每天知道
你的领域
出了什么

Transformer-based language models perform well on symbolic tasks, yet it remains unclear whether they learn generalizable rules or rely on statistical shortcuts. Mechanistic studies link algorithmic behavior to structured internal representations, motivating the hypothesis that robust reasoning benefits from separating values from the types that control their manipulation. Can making this separation an architectural primitive improve the learnability and generalization of logical mechanisms? We introduce STRAT (STratified Registers And Types), which partitions the residual stream into orthogonal Data and Type subspaces and uses Type-based attention and gating to govern Data transformations. Controlled arithmetic ablations identify three failure modes associated with data-control interference: the Linear Trap, Gradient Wall, and Open Gate Trap. Mechanistic analysis reveals interpretable logical structure, and in arithmetic, STRAT reduces median OOD error 35-fold relative to a Transformer baseline. On each of 11 datasets spanning 10 tasks, STRAT outperforms the Transformer baseline in mean accuracy, by 26 percentage points on average, with both models trained from 10 base examples per dataset using identical task-specific augmentation where applicable. Under distribution shift, STRAT's mean accuracy drops by only 2.39 percentage points, compared with 11.75 for the Transformer.

A two-hop question makes a language model retrieve twice: once to produce a bridge entity, and once to retrieve with it. Transformers resolve that entity in their early layers. Hybrid models replace most of the attention with a recurrent state, so where the key becomes usable, and where it is spent on an answer, is not known. Answering "Where is the ball?" from "the ball belongs to Alice" and "Alice is in the garden" turns on Alice, a name the question does not mention. A model could reach garden through Alice, the key it computed, or through where the fact sits in the prompt. Across twelve models, dense and hybrid, we move a hidden state from one story into another where the two routes lead to different places, and read off which place the model gives. An attention layer converts the key in every model we tested, and crossing that layer removes its usable effect, all of it where no attention follows. In sequential hybrids this makes the answer a handoff: recurrent layers carry the key forward, and attention spends it. Retrieval does not always stop there. Writing a different fact into the memory of the recurrent layers that follow the last attention layer can move the answer toward that fact, multiplying its odds by 1.3 to 2.7, and which hybrids do this is not settled by their architecture or training. That read is addressed by the key, not by position: recurrent state in a hybrid is not only a carrier, but a memory that later layers can query.

Language models (LMs) often hallucinate by committing to confident answers rather than abstaining, even when they do not have enough information to answer reliably. A large body of existing work mitigates hallucination through detection or abstention mechanisms, but leaves open how models internally arrive at the decision to commit or abstain in the first place. We study this decision through mechanistic analysis, framing hallucination as unsupported commitment: the model commits despite exhibiting signals of unanswerability. Using causal gating, we identify a Commit-Abstain Circuit (CAC), a sparse, causally localised subset of attention heads and MLP sublayers underlying this decision. Across ten LMs (3B-14B) from five families and three benchmarks, the CAC exhibits a recurring accumulate-yet-undercorrect pattern: commitment-promoting components build up commitment in earlier layers, while abstention-promoting components act later as corrective signals that are often insufficient to overturn the accumulated commitment. Building on this finding, a lightweight policy trained on CAC activations improves decision accuracy by 12.2 points over the model's intrinsic commit-abstain margin, reduces false abstentions by 2.5 times, transfers to unseen benchmarks, and extends to larger models (27B-35B). The CAC is both diagnostic, clarifying how models overcommit, and practical, enabling improved abstention decisions.

Multi-step agents are trained on flat action streams: SFT and RLVR weight every token uniformly and ignore the sub-procedures that recur across tasks, the hierarchy that lets humans plan top-down from reusable routines. This structure sits unused, and flat training uses each scarce trajectory less fully than its content allows. Recent agents do use that structure, but only as LLM-written skills in context, never in the weights, so their gains do not generalize beyond retrieval. We instead recover this hierarchy from the data itself and train on it, with no LLM calls. Following text tokenizers, which build a vocabulary by counting alone, we score action spans by reusability and merge canonicalized actions into a reusable eXperience tree (X-Tree). Each X-Tree node captures how a frequent and success-bearing skill is composed from sub-skills, guiding efficient generalization. We integrate X-Tree into three training settings: offline RL, with each node as a training instance; online RLVR, with an adaptive skill bonus; and on-policy self-distillation, with X-Tree as the self-teacher's privileged context. Across WebArena, ScienceWorld, and WebShop at three model scales, X-Tree improves over standard recipes at matched data and budget by up to 4.5% SR on WebArena, 5.8% SR on ScienceWorld and 4.1% success on WebShop. Matched analyses attribute the gains to the X-Tree structure and the three integrations.

