Kullback-Leibler Divergence

We establish novel structural and statistical results for entropy-regularized min-max inverse reinforcement learning (Min-Max-IRL) with linear reward classes in finite-horizon MDPs with Borel state and action spaces. On the structural side, we show that maximum likelihood estimation (MLE) and Min-Max-IRL are equivalent at the population level, and at the empirical level under deterministic dynamics. On the statistical side, exploiting pseudo-self-concordance of the Min-Max-IRL loss, we prove that both the trajectory-level KL divergence and the squared parameter error in the Hessian norm decay at the fast rate $\mathcal{O}(n^{-1})$, where $n$ is the number of expert trajectories. Our guarantees apply under misspecification and require no exploration assumptions. We further extend reward-identifiability results to general Borel spaces and derive novel results on the derivatives of the soft-optimal value function with respect to reward parameters.

Knowledge distillation is widely used to improve generalization in practice, yet its theoretical understanding remains elusive. In the standard distillation setting, a teacher model provides soft predictions to guide the training of a student model. We model teacher and student training as coupled stochastic processes and introduce a distillation divergence, defined as the Kullback-Leibler divergence between these two stochastic kernels. Within this framework, we derive two generalization bounds for the student model relative to the teacher's generalization gap: an upper bound under a sub-Gaussian assumption via algorithmic stability, and a lower bound under a central condition with sharper dependence on the distillation divergence. We further develop a loss-sharpness-aware bound with an explicit tightness regime, showing that the teacher's local flatness can strictly tighten the bound. Additionally, in a linear Gaussian case study, the distillation divergence admits an interpretable decomposition into bias, variance, and rank-bottleneck costs, yielding practical guidance for distillation design.

Classical communication systems fail not only through random noise but also when transmitter and receiver use incompatible operational codebooks. Variational autoencoders (VAEs) train an encoder $q_φ$ and decoder $p_θ$ jointly, and practitioners treat the resulting latent space as a discrete code -- for clustering, conditional generation, and mechanistic interpretability. Yet standard VAE diagnostics -- ELBO, active units, mutual information, and code histograms -- certify only whether this code is used, never whether the decoder reads each latent under the encoder's code. We close this gap with the neural codebook channel $K_{e\to d}(j\mid i)$, a coupled encoder-decoder diagnostic whose off-diagonal mass is bounded by an architecture-free Bernoulli-KL certificate $d_{\mathrm{bin}}(1-\mathcal{A} \,\|\, \barη_p) \le \barΔ$ controlled by the variational gap. The certificate is the operational specialization of the classical KL chain rule under disintegration to the encoder-decoder disagreement event, complemented by a constructive marginal-impossibility result: no combination of marginal histograms, entropies, active-code counts, or mutual information determines $K_{e\to d}$. We audit the certificate on four sklearn datasets (finite-grid exact, 5/5 seeds, 20/20 pairs satisfy the bound), a 2D model where the bound is non-vacuous at $2.71\times$ the observed disagreement and the four-term identity closes within $10^{-4}$, MNIST under importance-sampling control, and a VQ-VAE attaining the predicted limit $\hat{\mathcal{A}}=1.000$. The package $(K_{e\to d}, \mathcal{A}, R_{\mathrm{eff}}, R, \mathrm{AU})$ is an audit-ready reporting unit. More broadly, the framework makes mismatched decoding -- a failure mode classical communication theory named decades ago -- visible inside a single deep generative model.

We study the fundamental and timely problem of learning long sequences in autoregressive modeling and next-token prediction under model misspecification, measured by the joint Kullback--Leibler (KL) divergence. Our goal is to characterize how the sequence horizon $H$ affects both approximation and estimation errors in this joint-distribution, sequence-level regime. By establishing matching upper and lower bounds, we provide, to our knowledge, the first complete characterization of long-horizon error behavior under the natural joint KL objective, with improved rates and optimality justification relative to existing work. On the approximation side, we show that joint KL admits a horizon-free approximation factor, in sharp contrast to Hellinger-based analyses that exhibit an $Ω(H)$ dependence for computationally efficient methods; this isolates the choice of divergence as the source of approximation amplification. On the estimation side, we prove a fundamental information-theoretic lower bound of order $Ω(H)$ that holds for both decomposable policy classes and fully shared policies, matching the $\widetilde O(H)$ upper bounds achieved by computationally efficient algorithms. Our analysis clarifies the landscape of recent autoregressive learning results by aligning the log-loss training objective, the sequence-level evaluation metric, and the approximation metric {\color{black}through a sharp joint-KL oracle theory}. We further show that these joint-KL guarantees imply policy learning regret bounds at rates matching prior imitation learning literature.

Self-supervised learning (SSL) is recognized as an essential tool for building foundation models for Artificial Intelligence applications. The advances in SSL have been made thanks to vigorous arguments about the principles of SSL and through extensive empirical research. The aim of this paper is to contribute to the development of the underpinning theory of SSL, focusing on the deep clustering approach. By analogy to supervised learning, we formulate SSL as K-L divergence optimization. The mode collapse is prevented by imposing an optimisation constraint on the teacher distribution. This leads to normalization using inverse cluster priors. We show that using Jensen inequality this normalization simplifies to the popular batch centering procedure. Distillation and centering are common {heuristics-based} practices in SSL, {but our work underpins them theoretically.} The theoretical model developed not only supports specific existing successful SSL methods, but also suggests directions for future investigations.

