每天知道
你的领域
出了什么

Reinforcement learning from human feedback (RLHF) aligns large language models (LLMs) with human preferences, yet most pipelines learn a single reward model that overlooks individual differences in preferences. Personalized reward models (PRMs) address this by conditioning rewards on user-specific feedback, most commonly through in-context learning (ICL), where a user's historical comparisons are supplied as contextual preference pairs. However, we identify a key limitation of ICL-based PRMs: they fail to capture the preference relations conveyed by contextual pairs. To address this, we propose Preference-Aligned Test-Time Training (P-TTT), which explicitly encodes these relations into user-specific fast weights for personalized reward prediction. P-TTT introduces sequence-level update and apply operations to match the response-level granularity of preference feedback, together with a preference-aligned objective that directly uses pairwise preference relations to guide fast-weight adaptation. Notably, P-TTT is simple to implement and computationally efficient, updating fast weights within a single forward pass without inference-time backpropagation. Extensive experiments show that P-TTT more effectively captures historical preference relations and outperforms state-of-the-art methods by a large margin.

Object hallucination remains a major challenge for large vision-language models. While off-policy preference optimization proves to be an effective solution, on-policy reinforcement learning provides a more promising direction as it directly targets a model's current failure modes. However, we find that without fine-grained reward formulation and allocation, on-policy optimization often falls into an easy shortcut: reducing hallucinations merely by saying less---making fewer valid claims. To comprehensively resolve this, we propose a fine-grained alignment framework that couples dense reward signals at the data level with precise credit assignment at the algorithmic level. Specifically, we first construct the Dense Object Presence and Absence (DOPA) dataset to address sparse annotations that prevent valid object claims from being verified and rewarded. DOPA exhaustively annotates the deterministic presence and absence of every concept across an expanded vocabulary, significantly increasing the density of reliable reward signals during on-policy rollouts. Second, we propose Subsentence-level Credit Assignment for on-Policy Optimization (SCAPO) to prevent response-level shared advantages from allowing local hallucinations to compromise all other valid outputs within the same response. By assigning credit to each subsentence independently based on its object claims, SCAPO can precisely reinforce faithful generations and penalize hallucinations. Furthermore, we leverage the resulting faithful image descriptions as auxiliary context to transfer generative gains to discriminative tasks. Experiments demonstrate that our method produces highly informative, faithful descriptions in generative tasks while yielding clear performance gains on discriminative evaluation.

We present MEGATRON, a heterogeneous Edge GenAI System-on-Chip in 28nm FD-SOI CMOS technology combining a non-volatile analog in-memory-computing engine based on a 4Mi-cell phase-change memory (PCM) array with a digital RISC-V-based flexible neural processing unit. MEGATRON demonstrates up to 3.5 TOPS/W using the RISC-V processors and 57.5 TOPS/W with PCiM, at a storage density of 1.52 Mparam/mm${}^2$ with 4-bit effective weight precision.

Long-context training runs on sequences whose lengths span orders of magnitude, and dynamic context parallelism (DCP) gives each sequence its own CP degree. Existing DCP systems either do not scale or perform poorly on mainstream models, leaving Megatron-Core (Mcore) DCP as the only option at production scale. Mcore DCP, however, sizes each degree to fit memory, which grows linearly with length while attention grows quadratically, so comparable token counts hide unequal computation. In our production 256K-context training job on more than 11K GPUs under Mcore DCP, per-rank microbatch times differ by up to 5x at similar token counts. The skew leads to a 46% pipeline bubble and a 13% data-parallel bubble. We present HyDra, a scalable load-driven DCP system. Its scheduler balances computation by pulling every rank toward one load target, and balance in turn makes that target solvable in closed form. It places sequences with lazy heaps rather than whole-pool scans. That balance asks for CP degrees larger than memory requires, so its CP engine nests an inner Ulysses group in a shallow outer ring, letting a higher degree lower computation and communication together. Evaluation at both scales shows consistent gains. On a 512-GPU testbed, HyDra raises throughput over Mcore DCP by 1.18x on average at 32K context and 2.48x at 256K. On a 2,048-GPU production job, it shrinks the pipeline bubble from 36% to 14%, cuts scheduling time by 2.3x, and raises throughput by 1.10-1.43x (avg. 1.25x) over Mcore DCP and 1.33-1.90x (avg. 1.59x) over static CP.