Token-level certainty is widely used as a proxy for correctness in LLM training and inference. However, the performance of certainty-based methods depends both on the information in certainty scores and on how those scores are used. We therefore directly assess certainty's predictive ability through controlled empirical evaluations across models and tasks. We distinguish two prediction targets: identifying questions a model is more likely to answer correctly and distinguishing correct from incorrect responses to the same question. In our experiments, certainty is generally better at identifying questions a model is likely to answer correctly than at distinguishing correct from incorrect responses to the same question. Certainty also varies systematically across token types and positions within words, reflecting local properties of words and text form. Information about question difficulty appears early in generation, while the weaker information about answer correctness is more concentrated near the end. These findings show that the information certainty provides for decisions depends on the prediction target, the model, the certainty metric, and which token positions in the response are included in aggregation. We further demonstrate the practical value of these findings for test-time compute. We allocate the number of responses using certainty early in generation and weight answer votes using certainty near the end of each response. Compared with a fixed-sampling majority-voting baseline, this approach increases overall accuracy from 78.71% to 79.54% while reducing generated-token cost by 82.4%.

Multi-agent debate (MAD) reportedly improves reasoning and factuality over single-model inference, but prior work treats agents as symmetric peers, leaving open what drives the gains. We test the hypothesis that cognitive diversity among agents is the driver, in the setting where the question is still measurable: small open-weight models with benchmark headroom. Across 23 models from eleven vendor families, five tasks, and 5,500+ debate and control runs, we vary diversity along three axes - personas, sampling temperature, and model identity - pairing every debate configuration with a generation-budget-matched majority-vote control. The hypothesis is rejected on every axis. Debate beats single-agent inference (3--7 points where tasks have headroom) but at matched budget conditions it ties or even loses to self-consistency sampling at 1.6$\times$ the wall-clock and 3.4$\times$ the token cost. Persona prompting reduces accuracy and a dose-response experiment over each model's full combinatorial persona space shows the cost is a persona tax, not a diversity tax: redundant personas hurt most, while maximally-diverse teams recover part of the loss. Furthermore, mixed-model teams lose to majority votes over their own rosters, with accuracy tracking member capability rather than heterogeneity, and nearly all of debate's benefit comes from the first exchange of answers. We further identify a pervasive measurement hazard in which debate transcripts silently overflow serving context windows, whose correction alone moves our debate-versus-sampling comparison from $-1.8$ points to parity. Our results recast reported MAD gains as an ensemble-sampling effect and provide the budget-matched, contamination-checked baseline bar that future debate mechanisms should be required to clear.

Is human readability necessary for effective fine-tuning of large language models? We investigate whether model-conditioned training representations can preserve or improve adaptation utility without requiring a human-readable textual form. We propose Desired-Update-Aligned Synthetic Data (DASA), which uses activation-gradient feedback from a frozen reference model to guide the optimization of continuous synthetic input embeddings. Inspired by the role of activation gradients in local risk reduction, DASA targets useful adaptation updates rather than source-text reconstruction or linguistic fluency. The resulting embeddings are used directly for downstream fine-tuning; discrete token projections are employed only for qualitative inspection. Experiments on six models from the Llama and Qwen families, ranging from 1B to 32B parameters, cover six benchmarks spanning knowledge, mathematical reasoning, code generation, and commonsense reasoning. Under matched LoRA adaptation settings, DASA achieves performance comparable to the source natural-language data and surpasses it in multiple configurations, while outperforming GRADMM in most comparisons. Further experiments cover general-domain and task-specialized source data. Under the evaluated synthesis settings, DASA provides a $3.6$--$4.9\times$ speedup over GRADMM with comparable peak GPU memory.

