuning, often caused by the small batch sizes typically used in this regime, can destabilize training by introducing large oscillations in generation quality. A popular approach to mitigating this instability is to take an Exponential moving average (EMA) of weights throughout training. While EMA reduces stochasticity, thereby smoothing training, the introduction of bias from old iterates often creates a lag in optimization relative to vanilla training. In this work, we propose the Bias-Corrected Exponential Moving Average (BEMA), a simple and practical augmentation of EMA that retains variance-reduction benefits while eliminating bias. BEMA is motivated by a simple theoretical model wherein we demonstrate provable acceleration of BEMA over both a standard EMA and vanilla training. Through an extensive suite of experiments on Language Models, we show that BEMA leads to significantly improved convergence rates and final performance over both EMA and vanilla training in a variety of standard LM benchmarks, making BEMA a practical and theoretically motivated intervention for more stable and efficient fine-tuning.

cs.LG↗

arXiv2025

Is Best-of-N the Best of Them? Coverage, Scaling, and Optimality in Inference-Time Alignment

Inference-time computation offers a powerful axis for scaling the performance of language models. However, naively increasing computation in techniques like Best-of-N sampling can lead to performance degradation due to reward hacking. Toward a theoretical understanding of how to best leverage additional computation, we focus on inference-time alignment, which we formalize as the problem of improving the quality of responses drawn from a pre-trained policy, given a prompt of interest and access to an imperfect reward model. We analyze the performance of inference-time alignment algorithms in terms of (i) response quality, and (ii) compute, and provide new results that highlight the importance of the pre-trained policy's coverage over high-quality responses for performance and compute scaling: 1. We show that Best-of-$N$ alignment with an ideal choice for $N$ can achieve optimal performance under stringent notions of coverage, but provably suffers from reward hacking when $N$ is large, and fails to achieve tight guarantees under more realistic coverage conditions. 2. We introduce $\texttt{InferenceTimePessimism}$, a new algorithm which mitigates reward hacking through deliberate use of inference-time compute, implementing the principle of pessimism in the face of uncertainty via rejection sampling; we prove that its performance is optimal and does not degrade with $N$, meaning it is scaling-monotonic. We complement our theoretical results with an experimental evaluation that demonstrate the benefits of $\texttt{InferenceTimePessimism}$ across a variety of tasks and models.

cs.AI↗

arXiv2025

Stellar tidal streams around nearby spiral galaxies with deep imaging from amateur telescopes

Tidal interactions between massive galaxies and their satellites are fundamental processes in a Universe with L-Cold Dark Matter cosmology, redistributing material into faint features that preserve records of past galactic interactions. While stellar streams in the Local Group impressively demonstrate satellite disruption, they do not constitute a statistically significant sample. Constructing a substantial catalog of stellar streams beyond the Local Group remains challenging due to the difficulties in obtaining deep, wide-field images of galaxies. Despite their potential to illuminate dark matter distribution and galaxy formation processes, stellar streams remain underutilized as cosmological probes. The Stellar Tidal Stream Survey (STSS) addresses this observational gap by leveraging amateur telescopes to obtain deep, scientific-grade images of galactic outskirts, capable of building a more statistically meaningful sample of stellar streams. Over the last decade, the STSS has acquired deep (up to surface brightness limit 28.3 mag/arcsec^2 in the r-band), wide-field images of 15 nearby Milky Way analog galaxies using a network of robotic amateur telescopes, avoiding the issues associated with "mosaicing" smaller images taken with a professional telescope. Our survey has revealed a diverse range of previously unreported faint features related to dwarf satellite accretion - including stellar streams, shells, and umbrella-like structures. We discover an ultra-diffuse galaxy (NGC150-UDG1), which shows hints of tidal tails. The STSS demonstrates the suitability of modern amateur telescopes to detect and study faint, diffuse structures in large fields around nearby spiral galaxies. Their economic and accessibility advantages enable larger samples with deep imaging, essential for testing galaxy formation models and constraining the properties of minor merger events in the local Universe.

astro-ph.GA↗

arXiv2025

Computational-Statistical Tradeoffs at the Next-Token Prediction Barrier: Autoregressive and Imitation Learning under Misspecification