Dynamic routing creates severe load imbalance in large-scale expert-parallel Mixture-of-Experts (MoE) training, turning GPUs that host hot experts into stragglers. As each MoE layer waits for its slowest rank, these stragglers prolong the expert-parallel stage and reduce overall training efficiency. Existing expert-parallelism load-balancing (EPLB) systems commonly compute load-balancing plans on the CPU, incurring device--host data transfers and cross-rank synchronization that make scheduling at every layer and microbatch expensive. Their planning formulations also overlook the hierarchical communication costs of modern scale-up and scale-out GPU clusters. We present TopoEP, a GPU-native, topology-aware load-balancing system for large-scale MoE training. At each MoE layer and training microbatch, TopoEP converts the current routing result into hot-expert replication and token-rerouting decisions and executes the resulting plan without data-dependent host synchronization, reducing critical-path overhead. To generate these decisions, TopoEP uses a deterministic GPU solver that performs inter-node placement followed by intra-node refinement, allowing all ranks to independently produce bitwise-identical plans. On a 32-GPU NVIDIA H800 cluster, integrating TopoEP with Megatron-LM improves end-to-end training throughput by 6.2%--11.4% across three representative MoE models.

Pretraining a language model takes enormous compute, and the right optimizer can save a good part of it. Muon's spectral step gives a stronger direction than a sign step, but it is expensive. Every step runs Newton-Schulz iterations on the full matrix and, in distributed training, an extra all-reduce. Sign steps, as in Lion and Signum, are cheap and stay local to each device. We propose LionMuon, which takes one Muon step every $P$ iterations and Lion steps in between, with a single dual-EMA momentum buffer shared by both. Muon's compute and communication are paid once per $P$ steps, and the optimizer state is half of AdamW's. A single-EMA variant, SignMuon, already improves on Muon. We prove complexity bounds under heavy-tailed noise in which the period sets an interpolation between Muon's and Lion's smoothness and noise constants, and which say when LionMuon is faster than both. On 124M and 355M models trained on FineWeb, LionMuon with $P=2$ and $P=5$ reaches a lower loss than Muon, AdamW, Lion and Signum at the same number of tokens. Under 4-GPU data-parallel training it reaches Muon's final loss with a third less wall-clock on PCIe, and it beats the communication-efficient Muon variants Dion and MuonBP on loss at no more exposed communication, while keeping the exact gradient. Code: https://github.com/brain-lab-research/lion-muon

Looped architectures scale computation by reusing the same parameters across recurrent steps, and recent work shows that they substantially improve deep reinforcement learning policies on long-horizon tasks. Since recurrent depth directly controls computation, one may expect looped policies to naturally support elastic inference across recurrent depths. Surprisingly, we find that pretrained looped policies exhibit severe recurrent-depth specialization: reliable decisions are concentrated near the full trained depth, tying deployment computation to this depth even when less computation may suffice. Achieving depth elasticity, i.e., reliable decisions across recurrent depths with adaptive computation at deployment, therefore remains a key challenge. To address this, we propose FlexLoop, a novel post-training framework that converts pretrained fixed-depth looped policies into depth-elastic policies. FlexLoop keeps training on the original RL objective to preserve full-depth capability while performing adjacent-depth policy distillation to progressively transfer decision quality from deeper to shallower recurrent steps. The resulting policy supports reliable inference across recurrent depths and enables state-wise adaptive inference through recurrent-depth consistency. Experiments on $30$ online and offline long-horizon goal-conditioned environments show that FlexLoop preserves full-depth performance while making shallower depths effective. Keeping competitive performance, FlexLoop reduces average recurrent depth by up to $\bf{43%}$ and achieves up to $\bf{1.34\times}$ wall-clock speedup in a stress test.

On-Policy Self-Distillation (OPSD) has emerged as a promising paradigm for safety alignment, delivering dense, token-level supervision by distilling from a teacher conditioned on refusal-oriented privileged prompts. However, we reveal that this paradigm suffers from critical inefficiencies that degrade both training efficiency and general reasoning capabilities. Specifically, we diagnose two fundamental bottlenecks: (1) supervisory collapse over extended rollouts, where the teacher's corrective efficacy degrades precipitously as the student's generation prefix lengthens, injecting noisy gradients into late-stage tokens; and (2) gradient dilution from stylistic shifts, where the distillation objective is dominated by safety-irrelevant stylistic discrepancies induced by privileged prompting, washing out genuine safety signals and impairing base reasoning. To resolve these issues, we propose Efficient On-Policy Self-Distilled Safety Alignment (EOPSA), which concentrates computational and gradient budgets exclusively on reliably supervised, safety-critical tokens. EOPSA incorporates two coordinated mechanisms: (i) Adaptive Rollout Scheduling, which dynamically bounds the generation horizon guided by a novel Teacher Rescue Rate (TRR) metric to operate strictly within reliable supervision regimes; and (ii) Selective Distillation, which filters out safety-neutral tokens to restrict gradient updates exclusively to safety-pivotal transitions. Extensive evaluations across reasoning models up to 32B parameters demonstrate that EOPSA slashes rollout computation by $\sim$50% and backpropagates through merely $\sim$2% of tokens, consistently outperforming full-token distillation baselines in both safety compliance and reasoning retention.