Pathology vision-language models (VLMs) are conventionally evaluated by accuracy, but accuracy alone does not measure evidence use: it may conflate dataset contamination, prior knowledge, and image evidence. In a motivating study of lymph-node metastasis prediction, we found that most public pathology VLMs showed minimal differences when changing from feeding the models with whole-slide images, an annotated lesion, or no image at all. To better understand the specific features leveraged by these models, this paper presents two contributions aimed at disentangling these factors. First, we present CleanSlide, a TCGA-based VQA benchmark designed to eliminate image- and question-side contamination. It contains 149K audited multiple-choice questions over 9,985 slides, with patient- and tissue-source-disjoint splits. Every question is audited for option shortcuts, stem leakage, cross-split duplication, and blind solvability. Second, we propose Pair-DPO, a preference loss over counterfactual slide pairs from the same question and source. By controlling for shared confounding factors, Pair-DPO cancels out the question-attributable signal and leaves image evidence as the source of preference. Specifically, each pair consists of two real slides with opposite, verified findings, introducing neither editing artifacts nor unverified labels for diffuse or graded features such as invasion, necrosis, and tumor grade. Experiments show that our method gains 15.29% from image evidence on the CleanSlide, compared with 2.81% for the best published model. On the external CPTAC and BCNB cohorts, our method achieves accuracies of 57.6% and 59.0%, outperforming all other evaluated models by 9.7% and 3.4%, respectively. We will release the benchmark and code.

Language models are now routinely graded by other language models. If post-training makes a model's own writing more predictable, it may also teach the same model, acting as a judge, to reward predictable writing, so that progress on creativity would be invisible to automated evaluation. We follow two open model families, OLMo-2 and Zephyr (7B parameters each), through their public training stages and measure every stage twice, as a writer of short stories and as a judge of pairs of stories. As writers, the models drift as feared: each family's fully trained model finds the stories of its base, supervised fine-tuned (SFT) and preference-trained (DPO) stages progressively more familiar, in all ten prompts, and a model from the other family finds the trained stories 7.0 to 9.2 percent (OLMo-2) and 24 to 25 percent (Zephyr) less surprising per token. As judges, they barely move toward predictable writing. Asked which story is better, a question every judge can use to tell a story from its own words in scrambled order, no trained judge's estimated tilt toward the more predictable story grows by as much as one point in the probability of picking it. A post hoc one-sided 95% upper bound on that growth is 2.7 points on an average pair, about the size of the untrained OLMo-2 judge's own tilt. Asked which is more creative, no trained judge's estimate favors the predictable story more than its base's does. Training instead strengthens a preference for longer stories when the question is creativity, and for one answer slot, and it breaks "more creative" as a question: trained judges asked it no longer reliably prefer a story to its scrambled words. A follow-up could not build pairs that differ in predictability but not in quality, because the routes that made this writer's stories less predictable also broke some of them, often enough to fail a quality floor set in advance.

Reinforcement learning (RL) for code agents often uses executable tests to provide binary rewards. With these rewards, Group Relative Policy Optimization (GRPO) assigns identical advantages to test-passing trajectories within each rollout group, overlooking differences in implementation quality and adherence to task requirements. This leaves the policy without a learning signal that favors clean, targeted implementations over those containing unnecessary or out-of-scope changes. We introduce GAGAR, a framework for quality-aware credit redistribution in code agent RL. Built on dynamic sampling that retains groups containing both passing and failing trajectories, GAGAR places all trajectories from each group in a shared workspace, where an SFT-trained agentic grader jointly inspects them and ranks the test-passing candidates. Based on this ranking, we downweight lower-ranked trajectories and proportionally rescale the advantages of all test-passing trajectories to restore their original sum. This sum-preserving redistribution retains the relative weights established by quality-based downweighting while shifting credit toward higher-quality implementations. We evaluate GAGAR at industrial scale using pre-RL SFT checkpoints of MiMo-V2.6-Flash (310B total parameters) and MiMo-V2.6-Pro (1.02T total parameters). Controlled code-only Flash experiments show improved code agent performance, reduced trajectory-length growth, and more stable training. We further apply GAGAR in large-scale mixed-task RL with both Flash and Pro. Our results support combining test-based verification with groupwise agentic grading to improve the quality and stability of code agent RL.