Reinforcement Learning with Verifiable Rewards (RLVR) has become the standard paradigm for LLM mathematical reasoning, with Group Relative Policy Optimization (GRPO) serving as the dominant algorithm. We identify two overlooked inefficiencies inherent in GRPO. First, a fixed KL coefficient overly restricts policy exploration at moments when the model needs to diverge significantly from the reference policy. Second, uniform question sampling overlooks that moderately difficult problems produce the most informative gradient signals. We propose FG-ExPO, short for Frontier-Guided Exploration-Prioritized Policy Optimization, which integrates two lightweight components. Accuracy-Conditioned KL Scaling (AKL) adjusts the KL penalty strength through a smooth nonlinear function of batch average accuracy, loosening the constraint when the model performs poorly and strengthening it when the model achieves satisfactory results. Gaussian Curriculum Sampling (GCS) assigns sampling weights to questions following a Gaussian distribution centered at a moderate accuracy level around 0.5, focusing model training on its learning frontier. We conduct evaluations on DeepSeek-R1-Distill-Qwen-1.5B and Qwen3-8B-Base across six mainstream mathematical reasoning benchmarks. Experimental results demonstrate that FG-ExPO consistently outperforms vanilla GRPO. It delivers an absolute improvement of 13.34 on the AIME 2025 pass@32 metric, rising from 63.33 percent to 76.67 percent, and obtains an average pass@32 gain of 2.66 on the 8B model. The substantially larger performance gains observed on pass@32 compared to pass@1 verify that FG-ExPO enlarges the model's effective exploration space under a fixed inference budget.

On-policy distillation (OPD) and on-policy self-distillation (OPSD) have emerged as promising post-training methods for large language models, offering dense token-level supervision on trajectories sampled from the model's own policy. However, existing results on their effectiveness remain mixed: while OP(S)D has shown promise in system prompt and knowledge internalization, recent studies also report instability and degradation. In this work, we present a comprehensive empirical study of when OPD and OPSD work, when they fail, and why. We find that OPD on mathematical reasoning is highly sensitive to teacher choice and loss formulation, whereas OPSD fails in our tested settings due to test-time absence of instance-specific privileged information (PI). In contrast, OPSD is effective when PI represents a shared latent rule, such as a system prompt or alignment preference. We identify three failure mechanisms: (1) distribution mismatch between teacher and student caused by conditioning on student-generated prefixes, (2) optimization instability from biased TopK reverse-KL gradients, and (3) an OPSD-specific limitation where the student learns a PI-free policy that aggregates PI-conditioned teachers, which is insufficient when PI is instance-specific. We further show that stop-gradient TopK objectives, RLVR-adapted teachers, and SFT-stabilized students mitigate these failures.

This paper proposes a paradigm shift linking machine unlearning directly to the structure of the data distributions rather than a mere update of the neural network parameters. We show that inferring these distributions with precision enables distilling the exact unlearning signal induced by the modeling. Theoretical bounds on the Kullback-Leibler divergence from the ideal retrained model to our unlearned model, under verifiable admissibility criterion, reveal the soundness of our framework. This method is experimentally validated over three forgetting scenarios as reaching the closest classifier to the ideal retrained model when compared to competitors.

We propose a closed-form spectral framework for relative log-density estimation in linearly parameterized probabilistic models, including unnormalized and conditional models. This is achieved by representing the Kullback-Leibler (KL) divergence as an integral of weighted chi-squared divergences, converting KL estimation into a family of least-squares problems. We derive an explicit spectral formula based only on first- and second-order feature moments, yielding closed-form estimators of both divergences and log-density potentials for fixed features. The framework extends to a broad class of f-divergences and can be combined with kernelization or feature learning with neural networks. We prove convergence guarantees for the resulting estimators and empirically compare them on synthetic data with optimization-based variational formulations, including logistic and softmax regression for normalized conditional models.

Trellis-coded quantization sets the current 2-bit post-training frontier for LLMs (QTIP), but pushing below the PTQ ceiling requires quantization-aware training, and QAT on a trellis is obstructed by the non-differentiable Viterbi argmax. We introduce BCJR-QAT, a relaxation that replaces the argmax with the BCJR forward-backward sum-product algorithm at temperature $T$, producing a soft codeword equal to the Boltzmann expectation over trellis paths, exactly differentiable, recovering the hard QTIP code as $T \to 0$, and mathematically identical to the transfer-matrix computation for a 1D Ising-like spin chain. We contribute (i) a fused Triton kernel making BCJR tractable on a single consumer GPU ($6.57\times$ speedup, fp32 parity); (ii) a quantitative drift-budget theory of when BCJR-QAT can escape the QTIP-PTQ Voronoi basin, verified across four experiments; and (iii) a positive empirical result on Llama-3.2-1B at 2 bpw under end-to-end forward-KL distillation: with the right schedule (skip the high-$T$ phase to avoid an overshoot we diagnose), single-layer BCJR-QAT beats QTIP-PTQ by $\mathbf{-0.084}$ PPL on WikiText-2, and multi-layer compounding is super-additive.

Modern LLM APIs often reveal only top-$K$ logit scores and censor the remaining vocabulary. We study the per-position distribution-recovery limits of this access model. For censoring threshold $τ$, the compatible teacher distributions form an identified set whose total-variation diameter is exactly $U_K=(V-K)\exp(τ)/(Z_A+(V-K)\exp(τ))$, where $Z_A$ is the observed partition function. For KL recovery, we give a computable binary-endpoint lower bound and an asymptotically matching small-ambiguity upper bound, with an extension to reference-aware attackers. Experiments on a Qwen3 math-reasoning teacher reveal a layered extraction hierarchy: on-task top-$K$ distillation recovers 12% of private capability, full-logit distillation recovers 56% despite 99% KL closure, and generation-based extraction recovers 96%. Top-$K$ censoring therefore limits per-position distribution recovery but does not by itself prevent capability extraction, separating fidelity from transfer in prompt-only logit distillation.