World Action Models (WAMs) enable future-aware control by jointly modeling actions and environment dynamics. However, iterative diffusion or flow inference incurs substantial denoising latency. Prior inference state offers a natural opportunity for acceleration, yet changing planning contexts, observations, and intermediate representations can quickly render retained state stale. Preserving useful computation therefore requires adapting inference state rather than reusing it as-is. To this end, we present $\mathrm{WAM}{\scriptstyle\mathrm{ACHINE}}$, a training-free framework that accelerates WAM inference by preserving and adapting inference state for efficient and accurate continuation as the control loop evolves. Across closed-loop replans, Trajectory Remapping remaps replan state from the preceding replan to initialize the next replan, reducing redundant trajectory generation. Across denoising steps, Observation Rebinding performs anticipatory inference during action execution and rebinds retained denoising state to the real observation for continuation when consistency checks pass, reducing latency exposed to the control loop. Across Transformer layers, Residual Rescaling selectively rescales retained layer state and refreshes it through full computation of the middle layers when probe checks fail, reducing repeated Transformer computation. Evaluations of three representative WAM architectures on LIBERO and RoboTwin 2.0 show that $\mathrm{WAM}{\scriptstyle\mathrm{ACHINE}}$ achieves 1.47-3.05$\times$ speedups in observation-to-action latency and 2.23-3.27$\times$ speedups in GPU inference time per replan, while preserving 96.69-99.54% of native WAM task success.

Pretrained autoregressive Transformers represent a large sunk investment in compute. Existing conversion methods reuse that investment by changing either the architecture (attention to recurrence) or the objective (next-token prediction to denoising), never both. We convert Qwen3 teachers at 1.7B and 8B into attention-free, bidirectional, gated-delta-rule diffusion students in three stages, so that each capability can be traced to the stage that kept or lost it. Language modeling transfers only partially and in-distribution; in-context retrieval does not transfer. On a multi-query recall probe where the teachers score 0.34-0.58, both converted students score 0.000, and diffusion pretraining alone does not restore retrieval. A retrieval curriculum in the final stage, which gradually lengthens the gap between a key-value table and the queries that address it, restores it only stochastically: on a fixed schedule, one seed in three learns to retrieve. Advancing the gap only while a running accuracy estimate stays above a threshold works for all three of those seeds, holds on real text, and carries unchanged to 8B, where two of three seeds succeed. The third had not learned within its fixed 16k-step budget: retrieval switches on abruptly at a seed-dependent step (6.5k and 11k in the other two), so a fixed budget can cut a late run off. One boundary survives every intervention: every model that learns retrieval scores 0.000 on tokens that never appeared in a retrieval episode, and an arm that resamples the key and value tokens every batch shows this is a coverage limit, not memorization of particular bindings. Separately, we convert a 7B code model into a 3:1 recurrent-attention block-diffusion hybrid over 85k steps and report two negative training results.

Multi-teacher on-policy distillation (MOPD) has emerged as a popular post-training paradigm for integrating specialized capabilities in frontier language models. Existing OPD research has primarily focused on optimizing single-task distillation through objective design, distillation scope, and teacher signal construction, whereas MOPD must aggregate multiple capabilities in shared parameters and address the resulting capability seesaw, in which improving one domain suppresses capabilities acquired from another. Inspired by the distinctive update geometry of OPD, we find that parameter updates from different tasks rapidly concentrate in their respective low-dimensional subspaces during MOPD, providing a direct geometric basis for identifying and controlling cross-task interference. We therefore propose PMOPD (Projection-based Multi-Teacher On-Policy Distillation), which constructs subspace memories from the cumulative parameter displacements of different tasks and projects both gradients and optimizer updates to remove components that interfere with protected task directions. We further develop a lightweight conflict probe to characterize task interactions and guide task ordering, together with a cycling strategy that balances subspace estimation and timely task revisitation. Experiments on representative Code, Reason, and Math tasks show that PMOPD improves every evaluated capability over MOPD, raising the average score across the three tasks by 2.54 points on Qwen2.5-7B and 2.09 points on Llama-3.1-8B. These consistent gains establish geometry-aware optimization as an effective and transferable approach to balanced multi-teacher distillation.