ELBO-based reinforcement learning offers a sampler-agnostic approach to fine-tuning flow matching models with reward feedback. Timestep weighting in ELBO-based RL has large impact on performance, and it also provides a unified view (as we show in this work) to understand prediction losses heuristically chosen in prior work, yet it remains under-researched and is often chosen to inherit pretrain configs. We investigate impacts and dynamics of timestep weighting in ELBO-based RL. We show that effective weighting depends on both the reward landscape and stage of learning. (1) Through experiments on controlled CIFAR image generation, complemented by robotics, we investigate how weighting impacts reward-driven updates across noise levels. (2) Through gradient analysis, we reveal distinct patterns of cross-noise coordination across tasks and their evolution during training. These findings motivate the hypothesis that useful weighting depends on the gap between the policy's current behavior and the behavior favored by the reward. (3) Guided by this analysis, we study simple static weighting, budgeted profile selection, and dynamic schedules that improve performance beyond conventional target choices. Our results establish timestep weighting as an important design choice for flow-matching RL and motivate further research into methods that choose and adapt it throughout learning.

Reinforcement mid-training lets language models learn internal thoughts from unlabeled text, but efficient token-level credit assignment remains challenging. Existing group-relative methods require costly repeated generation. Learned critics offer single-rollout feedback, but accurate return prediction alone does not ensure reliable policy updates. Our analysis shows how training--inference mismatch and PPO clipping prevent a common offset in advantage estimates from cancelling out, introducing additional update drift. We propose \tfour{}, a twin-critic method that calibrates token-level advantages from a single generated trajectory. After warmup and held-out qualification, the critics provide two advantage estimates, combined using action-dependent weights learned through a conditional-moment saddle-point objective. This objective brings the average advantage at each prefix toward zero, while a signal-retention constraint prevents the correction from erasing the learning signal. Sharing information across text positions avoids repeated sampling of each prefix. Theoretically, we characterize optimal mixing under the signal-retention constraint and establish an upper bound on residual mean-induced drift. Experiments show that, compared with the state-of-the-art critic-free method, \tfour{} improves mean benchmark performance by 7.8% and reduces mean training-step time by up to 63.4%.

Online safe reinforcement learning (RL) seeks policies that maximize reward while satisfying safety constraints. Reward and safety can induce multimodal action distributions, challenging the prevailing primal-dual methods: Gaussian actors may collapse onto a single suboptimal mode, and optimization over the nonconvex Lagrangian landscape can be unstable. Diffusion and flow policies can represent such distributions, but recent work with a diffusion actor relies on estimating and matching the score of an augmented-Lagrangian target policy. Instead, we differentiate the augmented objective directly through the generation path of a flow policy, so no score needs to be estimated. Because a flow policy lacks a readily available action log-density for entropy regularization, we build on the density-free kinetic-energy regularizer of FLAC, a recent reward-only method, and propose Reparameterized Augmented-Lagrangian Flow Actor with Least Energy (RAFALE), an off-policy actor-critic method for safe RL. We formulate its update as a constrained one-ended generalized Schrödinger bridge and show that, for each source draw, this path-space problem is exactly an entropy-regularized problem in action space. At positive noise, its solution reweights the reward-only action distribution only where the estimated cost exceeds a threshold set by the Lagrange multiplier. As the noise vanishes, the optimal value converges to that of a least-energy map objective that the flow policy optimizes directly. Across seven Safety-Gymnasium tasks, RAFALE achieves competitive reward with mean final cost within budget on every task, whereas strong baselines trade one for the other; ablations support the necessity of both its augmented objective and its flow actor.