On-policy self-distillation (self-OPD) densifies reinforcement learning with verifiable rewards (RLVR) by letting a policy teach itself under privileged context. We find that when this guidance spans the full response, all-token KL spends gradients on mostly redundant positions and amplifies privileged-information leakage, causing entropy rise, shortened reasoning, and out-of-distribution degradation in long-horizon math training. We propose Token-Routed Alignment for Critical rEasoning (TRACE), which distills only on annotator-marked critical spans: forward KL on key spans of correct rollouts, optional reverse KL on localized error spans, and GRPO on all remaining tokens, with the KL channel annealed away after a short warm-up. Our analysis explains TRACE through two effects: forward KL provides non-vanishing lift to teacher-supported tokens that the student under-allocates, while span masking and decay keep cumulative privileged-gradient exposure finite. On four held-out math benchmarks plus GPQA-Diamond, TRACE improves over GRPO by 2.76 percentage points on average and preserves the Qwen3-8B base OOD score on GPQA-Diamond, where GRPO and all-token self-OPD baselines degrade. Gains persist under online self-annotation (+1.90 percentage points, about 69% of the strong-API gain), reducing the concern that TRACE merely imports external annotator capability. Across scales, the best routed action is base-dependent: on Qwen3-8B it is forward KL on key spans, while on Qwen3-1.7B it shifts to reverse KL on error spans.

In this paper, we study the problem of sampling from a distribution under the constraint of differential privacy (DP). Prior works measure the utility of DP sampling with density ratio-based measures such as KL divergence. However, such formulations suffer from two key limitations: 1) they fail to capture the geometric structure of the support, and 2) they are not applicable when the supports of the distributions differ. To deal with these issues, we develop a novel framework for DP sampling with Wasserstein distance as the utility measure. In this formulation, we propose Wasserstein Projection Mechanism (WPM), a minimax optimal mechanism based on Wasserstein projection. Furthermore, we develop efficient algorithms for computing the proposed mechanisms approximately and provide convergence guarantees.

Reinforcement Learning with Verifiable Rewards (RLVR) has become the standard paradigm for LLM mathematical reasoning, where Group Relative Policy Optimization (GRPO) serves as the mainstream algorithm. We point out two understudied inefficiencies existing in GRPO. First, the fixed KL penalty coefficient overly restricts policy exploration at stages where the model requires significant deviation from the reference policy. Second, uniform sampling of training questions ignores that moderately difficult problems provide the most informative gradient signals for optimization. We propose Exploration-Prioritized Policy Optimization (EXPO) with two lightweight plug-in modules. The Accuracy-Conditioned KL Scaling (AKL) dynamically adjusts KL regularization strength through a smooth nonlinear function of batch average accuracy, relaxing the penalty when the model underperforms and strengthening it when the model achieves good results. The Gaussian Curriculum Sampling (GCS) assigns sampling weights to questions following a Gaussian distribution centered at moderate accuracy around 0.5, focusing training on the model's learning frontier. We conduct extensive experiments on DeepSeek-R1-Distill-Qwen-1.5B and Qwen3-8B-Base over six mathematical reasoning benchmarks. The results show EXPO steadily surpasses vanilla GRPO. It obtains an absolute gain of 13.34 on AIME 2025 pass@32, rising from 63.33 percent to 76.67 percent, and achieves an average pass@32 improvement of 2.66 on the 8B model. The much larger performance gains on pass@32 compared with pass@1 demonstrate that EXPO effectively enlarges the model's exploration boundary under a fixed inference cost budget.

Diffusion large language models (dLLMs) offer a promising paradigm for parallel text generation, but in practice they face an accuracy-parallelism trade-off, where increasing tokens per forward (TPF) often degrades generation quality. Existing acceleration methods often gain speed at the cost of accuracy. To address this limitation, we propose TAD, a Temporal-Aware trajectory self-Distillation framework. During data construction, we condition a teacher model on both the prompt and the ground-truth response to generate decoding trajectories, recording the intermediate masked states throughout the process. Based on how many decoding steps remain before each masked token is revealed, we partition masked positions into near and distant subsets. For near tokens, we train the student with a hard cross-entropy loss using the teacher trajectory tokens as labels, encouraging confident predictions for tokens that are about to be decoded. For distant tokens, we apply a soft KL divergence loss between the teacher and student token distributions, providing softer supervision and preserving future planning knowledge. This temporal-aware partition naturally gives rise to two deployment configurations: a Quality model that prioritizes accuracy and a Speed model that favors more aggressive acceleration. Experiments show that TAD consistently improves the accuracy-parallelism trade-off. On LLaDA, it raises average accuracy from 46.2% to 51.6% with the Quality model and average AUP from 46.2 to 257.1 with the Speed model. Our code is available at: https://github.com/BHmingyang/TAD

Flow Matching has recently emerged as a popular class of generative models for simulating a target distribution $μ_1$ from samples drawn from a source distribution $μ_0$. This framework relies on a fixed coupling between $μ_0$ and $μ_1$, and on a deterministic or stochastic bridge to define an interpolating process between the two distributions. The time marginals of this process can then be approximately sampled by estimating the transition rates, or more generally the generator, of its Markovian projection. This framework has recently been extended to the case of discrete source and target distributions, under the name Discrete Flow Matching (DFM). However, theoretical guarantees for such models remain scarce. In this paper, we study two DFM models on $\mathbb{Z}_m^d = \{0,\ldots,m-1\}^d$, sampled through time discretization, and derive non-asymptotic associated bounds for both of them. In contrast to previous work, we establish non-asymptotic bounds in Kullback--Leibler divergence for the early-stopped version of the target distribution. We also derive explicit convergence guarantees in total variation distance with respect to the true target distribution. Importantly, these bounds rely only on an approximation error assumption, relaxing standard score assumptions used in earlier works, while also yielding improved dependence on the vocabulary size $m$ and the dimension $d$.