Representation alignment speeds up diffusion transformer training by pulling an intermediate block of the model (student) toward features of a frozen pretrained encoder (teacher). Which teacher layer to align, and for how long, is still set by convention, and each alternative costs a training run. We find that alignment helps where the student cannot linearly recover the teacher's features, not where it already resembles them. Since a deep teacher layer is largely predictable from the one below, we isolate what each layer adds, its increment, and measure how much of it an unaligned student recovers. The student fills the teacher's hierarchy from the bottom up and stalls near the top, which we call hierarchy filling: even after 400K steps it recovers almost none of the deepest. The recoverability gap is the unrecovered share of an increment, read from one unaligned checkpoint. In short runs that each align one teacher layer at one block, the gap nearly reproduces their ranking by FID improvement, and CKA, a measure of feature similarity, largely reverses it. Representation Alignment and Recoverability Estimation (RARE) picks the teacher layer with the largest gap before training. During training, it tracks each token's remaining distance to that layer, the online counterpart of the gap, weights tokens by it, and phases out the loss once the average distance stops falling. With SiT-B/2 on ImageNet $256\times256$, RARE reaches an FID of 18.02 without guidance and 4.46 with it, ahead of seven alignment baselines including REPA, iREPA and HASTE. It also trains in 14% fewer GPU-hours than iREPA. Its FID stays below iREPA's across model scales, teachers, datasets and backbones.

RLVR provides reliable trajectory-level credit, while OPSD offers dense supervision for token-level credit. This exposes a fundamental coupling when updating step-level credit direction and magnitude with teacher supervision, preventing steps from receiving reliable credit directions and contribution magnitudes, while making both vulnerable to teacher judgment errors and preference variance, as supported by our theoretical analysis. To separate credit direction from its contribution magnitude, we introduce Decoupled Credit Self-Distillation (DCSD), which theoretically decouples credit direction and magnitude into two reliable signals and uses them to calibrate privileged teacher supervision. Specifically, we design belief-margin probing to determine credit direction and marginal information gain to quantify credit magnitude, enabling step-to-token credit assignment for policy optimization. Across 11 benchmarks, DCSD achieves the best overall scores against GRPO, OPSD, RLSD, and RLCSD. Compared with base models, DCSD improves the overall score by 8.45 points on mathematical reasoning and 7.01 points on multimodal reasoning, while correcting the credit direction for 6% of tokens and yielding a 1.5$\times$ reduction in token credit magnitude.

Reinforcement learning with verifiable rewards provides a sparse post-training signal: a single binary outcome evaluates the entire rollout, and every token receives the same sequence-level advantage regardless of its individual contribution. To complement this sparse supervision, a growing family of methods adds a scalar-weighted teacher KL term to the policy-gradient objective, providing dense token-level guidance that may be unreliable at some positions. Despite the benefits of combining these signals, their interaction during optimization can destabilize joint training. To understand how this instability develops, we study the learning dynamics of hybrid reward--distillation training through a neural tangent kernel (NTK) analysis. We introduce the cross-signal NTK $K_{DR}(n)$, a token-level statistic that measures the alignment between reward and distillation gradients at position n. Through this analysis, we identify two failure modes: 1 Magnitude drowning, where the reward gradient exceeds the distillation gradient by orders of magnitude, so that even weak directional conflict can cause the distillation loss to rise despite its explicit inclusion in the training objective; and 2 Localized directional conflict, where the sequence-level advantage and the teacher's position-specific distribution induce opposing updates at the same token ($K_{DR}(n)\!<\!0$). The severity of these effects depends on the optimization regime: the gradient-norm ratio $κ\!=\!\|\nabla\mathcal{L}_R\|/\|\nabla\mathcal{L}_D\|$ varies by roughly an order of magnitude across tasks, and our experiments reveal an empirical threshold beyond which naive mixing can lead to persistent training collapse. Motivated by these findings, we introduce the M3 family, which combines magnitude normalization with three strategies...

Algorithm Distillation (AD) has demonstrated the remarkable ability of Transformers to perform in-context reinforcement learning without explicit weight updates. However, capturing long-term learning progress necessitates expansive context windows, which incur prohibitive memory costs and limit scalability in complex, long-horizon tasks. To address this bottleneck, we propose Recurrent Algorithm Distillation (RAD). RAD employs a dual-component architecture: a Compression Transformer that distills extended interaction histories into compact latent tokens, and an AD Transformer that auto-regressively generates actions using a hybrid context of these compressed memories and recent transitions. By maintaining a fixed-size latent buffer, RAD decouples the effective history length from computational complexity, functionally providing the model with a long-horizon memory. Empirical evaluations across diverse environments demonstrate that RAD matches the asymptotic performance of standard AD with significantly reduced context window sizes, offering a scalable solution for efficient in-context decision-making.