Data shape determines how features are structured, how patterns are separated, and how distributions cover the underlying domain. Poor data shape can make models learn noise rather than generalizable structure. This paper studies robust feature-centric data reshaping: generating feature transformations that remain useful, stable, and reproducible under noisy evaluation and imperfect data conditions. We view reshaping operation sequence search as reward-guided diffusion generation, and robust reshaping as searching for regions in the latent reward landscape rather than isolated high-reward transformations. The key challenge is dual instability: noisy evaluators distort local reward guidance, while stochastic generative trajectories can converge to inconsistent solutions. We propose FCDiff, a flat-consensus diffusion framework that addresses both failures through a micro-macro decomposition. The micro layer replaces point-estimate reward guidance with Gaussian-smoothed, Monte Carlo averaged gradients, steering generation toward locally flat reward regions. The macro layer aggregates independently guided trajectories with a weighted Frechet-mean barycenter, selecting consensus-supported basins and filtering stochastic outliers. Across an 8-dataset headline cohort under heavy-tailed evaluator noise, FCDiff attains the best aggregate rank on lower-tail reliability and robustness against both search-based AutoFE and robustness-oriented generative baselines, with statistically significant accuracy gains over every generative baseline. Our results show that robust data reshaping requires searching for flat, consensus-supported regions rather than sharp single-trajectory optima.

Reinforcement learning with verifiable rewards has substantially improved mathematical reasoning. However, terminal correctness alone provides limited insight into the quality of high-level strategies, such as theorem selection and subgoal decomposition, when considered separately from their subsequent execution. This paper studies strategy utility, which is defined as the likelihood that a strategy supports a correct downstream solution under a given executor. We introduce SURE, a framework for learning and leveraging relative strategy utility. In this framework, high-level strategies are separated from their detailed reasoning. Based on the pairwise preferences constructed from strategy-conditioned rollouts and teacher-generated contrasts, a Strategy Reward Model is learned to estimate relative strategy utility. During reinforcement learning, the frozen reward model reads only the extracted strategy, whose score is combined with the correctness and format rewards in a sequence-level GRPO objective. Compared with outcome-and-format GRPO baselines, experiments show that SURE improves average pass@1 by 1.87%, 2.64%, and 2.93% across three policy backbones. Our method also achieves competitive or better accuracy than stronger reward baselines while requiring substantially lower GRPO-stage compute.

Long-horizon software engineering (SWE) agents trained with reinforcement learning with verifiable rewards (RLVR) typically receive only terminal outcome supervision, making it difficult to distinguish productive actions from redundant exploration or functional regressions. We propose SWE-MILE, an asynchronous potential-induced milestone credit assignment framework that derives fine-grained process supervision from workflow runtime, without auxiliary reward models or external evaluators. SWE-MILE quantifies task-relevant file exposure and test-state alignment as navigation and verification potentials, respectively. Differences in these potentials attribute milestone progress and regressions to individual actions, while discounted backward credit propagates supervision to preceding steps. To efficiently acquire intermediate verification states, SWE-MILE further introduces asynchronous shadow probing, which replays repository-changing actions in an isolated sandbox and runs verification in parallel with the agent's primary interaction, largely hiding verification latency. The resulting process credit augments terminal outcome advantages and provides informative learning signals. Experiments on two representative long-horizon SWE tasks demonstrate substantial improvements in agent performance, highlighting workflow runtime signals as a practical source of process supervision for long-horizon SWE agents.

Comparative Process Reward Models (PRMs) provide critical step-level guidance for autonomous web agents by evaluating state-conditioned preferences between candidate actions. However, existing preference training data synthesized via multi-policy sampling suffers from a severe scarcity of Grounded Minimal Contrastive Pairs (GMCPs)-where competing candidates target genuine on-page elements with identical action types. In representative baselines preference data (namely, WebArbiter), GMCPs account for merely 24.19%, biasing PRMs during training to rely on shallow shortcuts (such as element hallucinations and action type mismatches) rather than acquiring genuine contextual decision semantics. To address these challenges, we propose SURFPRM, a graph-guided process preference synthesis framework for comparative Web PRMs. SURFPRM structures web demonstrations into a persistent Interaction Element Graph that acts as an environment-grounded negative action proposal mechanism, systematically synthesizing contrastive negative actions across spatial, temporal, and spatiotemporal confusion axes. This elevates the GMCP proportion from 24.19% to 74.60%, producing the curated SURFPRM-DATA dataset. Across six open-source backbones (3B to 9B parameters), PRMs trained on SURFPRM-DATA outperform baseline-trained models on average on WEBPRMBENCH and rival leading proprietary LLMs. In downstream reward-guided trajectory search on WEBARENA-LITE, SURFPRM provides step-level guidance for both GPT-4o (+14.21%) and GPT-4o-mini (+12.83%) policies, yielding substantial improvements in complex web task success rates.