On-Policy Distillation (OPD) has emerged as a dominant post-training paradigm for large language models, especially for reasoning domains. However, OPD remains unstable in practice due to the high gradient variance of its single-sample Monte Carlo estimator, and recipes for stable training are still immature. We propose vOPD (On-Policy Distillation with a control variate baseline), which casts OPD as policy-gradient RL and stabilizes it by introducing a control variate baseline-canonically a value function -- from the RL literature. We show that the OPD value function admits a closed form as the per-token negative reverse KL divergence between the student and the teacher, available directly from the already-computed forward pass with no additional critic or inference. Existing stabilization methods either compute the full token-level reverse KL over the entire vocabulary, adding significant overhead, or restrict it to a top-k support, biasing the objective. vOPD instead preserves the lightweight single-sample estimator, subtracting the value function as a detached baseline to keep the gradient unbiased while reducing variance. Furthermore, we show that a top-k approximation of the baseline further lowers cost without compromising performance. Across mathematical and scientific reasoning benchmarks, vOPD consistently outperforms vanilla OPD and matches the most expensive full-vocabulary baseline, offering an efficient stabilization of On-Policy Distillation through principled RL variance reduction.

Optimizers that exploit the matrix structure of gradients are central to modern LLM pre-training, with two distinct frontiers: explicit Kronecker-factored preconditioning -- most recently KL-Shampoo, which estimates the preconditioner via KL divergence minimization -- and orthogonalization of the gradient momentum, exemplified by Muon and analyzed as steepest descent under the spectral norm. The two routes are typically developed in isolation. We make a structural observation about KL-Shampoo's Kronecker preconditioners: their eigenvalue spectra exhibit a \emph{spike-and-flat} shape -- a few dominant eigenvalues followed by an approximately uniform tail -- across layers and training stages, holding exactly under a rank-$ρ$ signal-plus-noise gradient model. We exploit this structure by restricting one of KL-Shampoo's Kronecker factors to a parametric family aligned with the spike-and-flat shape: full spectral structure on a tracked $r$-dimensional subspace, single shared eigenvalue across the remaining $n-r$ directions. On these directions, we apply orthogonalization. An identity shows that this orthogonalization recovers the algebraic form of full KL-Shampoo's preconditioner. On four pre-training scales (GPT-2 124M / 350M, LLaMA 134M / 450M), Pro-KLShampoo consistently outperforms KL-Shampoo at every subspace rank we test in validation loss, peak per-GPU memory, and wallclock time to reach each loss level.

Large language models (LLMs) remain vulnerable to adversarial prompting despite advances in alignment and safety, often exhibiting harmful behaviors under novel attack strategies. While adversarial training can improve robustness, existing approaches are computationally expensive and difficult to scale. Recent continuous adversarial training methods, such as Continuous adversarial training (CAT) and Continuous Adversarial Preference Optimization (CAPO), address this challenge by leveraging gradient-based perturbations in the embedding space, enabling more efficient and expressive attacks. Building on this paradigm, we propose WARDEN, a distributionally robust adversarial training framework for LLMs that dynamically reweights adversarial examples through an f -divergence ambiguity set around the empirical training distribution. Our method optimizes the worst-case adversarial loss within a divergence ball around the empirical data distribution, automatically emphasizing harder adversarial examples. Using the convex dual formulation, the objective reduces to a log-sum-exp form under the KL divergence, with a dynamical parameter controlling the strength of reweighting. This study leads to a new class of information-theoretic objectives that significantly reduce attack success rates while maintaining model utility. Across multiple LLMs and attack settings, WARDEN substantially reduces attack success rates with computational and utility costs comparable to CAT-, CAPO-, and MixAT-based baselines, making it a practical approach for scalable robust alignment.

On-policy distillation is an efficient alternative to reinforcement learning, offering dense token-level training signals. However, its reliance on a stronger external teacher has driven recent work on on-policy self-distillation, where the same model serves as both teacher and student under different prompt contexts. Yet, existing self-distillation methods largely reduce learning to KL matching toward the context-augmented teacher model. This approach often suffers from training instability and can degrade reasoning performance over time. Moreover, self-distillation from the same model with prompt augmentation lacks the exploratory diversity provided by a genuine external teacher. To address these limitations, we move beyond fixed-teacher KL matching and propose \textbf{P}reference-\textbf{B}ased \textbf{S}elf-\textbf{D}istillation (\textbf{PBSD}), which revisits on-policy self-distillation through a reward-regularized perspective. Instead of directly matching the teacher distribution, we derive a reward-regularized objective whose analytic optimum is a reward-reweighted teacher distribution, yielding a target policy provably superior to the original teacher under this objective. Practically, PBSD optimizes preference gaps between teacher and student samples while maintaining on-policy student sampling. We support this framework with a statistical analysis of the induced preference-learning problem, formally establishing when on policy self-distillation is preferable to learning from an external teacher in our setting. Experiments on mathematical reasoning and tool-use benchmarks across multiple model scales demonstrate that PBSD consistently achieves the strongest average performance among comparable baselines, showing improved training stability over prior self-distillation baselines while preserving token efficiency.

Reinforcement learning is crucial for aligning large language models to perform complex reasoning tasks. However, current algorithms such as Group Relative Policy Optimization suffer from coarse grained, sequence level credit assignment, which severely struggles to isolate pivotal reasoning steps within long Chain of Thought generations. Furthermore, the standard unbounded Kullback Leibler divergence penalty induces severe gradient instability and mode seeking conservatism, ultimately stifling the discovery of novel reasoning trajectories. To overcome these limitations, we introduce Distribution Guided Policy Optimization, a novel critic free reinforcement learning framework that reinterprets distribution deviation as a guiding signal rather than a rigid penalty. DGPO replaces the volatile KL divergence with the bounded Hellinger distance to safely quantify token level exploration without the risk of gradient explosion. To effectively distinguish genuine reasoning breakthroughs from hallucinatory noise, we propose an entropy gating mechanism that scales this deviation by the policy`s epistemic uncertainty. By dynamically redistributing the coarse sequence-level advantage to individual tokens based on these gated scores, DGPO heavily incentivizes critical exploratory steps while suppressing unwarranted, low-entropy deviations. Consequently, DGPO completely eliminates the traditional token-level KL penalty and achieves fine-grained credit reallocation without the computational overhead of an additional value network. Extensive empirical evaluations demonstrate that DGPO sets a new state-of-the-art for critic free alignment. Notably, on the Qwen2.5-32B architecture, DGPO achieves 60.0% Avg@32 accuracy and 46.0% Avg@32 accuracy on the challenging AIME2024 and AIME2025 benchmarks respectively, substantially outperforming competitive baselines like DAPO.