Frozen adaptive cruise control (ACC) policies can violate constraints when deployment dynamics differ from their training conditions. We propose residual-aware conformal action filtering (RACF), which calibrates residuals of a fixed nominal predictor and converts their quantile into an operating margin for finite-model action projection. Completed transitions update margins and candidate selection without retraining the policy. In a registered comparison over 2,400 controller-trial units, Adaptive RACF achieves 94.3% episode safety, improving by 19.9 percentage points over the evaluated nominal CBF-QP baseline while reducing projection frequency from 8.11% to 6.63%. A controlled study isolates a 4.54-point improvement from residual-margin injection. In a separate matched-hardware evaluation, Adaptive reduces mean amortized rollout time by 21.2% relative to Robust CBF-QP, with 161/180 versus 170/180 safe episodes. We characterize conditions linking one-step residual coverage to constraint satisfaction and quantify the observed safety-computation trade-offs.

Target Speaker Extraction (TSE) is pivotal in speech communication and human-computer interaction, enabling the isolation of a specific speaker's voice from complex acoustic environments, i.e., the cocktail party scenario. Although traditional TSE systems conditioned on enrollment speech have progressed substantially, enrollment speech as a cue has inherent limitations. Its reliability degrades when the target and interfering speakers have similar voice characteristics, when intra-speaker variability (e.g. changes in emotion or speaking style) creates a mismatch between the enrollment and target speech, or when the enrollment itself is contaminated by noise or competing speakers. This review surveys deep-learning-based TSE from the perspective of auxiliary target cues drawn from multiple modalities. We organize existing methods according to five types of information used to isolate the target speaker: audio enrollment, visual, spatial, textual/semantic, and neural cues. We also trace the evolution from discriminative estimators to variational, diffusion, flow, codec, and foundation-model-based systems and summarize representative datasets and evaluation metrics. We review the benefits and limitations of different cues and discuss challenges involving synchronization, missing or unreliable observations, data scarcity, privacy, computational cost, and real-time operation. Finally, we summarize future directions concerning adaptive cue fusion, instruction-driven extraction, realistic evaluation, and trustworthy deployment. By jointly reviewing cue design, model architecture, training objectives, datasets, and evaluation metrics, this article provides an overview of the current landscape and open problems in multimodal TSE.

Large Vision-Language Models (LVLMs) have shown strong promise for multimodal reasoning, yet often struggle with tasks requiring concepts beyond what is directly observable in the input image. Existing methods generate intermediate images or latent visual tokens to guide reasoning, but these representations can introduce errors and increasingly interfere with textual reasoning as reasoning progresses. We propose Visual-to-Text Chain-of-Thought Distillation (V2T), a framework that enables LVLMs to internalize visual reasoning without generating intermediate visual representations at inference time. V2T first trains a teacher LVLM using interleaved visual and textual chains of thought, and then uses knowledge distillation to train a student LVLM using the teacher's logits and cross-entropy supervision from ground-truth textual reasoning. When reasoning images can be mapped to the original image, V2T can additionally distill the teacher's attention to corresponding regions, while ground-truth bounding boxes can further guide a subsequent reinforcement learning stage. Experiments across multiple multimodal reasoning benchmarks show that V2T consistently outperforms the teacher and existing baselines, improving average accuracy by 14.3% on a held-out set and 2.7% on the broader visual evaluation suite. Moreover, lightweight SFT and substantially reduced RL make V2T up to 42x faster to train than state-of-the-art baselines.

Data pruning reduces the training cost of knowledge distillation (KD). However, the preferred teacher capacity changes with the data budget: smaller teachers can outperform larger ones when limited training data are available. Understanding what drives this shift is important not only for teacher choice but also for identifying which samples are useful for distillation. We analyze teacher supervision by decomposing it into relational ordering---the ranking of classes---and score geometry---the magnitudes and margins of class probabilities---and show that the small-teacher advantage in the low-data regime arises not only from score geometry but also from relational ordering. Beyond understanding teacher capacity, our analysis reveals two properties of effective subsets: samples should match the difficulty appropriate for the available budget, and their relational signals should be diverse rather than redundant. Based on these findings, we propose DVA (Difficulty- and Volume-Aware data selection for KD), a training-dynamics-free method, which uses a small teacher as a proxy for budget-aware difficulty filtering and class-conditional relational volume maximization. Despite requiring no training dynamics statistics, our method remains competitive with training-dynamics-based methods while consistently outperforming training-dynamics-free baselines.