Next-token prediction with the logarithmic loss is a cornerstone of autoregressive sequence modeling, but, in practice, suffers from error amplification, where errors in the model compound and generation quality degrades as sequence length $H$ increases. From a theoretical perspective, this phenomenon should not appear in well-specified settings, and, indeed, a growing body of empirical work hypothesizes that misspecification, where the learner is not sufficiently expressive to represent the target distribution, may be the root cause. Under misspecification -- where the goal is to learn as well as the best-in-class model up to a multiplicative approximation factor $C\geq 1$ -- we confirm that $C$ indeed grows with $H$ for next-token prediction, lending theoretical support to this empirical hypothesis. We then ask whether this mode of error amplification is avoidable algorithmically, computationally, or information-theoretically, and uncover inherent computational-statistical tradeoffs. We show: (1) Information-theoretically, one can avoid error amplification and achieve $C=O(1)$. (2) Next-token prediction can be made robust so as to achieve $C=\tilde O(H)$, representing moderate error amplification, but this is an inherent barrier: any next-token prediction-style objective must suffer $C=Ω(H)$. (3) For the natural testbed of autoregressive linear models, no computationally efficient algorithm can achieve sub-polynomial approximation factor $C=e^{(\log H)^{1-Ω(1)}}$; however, at least for binary token spaces, one can smoothly trade compute for statistical power and improve on $C=Ω(H)$ in sub-exponential time. Our results have consequences in the more general setting of imitation learning, where the widely-used behavior cloning algorithm generalizes next-token prediction.

cs.LG↗

arXiv2025

Small Loss Bounds for Online Learning Separated Function Classes: A Gaussian Process Perspective

In order to develop practical and efficient algorithms while circumventing overly pessimistic computational lower bounds, recent work has been interested in developing oracle-efficient algorithms in a variety of learning settings. Two such settings of particular interest are online and differentially private learning. While seemingly different, these two fields are fundamentally connected by the requirement that successful algorithms in each case satisfy stability guarantees; in particular, recent work has demonstrated that algorithms for online learning whose performance adapts to beneficial problem instances, attaining the so-called small-loss bounds, require a form of stability similar to that of differential privacy. In this work, we identify the crucial role that separation plays in allowing oracle-efficient algorithms to achieve this strong stability. Our notion, which we term $ρ$-separation, generalizes and unifies several previous approaches to enforcing this strong stability, including the existence of small-separator sets and the recent notion of $γ$-approximability. We present an oracle-efficient algorithm that is capable of achieving small-loss bounds with improved rates in greater generality than previous work, as well as a variant for differentially private learning that attains optimal rates, again under our separation condition. In so doing, we prove a new stability result for minimizers of a Gaussian process that strengthens and generalizes previous work.

cs.LG↗

arXiv2025

Rate of convergence of the smoothed empirical Wasserstein distance

Consider an empirical measure $\mathbb{P}_n$ induced by $n$ iid samples from a $d$-dimensional $K$-subgaussian distribution $\mathbb{P}$ and let $γ= N(0,σ^2 I_d)$ be the isotropic Gaussian measure. We study the speed of convergence of the smoothed Wasserstein distance $W_2(\mathbb{P}_n * γ, \mathbb{P}*γ) = n^{-α+ o(1)}$ with $*$ being the convolution of measures. For $K<σ$ and in any dimension $d\ge 1$ we show that $α= {1\over2}$. For $K>σ$ in dimension $d=1$ we show that the rate is slower and is given by $α= {(σ^2 + K^2)^2\over 4 (σ^4 + K^4)} < 1/2$. This resolves several open problems in [GGNWP20], and in particular precisely identifies the amount of smoothing $σ$ needed to obtain a parametric rate. In addition, for any $d$-dimensional $K$-subgaussian distribution $\mathbb{P}$, we also establish that $D_{KL}(\mathbb{P}_n * γ\|\mathbb{P}*γ)$ has rate $O(1/n)$ for $K<σ$ but only slows down to $O({(\log n)^{d+1}\over n})$ for $K>σ$. The surprising difference of the behavior of $W_2^2$ and KL implies the failure of $T_{2}$-transportation inequality when $σ< K$. Consequently, it follows that for $K>σ$ the log-Sobolev inequality (LSI) for the Gaussian mixture $\mathbb{P} * N(0, σ^{2})$ cannot hold. This closes an open problem in [WW+16], who established the LSI under the condition $K<σ$ and asked if their bound can be improved.

math.PR↗