The Distributional Alignment Game framework provides a powerful variational perspective on Answer-Level Fine-Tuning (ALFT). However, standard algorithms for these games rely on estimating logarithmic rewards from small batches, introducing a systematic bias due to Jensen's inequality that can destabilize training. In this paper, we systematically resolve this structural estimation bias. First, we generalize the alignment game to arbitrary Bregman divergences, showing that for a family of geometries inducing polynomial rewards, we can construct provably exact and unbiased estimators using U-statistics. Second, for the canonical KL divergence game where an exact solution is impossible, we derive a globally robust minimax polynomial estimator that is provably optimal, achieving the fundamental statistical error limit of $Θ(1/K^2)$, which we establish via the Ditzian-Totik theorem. Finally, we synthesize these two approaches to propose a novel Variance-Optimal Augmented Polynomial Optimization Program (AQP) Estimator, proving that by systematically reducing variance, our method achieves not only optimal bias but also provably accelerated game convergence, leading to more efficient and stable training with zero online computational overhead.

Reinforcement learning with verifiable rewards (RLVR) has become a standard approach for improving reasoning in language models, yet models trained with RLVR often suffer from diversity collapse: while single-sample accuracy improves, multi-sample coverage degrades, sometimes falling below the base model. We provide a structural account of this phenomenon grounded in the properties of binary rewards. Binary rewards create a fundamental degeneracy for policy gradient methods: the set of distributions maximizing expected reward is infinite, with no distinguished element. KL-control resolves this degeneracy by selecting, in the limit $β\to 0$, the filtered model $p_*:=a(\cdot\mid\mathcal{Y}_1)$ -- the base model conditioned on validity -- which is the unique fully valid distribution closest to the base model in KL divergence. This selection operates through a nontrivial asymmetry: the tilted distribution $p_{[β]}\propto a(y)\,e^{v(y)/β}$ converges to $p_*$ in forward KL as $β\to 0$, yet $p_*$ cannot serve as a direct optimization target because $\mathrm{KL}(q\,\|\,p_*)$ is infinite for any full-support policy $q$. We develop explicit formulas relating the hyperparameter $β$ to the more interpretable target validity rate $μ$. Under model misspecification -- the typical practical regime -- the pressure to decrease $β$ drives the optimizer toward highly concentrated distributions over a small number of valid outputs, collapsing toward ever fewer as $β$ decreases, rather than toward the filtered model. We illustrate this mechanism on a toy autoregressive experiment and discuss how alternative divergences that target $p_*$ directly -- as pursued empirically by \citet{kruszewski_whatever_2026} -- avoid this failure mode by rewarding coverage of $p_*$'s support rather than concentration on high-validity outputs.

Exponential families encompass the distributions central to modern machine learning -- softmax, Gaussians, and Boltzmann distributions -- and underlie the theory of variational inference, entropy-regularized reinforcement learning, and RLHF. We isolate a simple identity for exponential families that expresses the KL difference $\mathrm{KL}(q \| p_{λ_2}) - \mathrm{KL}(q \| p_{λ_1})$ in terms of the log-partition function $A(λ)$ and the moment $μ_q$. Remarkably, this identity together with the single fact that $\mathrm{KL} \geq 0$ (with equality iff $p = q$) suffices, by direct substitution and rearrangement, to derive a cluster of results that are classically obtained by separate, heavier arguments: a generalized three-point identity for arbitrary reference distributions, Pythagorean theorems for I-projections and reverse I-projections, convexity of the log-partition function, identification of its Legendre dual in KL terms, the Gibbs variational principle, and the explicit optimizer in KL-regularized reward maximization, including the exponential tilting formula underlying entropy-regularized control and RLHF. Beyond these purely algebraic consequences, standard analytic arguments recover the gradient formula for the log-partition function, the Bregman representation of within-family KL divergence, and the surjectivity of the moment map. The note is self-contained.

We analyse three KV cache quantization schemes under a fair bit budget: \textbf{KV} (scalar MSE baseline), \textbf{KQV} (WHT + MSE on $K$; WHT + MSE + QJL on $V$), and \textbf{QKQV} (WHT + MSE + QJL on both). Starting from the Beta distribution on the hypersphere, we trace how QJL on $K$ inflates inner product variance by $π/2$, which softmax amplifies nonlinearly via Jensen's inequality, and we present statistical inference and information metrics to highlight practical differences. Three empirical findings emerge. (1)~At $n=4$ (the practically dominant budget), KQV wins on every measure -- KL divergence, geometric $K$ error, and 6D distance -- across all distributions and ranks tested. (2)~The K--V asymmetry is unconditional: QKQV is consistently worse than KQV in KL divergence at every budget and distribution. (3)~A budget-dependent crossover exists: QKQV achieves better geometric $K$ reconstruction at $n \in \{2,3,5\}$, KQV at $n \in \{4,6\}$, invariant to rank and tail weight -- an open rate-distortion problem. $\mathrm{KL}(p_{\mathrm{ref}} \| p_{\mathrm{quant}})$, K-only by construction, bridges K direction error to routing corruption and output collapse. We present a sufficient condition when the Jensen mechanism amplifies superlinearly through the softmax. At $n \in \{2,3,5\}$, QKQV wins geometrically because this assumption does not bind. At $n=4$, elevated K error and KL divergence for QKQV strongly suggest the Jensen mechanism is the operative cause of the crossover, providing a new perspective and explanation.