Socially intelligent language agents must negotiate, coordinate, and resolve conflicting preferences while respecting the time and attention of both participants. Balancing these demands is challenging because agents must convey enough to address a partner's constraints and advance their goals without adding words that do not help the interaction. In this paper, we propose Teacher-Assisted Communication Training (TACT) to improve social goal attainment while reducing communication cost, making interactions with agents more productive and less demanding. We first characterize communication efficiency in terms of action strategy and expression, whose effects extend beyond the current utterance to the partner's response and subsequent exchanges. We design TACT to revise student-generated actions, test the revisions through partner responses, and distill useful feedback into the student. An expression specialist removes unnecessary detail while preserving the intended action, while a strategy specialist proposes alternatives that may better address the partner's constraints. To determine which revision helps, TACT samples a partner response for each candidate and selects a teacher reference by balancing local goal support against action-token cost. That reference guides on-policy distillation on the student's own generation prefixes, allowing the student to act independently at deployment. We evaluate TACT on SOTOPIA and AgentSense. On SOTOPIA, it achieves the highest Goal among the evaluated methods on All and Hard while using substantially fewer target tokens than SFT+SDPO. On AgentSense, it improves goal success over the initial student while reducing target tokens and interaction messages.

On-policy distillation (OPD) is becoming an important component of large language model (LLM) post-training for transferring the reasoning capability of a strong teacher LLM to a weaker student LLM. OPD trains the student by minimizing the reverse KL divergence between the teacher and the student via rollouts generated by the student's policy. However, estimating the gradient of the reverse KL divergence in OPD remains a challenge. Using only the sampled token from the student-generated rollout is computationally cheap but provides limited distributional supervision, which will degrade accuracy. In addition, using the full vocabulary provides complete distributional supervision but is computationally expensive. Therefore, recent works propose Top-$k$ OPD (TK-OPD) that use selected top-$k$ tokens, which provides richer distributional supervision than sampled-token estimation at substantially lower computational cost than full-vocabulary estimation. Unfortunately, using only the selected top-$k$ tokens induces bias, leading to accuracy degradation, as the probability mass outside the selected top-$k$ tokens is discarded. To address the bias of TK-OPD, we propose Tail-Corrected Top-$k$ On-Policy Distillation (TT-OPD). It preserves the advantages of TK-OPD, including rich distributional supervision and low computational cost, while providing an unbiased estimator of the gradient of the reverse KL divergence. The key insight of TT-OPD is to use not only the selected top-$k$ tokens, but also the sampled token from the student-generated rollout, thereby recovering the discarded probability mass in expectation, avoiding the bias. Experimental results demonstrate that TT-OPD significantly outperforms other tested OPD variants.

Fine-grained visual understanding requires models to recognize small details within complex images. Multimodal on-policy self-distillation (OPSD) addresses this challenge by using a teacher conditioned on evidence-centered crops to supervise a student conditioned on original images along student-generated trajectories. Ideally, teacher corrections, the distributional changes from the student toward the privileged teacher, should be driven by task-relevant visual evidence. However, the designs that make the teacher effective also introduce other interference. Using a lagged or frozen teacher improves training stability but introduces a model-state gap from the evolving student, while cropping enhances task-relevant evidence but also loses the visual context. These two sources of interference make the teacher corrections not purely rely on the visual evidence. We introduce Evidence-Aligned multimodal on-policy self-Distillation (EAD), which retains the crop-conditioned teacher as the target but constructs a separate evidence reference for weighting the corrections. To exclude the effect of lagged model-state from this reference, EAD measures prediction changes using the current student. To avoid crop-induced context changes, EAD masks the evidence region in the original image while preserving the other visual context. The change from the student's masked-image prediction to its original-image prediction provides a controlled reference for the direction in which the visual evidence shifts the student's prediction. EAD weights each teacher correction by its cosine alignment with the reference, i.e., retaining aligned corrections and downweighting the rest. Retaining only 6% of the supervision mass of dense OPSD, EAD consistently outperforms previous state-of-the-art methods.