Speech-only spoken language models (SLMs) lag behind text and text-speech models in performance, with recent discrete autoregressive (AR) SLMs indicating significant computational and data demands to match text models. Since discretizing continuous speech for AR creates bottlenecks, we explore whether continuous diffusion (CD) SLM is more viable. To quantify the SLMs linguistic quality, we introduce the phoneme Jensen-Shannon divergence (pJSD) metric. Our analysis reveals CD SLMs, mirroring AR behavior, exhibit scaling laws for validation loss and pJSD, and show optimal token-to-parameter ratios decreasing as compute scales. However, for the latter, loss becomes insensitive to choice of data and model sizes, showing potential for fast inference. Scaling CD SLMs to 16B parameters with tens of millions of hours of conversational data enables generation of emotive, prosodic, multi-speaker, multilingual speech, though achieving long-form coherence remains a significant challenge.

On-policy distillation (OPD) has shown strong potential for transferring reasoning ability from frontier or domain-specific models to smaller students. While effective on static single-turn tasks, its behavior in multi-turn agent settings remains underexplored. In this work, we identify a key limitation of vanilla OPD in such settings, which we term Trajectory-Level KL Instability. Specifically, we observe that KL divergence increases together with a drop in success rate, and even after convergence, the KL remains high, leading to unstable training. This instability arises from inter-turn error compounding: as errors accumulate, the student is driven beyond the teacher's effective support, rendering the supervision signal unreliable. To address this, we propose TCOD (Temporal Curriculum On-Policy Distillation), a simple yet effective framework that controls the trajectory depth exposed to the student and progressively expands it from short to long with a curriculum schedule. Experimental results across four student-teacher pairs on three multi-turn agent benchmarks (ALFWorld, WebShop, ScienceWorld) show that TCOD mitigates KL escalation and enhances KL stability throughout training, improving agent performance by up to 18 points over vanilla OPD. Further evaluations show that TCOD can even surpass the teacher's performance and generalize to tasks on which the teacher fails. Our code is available at https://github.com/kokolerk/TCOD.

When approximating an intractable density via variational inference (VI) the variational family is typically chosen as a simple parametric family that very likely does not contain the target. This raises the question: Under which conditions can we recover characteristics of the target despite misspecification? In this work, we extend previous theoretical results on robust VI with location-scale families under target symmetries in two substantial ways: (1) We open them up to a wider range of divergences by providing sufficient conditions for exact recovery of the target mean and correlation matrix when using the forward Kullback-Leibler divergence and $α$-divergences. (2) By doing so, we find that we can drop the restrictive assumption of a log-concave target made in previous work, allowing us to give guarantees for a wider range of targets, including multi-modal ones. In our experiments, we show how our guarantees can serve as guidelines for the choice of the variational family and $α$-value and we illustrate on a diverse set of examples how and why optimization can fail in the absence of our sufficient conditions.

Trajectory Inference (TI) seeks to recover latent dynamical processes from snapshot data, where only independent samples from time-indexed marginals are observed. In applications such as single-cell genomics, destructive measurements make path-space laws non-identifiable from finitely many marginals, leaving held-out marginal prediction as the dominant but limited evaluation protocol. We introduce a general framework for estimating the Kullback-Leibler divergence (KL) divergence between probability measures on function space, yielding a tractable, data-driven estimator that is scalable to realistic snapshot datasets. We validate the accuracy of our estimator on a benchmark suite, where the estimated functional KL closely matches the analytic KL. Applying this framework to synthetic and real scRNA-seq datasets, we show that current evaluation metrics often give inconsistent assessments, whereas path-space KL enables a coherent comparison of trajectory inference methods and exposes discrepancies in inferred dynamics, especially in regions with sparse or missing data. These results support functional KL as a principled criterion for evaluating trajectory inference under partial observability.

We present a systematic empirical study of transformer compression through over 40 experiments on GPT-2 (124M parameters) and Mistral 7B (7.24B parameters). Our analysis covers spectral compression, block-level function replacement, rotation-based quantization, activation geometry, and adaptive early exit. We identify five structural properties relevant to compression. (1) Variance is not importance: high-variance activation directions are approximately 96 percent uncorrelated with predictive directions (measured via CCA), and projecting onto these subspaces preserves over 90 percent of variance while degrading perplexity. (2) Block linearity is conditional: transformer blocks are approximately linear (R^2 ~ 0.95 on GPT-2, 0.93 on Mistral block 31) only under the correct upstream distribution; modifying earlier blocks induces distribution shift that degrades downstream approximations. (3) The reconstruction wall: approaches that factor weights into quantized components amplify errors through cross-terms, making direct quantization strictly superior. (4) Linearity increases with depth: Mistral 7B exhibits a progression from R^2 = 0.17 (block 0) to R^2 = 0.93 (block 31), indicating a division between nonlinear feature construction and linear refinement. (5) Approximately 30 percent of tokens are computationally easy, confirmed via exit heads and KL divergence sensitivity. We demonstrate that single-block linear replacement achieves 34x compression with a 1.71 perplexity increase on the final block of Mistral 7B, while multi-block replacement fails due to residual error accumulation and distribution shift. These findings suggest fundamental limits to static post-training compression and motivate adaptive, per-token computation as a more effective direction.

We consider the problem of sampling from a probability distribution $π$. It is well known that this can be written as an optimisation problem over the space of probability distributions in which we aim to minimise the Kullback--Leibler divergence from $π$. We consider the effect of replacing $π$ with a sequence of moving targets $(π_t)_{t\ge0}$ defined via geometric tempering on the Wasserstein and Fisher--Rao gradient flows. We show that convergence occurs exponentially in continuous time, providing novel bounds in both cases. We also consider popular time discretisations and explore their convergence properties. We show that in the Fisher--Rao case, replacing the target distribution with a geometric mixture of initial and target distribution never leads to a convergence speed up both in continuous time and in discrete time. Finally, we explore the gradient flow structure of tempered dynamics and derive novel adaptive tempering schedules.

For a broad family of discriminative models that includes autoregressive language models, identifiability results imply that if two models induce the same conditional distributions, then their internal representations are equal up to an invertible linear transformation. We ask whether an analogous conclusion holds approximately when the distributions are close instead of equal. Building on the observation of Nielsen et al. (2025) that closeness in KL divergence need not imply high linear representational similarity, we study a distributional distance based on logit differences and show that closeness in this distance does yield linear similarity guarantees. Specifically, we define a representational dissimilarity measure based on the models' identifiability class and prove that it is bounded by the logit distance. We further show that, when model probabilities are bounded away from zero, KL divergence upper-bounds logit distance; yet the resulting bound fails to provide nontrivial control in practice. As a consequence, KL-based distillation can match a teacher's predictions while failing to preserve linear representational properties, such as linear-probe recoverability of human-interpretable concepts. In distillation experiments on synthetic and image datasets, logit-distance distillation yields students with higher linear representational similarity and better preservation of the teacher's linearly recoverable concepts.

Diffusion models over discrete spaces have recently shown striking empirical success, yet their theoretical foundations remain incomplete. In this paper, we study the sampling efficiency of score-based discrete diffusion models under a continuous-time Markov chain (CTMC) formulation, with a focus on $τ$-leaping-based samplers. We establish sharp convergence guarantees for attaining $\varepsilon$ accuracy in Kullback-Leibler (KL) divergence for both uniform and masking noising processes. For uniform discrete diffusion, we show that the $τ$-leaping algorithm achieves an iteration complexity of order $\tilde O(d/\varepsilon)$, with $d$ the ambient dimension of the target distribution, eliminating linear dependence on the vocabulary size $S$ and improving existing bounds by a factor of $d$; moreover, we establish a matching algorithmic lower bound showing that linear dependence on the ambient dimension is unavoidable in general. For masking discrete diffusion, we introduce a modified $τ$-leaping sampler whose convergence rate is governed by an intrinsic information-theoretic quantity, termed the effective total correlation, which is bounded by $d \log S$ but can be sublinear or even constant for structured data. As a consequence, the sampler provably adapts to low-dimensional structure without prior knowledge or algorithmic modification, yielding sublinear convergence rates for various practical examples (such as hidden Markov models, image data, and random graphs). Our analysis requires no boundedness or smoothness assumptions on the score estimator beyond control of the score entropy loss.

This paper introduces an analytical approach to quantifying and optimizing the distributional discrepancy in generative diffusion models. For a multivariate Gaussian source, we explicitly derive the closed-form evolution trajectory and the resulting Kullback-Leibler (KL) divergence between the distributions of the source data and the reversely sampled data. Asymptotic analysis via the Euler-Maclaurin expansion characterizes the convergence behavior of this KL divergence, extracting its dominant term as an explicit functional of the noise schedule. Minimizing this dominant term via the calculus of variations yields a noise schedule described by a tangent law, inherently determined by the source covariance spectrum. We further prove that the Gaussian source exhibits an extremal property for the KL divergence among general source distributions with a given covariance. We also utilize the analytical KL divergence as a principled metric to identify efficient time discretization strategies for pretrained diffusion models, and demonstrate via experiments over diverse datasets that the identified strategies consistently outperform established baselines, particularly under constrained function evaluation budgets.

We inspect the deductive connection between the neural scaling law and Zipf's law -- two statements discussed in machine learning and quantitative linguistics. The neural scaling law describes how the cross entropy rate of a foundation model -- such as a large language model -- changes with respect to the amount of training tokens, parameters, and compute. By contrast, Zipf's law posits that the distribution of tokens exhibits a power law tail. Whereas similar claims have been made in more specific settings, we show that the neural scaling law is a consequence of Zipf's law under certain broad assumptions that we reveal systematically. The derivation steps are as follows: We derive Heaps' law on the vocabulary growth from Zipf's law, Hilberg's hypothesis on the entropy scaling from Heaps' law, and the neural scaling from Hilberg's hypothesis. We illustrate these inference steps by a toy example of the Santa Fe process that satisfies all four statistical laws.

Shampoo and its efficient variant, SOAP, employ structured second-moment estimations and have shown strong performance for training neural networks (NNs). In practice, however, Shampoo typically requires step-size grafting with Adam to be competitive, and SOAP mitigates this by applying Adam in Shampoo's eigenbasis -- at the cost of additional memory overhead from Adam in both methods. Prior analyses have largely relied on the Frobenius norm to motivate these estimation schemes. We instead recast their estimation procedures as covariance estimation under Kullback-Leibler (KL) divergence minimization, revealing a previously overlooked theoretical limitation and motivating principled redesigns. Building on this perspective, we develop $\textbf{KL-Shampoo}$ and $\textbf{KL-SOAP}$, practical schemes that match or exceed the performance of Shampoo and SOAP in NN pre-training while achieving SOAP-level per-iteration runtime. Notably, KL-Shampoo does not rely on Adam to attain competitive performance, eliminating the memory overhead introduced by Adam. Across our experiments, KL-Shampoo consistently outperforms SOAP, Shampoo, and even KL-SOAP, establishing the KL-based approach as a promising foundation for designing structured methods in NN optimization. An implementation of KL-Shampoo/KL-SOAP is available at https://github.com/yorkerlin/KL-Methods