Multi-teacher on-policy distillation (MOPD) combines independently developed domain teachers into a single student by distilling their predictions on student-generated samples. We study a setting where teachers share a reference model but undergo different post-training procedures, and find that MOPD can struggle to recover some teacher capabilities. Because distillation occurs on student-generated prefixes, the student initialization can strongly affect subsequent recovery. However, initial benchmark performance is not a reliable predictor of a good MOPD initialization. For example, merge initialization can start below SFT warm-up yet finish higher after MOPD. We further find that effective merging depends on both the relative teacher contributions and the overall merge scale, with some strong configurations lying outside the simplex of convex parameter averaging. Thus, selecting a good merge initialization requires evaluating not only its immediate performance but also the learning it enables under MOPD, making one-shot coefficient search difficult. We propose Iterative Merging for MOPD (IM-MOPD), which starts from a uniform merge and progressively adds task-vector increments for under-recovered domains during distillation. In a 5-domain setting, IM-MOPD achieves higher average normalized recovery than MOPD with either uniform merge initialization or SFT warm-up, showing that effective teacher contributions can be determined progressively during training.

On-policy distillation (OPD) is a widely adopted post-training technique for LLM reasoning. It is commonly believed to transfer knowledge from a stronger teacher, yet what OPD actually distills into the student's internal representations remains unclear. We study this question with sparse crosscoders, which learn one feature dictionary shared by the student before and after OPD and the teacher. Standard crosscoder analyses, however, identify model-specific features but cannot tell how a model's use of its features changes, since all models are encoded into one set of feature activations. We therefore propose the swap readout, which reads each student checkpoint's feature activations on its own, measuring how training changes the student's use of each feature, even for checkpoints unseen by the crosscoder. Across three OPD settings, we find that OPD neither creates features nor passes on the teacher's own, and leaves the firing rates of over 98% of the student's frequently used features within 20%. We further examine the SFT warm-up on the teacher's rollouts that commonly precedes OPD and makes it more effective. Rather than adding features, the warm-up reweights the shared ones in two ways. First, it already raises and lowers many of the features that OPD later raises and lowers, doing part of OPD's work in advance. Second, it changes features that OPD alone would not, notably those for conversation format, reasoning style, and mathematical notation, and these changes persist through OPD. Imposing this reweighting on a directly distilled student's features, without changing its weights, brings its accuracy close to that of the warmed-up student, whereas the same change on shuffled features does not. Together, these findings suggest that OPD reweights existing features rather than acquiring new ones: the student learns from the teacher how to use the features they already share.

On-policy learning has been argued to reduce catastrophic forgetting, produce sparser parameter updates, and improve generalisation. However, existing comparisons between supervised fine-tuning and reinforcement learning vary many factors simultaneously, making the contribution of rollout policy difficult to isolate. We study the effect of rollout policy in a controlled strong-to-weak distillation setting, by independently varying rollout policy, token-level KL direction, and learning rate across the Llama3 and Qwen2.5 model families and reasoning tasks spanning scientific, medical, and arithmetic domains. Our analysis reveals a nuanced picture of distillation dynamics in which rollout policy does not necessarily play a central role. Instead, token-level KL direction more clearly shapes task performance and output coverage, while learning rate governs forgetting and update sparsity. Analysis of KL gradients and experiments along a continuous student-teacher rollout-policy spectrum explain this pattern: forward KL is remarkably robust to rollout policy, with its performance stable and strong despite changes to the rollout policy, whereas reverse KL is substantially more sensitive and favours student-generated rollouts. On-policy data nevertheless improves generalisation to harder variants of the Countdown arithmetic task under both KL directions, although this advantage does not reliably persist after subsequent RLVR. Our broader conclusions remain robust to removing gradient clipping, using sampled KL estimators, and training on tasks requiring longer reasoning chains. Overall, our results challenge the view that on-policy rollouts are inherently preferable and show that their value depends critically on the objective, evaluation setting, and optimisation hyperparameters.

LLM personalization aims to generate responses aligned with individual users' preferences and needs. User-specific rubrics make these expectations explicit, providing direct supervision on what a satisfactory answer should cover. Existing rubric-guided approaches, however, exploit such guidance only at a coarse granularity, either by using rubrics to supervise the prediction of relevant aspects for subsequent generation or by reducing aspect coverage to a single response-level reward for reinforcement learning. This leaves a gap between specifying what a personalized answer should contain and teaching the model how to generate it. To bridge this gap, we propose GRASP, a rubric-aware on-policy self-distillation framework for LLM personalization that turns user-specific rubric aspects into fine-grained, token-level supervision. Specifically, GRASP pairs a rubric-free student with a rubric-informed teacher that additionally receives the target user-specific rubrics. By aligning their next-token distributions along on-policy trajectories generated by the student, GRASP transfers the teacher's rubric-conditioned guidance into the student, translating user-specific semantic requirements into dense token-level supervision. Since rubric-informed teachers can still produce inadequate supervision, we further introduce Rubric-based Teacher Validation (RTV), which retains only instances where the teacher sufficiently covers the target aspects, improving both supervision quality and training efficiency. Experiments on the LaMP-QA benchmark for personalized question answering demonstrate that GRASP achieves state-of-the-art performance across multiple backbones, supporting the effectiveness of rubric-guided token-level supervision for personalization. To ensure reproducibility, our code is available at https://github.com/SnowCharmQ/GRASP.