Large Language Models have achieved remarkable performance across a large number of tasks, but face critical deployment and usage barriers due to substantial computational requirements. Model compression methods, which aim to reduce model size while preserving its capacity, are an important means to mitigate these issues. Promising approaches along these lines, such as structured pruning, typically require costly manual hyperparameter exploration or rely on local heuristics that may run the risk of ignoring better solutions. In this work we introduce GeLaCo, an evolutionary approach to LLM compression via layer collapse. Our approach supports an efficient exploration of the compression solution space via population-based search and a novel layer collapse formulation based on parametrized weight merging, with a fitness function based on similarity over residual updates and language modeling KL divergence. GeLaCo also supports both single and multi-objective evolutionary compression search, establishing the first Pareto front estimation along compression and quality axes. We evaluate GeLaCo solutions via both perplexity-based and generative evaluations over foundational and instruction-tuned models, outperforming state-of-the-art alternatives.

Heavy-tailed posteriors are routine in Euclidean variational autoencoders, where the Student family relaxes the Gaussian without new machinery. The sphere has had no comparable option. Von Mises-Fisher distribution needs modified Bessel functions and a rejection sampler, and Power Spherical buys its closed forms by forcing the density to vanish at the antipode. We develop the spherical Cauchy distribution as a hyperspherical posterior that needs neither compromise. Stereographic projection carries it to a multivariate Student law, and a Möbius transformation turns a uniform spherical draw into an exact posterior sample from inner products, norms, and scalar arithmetic. The same transformation settles the regularizer. Evaluating the density along the sampling map reduces the Kullback-Leibler (KL) divergence to the uniform prior to a scalar expectation whose expansion terminates in every even ambient dimension, leaving one logarithm and a polynomial with finitely many correction terms. Odd dimensions admit certified truncation of value and gradient, the KL is increasing and convex in concentration, and the same function gives the pairwise KL. At matched modal curvature it has broader angular tails and a smaller KL penalty than both alternatives, so equal local precision costs less regularization. In dimension 128 the fused evaluator runs 1.5 times faster per latent-layer step than Power Spherical and 4.2 times faster than robust von Mises-Fisher on CPU, with factors of 1.6 and 5.4 on CUDA. Across five paired seeds it attains the lowest MNIST reconstruction loss at every tested dimension and lowers held-out viewpoint-gap negative log-likelihood on smallNORB by 3.6 percent.

Sampling in score-based diffusion models can be performed by solving either a reverse-time stochastic differential equation (SDE) parameterized by an arbitrary stochasticity function or a probability flow ODE, corresponding to setting this stochasticity function to zero. In this work, we investigate the effect of this stochasticity on the generation process through the evolution of Kullback-Leibler (KL) divergences, obtaining general KL divergence bounds and a novel analysis of the impact of the time-profile of the score error on model performance. For exact score functions, stochasticity has a contractive effect, decreasing KL divergence along the sampling trajectory. For approximate scores, however, a trade-off arises between correcting accumulated errors and amplifying current score errors, meaning stochasticity can either improve or degrade generation performance. Theoretical considerations indicate that the gain from stochasticity depends on the time-localization of the trained model error. We test this in experiments on both toy and benchmark data sets, also comparing the KL divergence evolution with the obtained bounds. We also present a fully analytical example, where all the relevant quantities can be computed, and the optimal stochasticity function can be characterized via an optimal control analysis.

We propose a framework for generative modeling of continuous-time processes from irregularly and asynchronously recorded data. It is based on the matching of generators and accommodates discontinuous trajectories. Analytical formulas for diffusion and jump bridges yield a family of reference generators that a neural network is trained to match. The key ingredient is that, for our constructed jump bridge, a parametrization of the jump kernel densities by scaled Gaussians admits closed-form expressions for the Kullback-Leibler divergence, allowing simulation-free training.

Otto's Wasserstein gradient flow of the inclusive (forward) Kullback--Leibler (KL) divergence offers a principled framework for analyzing statistical inference algorithms, yet algorithms targeting the exclusive (reverse) KL divergence are rarely studied with such tools. We establish a unified gradient-flow and PDF framework for inclusive KL inference. We show that maximum mean discrepancy minimization can be viewed as inclusive KL inference with an approximate gradient estimator, and we develop the Fisher--Rao and Wasserstein--Fisher--Rao gradient flows that directly target the inclusive KL divergence. Restricting these flows to the manifold of Gaussian distributions yields explicit gradient-flow ODEs, providing a foundation for Gaussian variational inference. Building on this viewpoint, we further introduce a local-estimator Wasserstein gradient flow whose velocity is obtained by local nonparametric regression, free of density-ratio evaluation or kernel gradients, improving the algorithmic performance over the MMD-based particle method.

Post-training quantization (PTQ) is essential for deploying large language models (LLMs) under strict resource constraints. State-of-the-art PTQ methods quantize each layer with a single closed-form second-order solver: to remain analytically tractable, they heavily approximate the global loss (dropping cross-channel coupling, pooling output rows into groups), and they then freeze the resulting Hessian across the entire layer, with no way to refresh it as the loss landscape shifts column by column--a phenomenon we call information misalignment. We propose REAL-Q (Real-time E2E-loss Aligned LLM Quantization), a novel PTQ paradigm that breaks this compromise: instead of diluting the objective for the sake of analytic tractability, REAL-Q targets an end-to-end-aligned surrogate of the global loss and refines it via fine-grained, dynamic Block-wise Gradient Descent applied after every column block (128 columns). By coupling this fine-grained correction with a sliding window mechanism for smooth cross-layer transitions, REAL-Q effectively mitigates error propagation across the network. On LLaMA-3.1 (8B and 70B) and Qwen3 (0.6B-32B) at W4A16, REAL-Q reduces end-to-end KL divergence by up to ~49% relative to state-of-the-art globally-guided methods.