Reinforcement learning can turn one language model into several specialists, each excellent at a single skill such as mathematics, coding or following instructions, but users need one model with all of these skills. Multi-teacher on-policy distillation (MOPD) merges them by letting the specialists teach one student: the student answers each prompt, and the specialist for that prompt's domain gives feedback on every token. This routing decides which specialist teaches, but not how strongly its feedback moves the shared student. In Qwen3.5 models at three sizes, we find that MOPD's student does not beat one taught by the best single specialist and gains little of the mathematics specialist's advantage. The feedback is unbalanced: instruction-following feedback is several times more spread out than mathematics feedback and dominates the student's updates. We propose Domain-Normalized MOPD (DN-MOPD), which keeps the routing and rescales each domain's feedback by its measured spread. On six public benchmarks, DN-MOPD improves the average score over MOPD at every size, across three random seeds and under two answer-length limits, and recovers most of the lost mathematics gain. Controls with fixed domain weights show that the gain comes mainly from turning down instruction-following feedback rather than turning up mathematics alone, and that fixed weights close to those DN-MOPD measures perform comparably. Combining specialists therefore requires deciding not only which one teaches, but also how strongly its feedback counts.

Verbal feedback can identify errors and prescribe corrections, providing rich supervision for language-model post-training even when reliable programmatic verifiers are unavailable. Such feedback, often generated by a capable model, can be used to condition the teacher in on-policy distillation, which trains the student to match the teacher's predictions on student-generated rollouts. However, this approach can transfer teacher preferences that the feedback did not motivate, while leaving much of the feedback's guidance unused. We find that both problems come from the standard on-policy distillation objective, specifically the divergence it minimizes and the distribution it uses as its target. Our proposed method addresses both limitations. First, to isolate the information conveyed by the feedback from the teacher's inherent preferences, we treat verbal feedback as evidence for or against the hypothesis that a particular token comes next at a given prefix. We then adopt a probabilistic confirmation framework which uniquely determines an ordering over the vocabulary based on the teacher's predictions before and after it receives feedback. Using a confirmation score consistent with this ordering, we construct a target distribution within a trust region of the student. Second, to learn from guidance that student rollouts can leave unused, we derive a simple shared-rollout estimator of a symmetric divergence between the student and target distributions over rollouts, reusing student and feedback-conditioned teacher rollouts in both directions through importance weighting. Empirical evaluations show that our method outperforms the common on-policy distillation recipe and a recent contrastive variant on knowledge-based and agentic benchmarks.

Active grounding of a frozen diffusion prior requires jointly determining where new measurements should be taken and how they should be used to refine the current reconstruction. Posterior-ensemble-based methods can estimate acquisition utility from generated samples, but require repeated ensemble generation as observations accumulate and capture posterior geometry only through empirical statistics. This paper proposes GPARA, which learns a context-dependent graph surrogate over diffusion prediction residuals, inducing an explicitly reusable posterior response operator that propagates measurement innovations to unobserved variables and evaluates candidate measurements through weighted posterior-risk reduction. Under the matched surrogate, we show that the same response operator also determines expected one-step acquisition benefit and yields an analytic ranking consistent with expected reconstruction improvement. A bounded learned residual calibrates the analytic utility to account for surrogate mismatch, while a small prior ensemble is generated once and reconditioned to update risk weights without repeated diffusion posterior sampling during acquisition. Experiments on two reconstruction tasks spanning physical field and computer vision show consistent improvements in refinement and active acquisition over the evaluated baselines. Ablations further support the complementary roles of step-wise graph refinement, adaptive risk weighting, and analytically anchored calibration.

Flow policies can model rich action distributions, but their iterative sampling limits decision speed. Adjoint matching uses the critic's action gradient to improve a flow policy without backpropagating through its sampling trajectory, yet its supervision is defined for instantaneous velocities. We propose QAMM, a method that turns the critic-derived adjoint signal into supervision for MeanFlow's average velocity. The resulting policy learns finite-interval transport directly and generates actions with few network evaluations. We derive the adjoint MeanFlow target, specify its gradient boundaries, and train it with an offline actor-critic. On ten HumanoidMaze tasks, QAMM produces effective two-call policies and achieves competitive performance against strong flow-policy baselines. These results show that adjoint-based Q optimization can be combined with average-velocity learning to obtain expressive offline policies with few-step action generation.