Complex video reasoning often depends on evidence scattered across distant moments, entities, and events, yet a correct answer alone does not reveal whether a model relied on the right parts of the video. We introduce ASSEMBLE, a framework that makes supporting evidence explicit throughout long-video reasoning. ASSEMBLE organizes local observations and cross-clip narratives into timestamped evidence catalogs traceable to the source video. A grounding-aware reader then composes question-specific atomic skills whose structured outputs contain explicit evidence references and support assessments. We use correctness-gated citation alignment as a direct grounding signal: after teacher-supervised fine-tuning, Group Relative Policy Optimization (GRPO) jointly optimizes answer correctness and citation alignment. This produces inspectable intermediate traces while keeping final predictions linked to explicit supporting evidence. Using a 9B reader supervised by a 235B teacher and shared precomputed evidence catalogs, ASSEMBLE achieves 59.2% macro-averaged answer accuracy across three long-video reasoning benchmarks, compared with 58.3% for Gemini-2.5-Pro, while improving macro-averaged overlap-based Grounded accuracy by 6.7%, with gains on all three benchmarks. Ablations further show that, with the same post-trained reader and inference budget, structured skill inference improves Grounded accuracy over free-form reasoning. Together, these results show that explicit evidence grounding can be integrated directly into long-video reasoning without sacrificing answer accuracy.

Reward models score responses from large language models (LLMs) and guide LLM training toward human preferences. However, reward models can favor superficial attributes such as length or confidence, leading LLMs to produce higher-scoring but not more correct responses. Existing mitigation methods either retrain the reward model or apply a fixed correction to one known bias, such as a preference for longer responses. Retraining requires additional data and computational resources, while existing editing methods require the target bias to be specified in advance and use a fixed edit for that bias. To this end, we propose BiasReducer, a lightweight framework that edits only the linear reward head and selects the relevant edits for each new dataset. First, BiasReducer uses a sparse autoencoder (SAE)-style encoder to learn which attributes (e.g., length and confidence) the reward model is sensitive to. Second, it learns how to reduce the reward model's dependence on each attribute by determining which direction to adjust the reward head and how much to adjust it. Third, for a new dataset, it ranks the attributes by their influence on reward scores, selects the relevant ones, and edits the reward model accordingly. BiasReducer consistently improves reward-model robustness to biases toward superficial response attributes. Across five reward models, BiasReducer-M improves the three benchmarks by 8.3, 18.0, and 6.9 percentage points on average, outperforming the two training-based baselines. The gains transfer downstream, reducing unnecessary verbosity and sycophancy while maintaining comparable judged quality.

Reinforcement Learning (RL) has proven effective in aligning flow-based generative models with human preferences. Recently, Flow-GRPO has emerged as an efficient critic-free paradigm by calculating advantages over sampled candidate trajectories. However, standard Flow-GRPO applies a uniform scalar advantage across both temporal denoising steps and spatial latent dimensions, without explicitly accounting for the spatial structure of generated images, which may lead to sub-optimal policy updates. To address this, we propose a novel gradient-guided spatial credit assignment framework tailored for Diffusion Transformers (DiTs). We first reformulate the transition-level log-likelihood in Flow-GRPO into a token-wise representation natively aligned with DiT patch architectures, constructing spatially fine-grained importance sampling ratios. To allocate localized credit without rigid, boundary-sensitive segmentation heuristics, we introduce a continuous spatial credit map derived from reward gradients. Crucially, we employ an outlier-robust normalization scheme based on Median Absolute Deviation (MAD) coupled with temperature scaling, effectively eliminating gradient noise while highlighting functional prompt-aligned regions. Extensive evaluations on GenEval show that our approach delivers SOTA alignment quality, achieving a convergence rate comparable to top-tier methods like DiffusionNFT while substantially improving upon Flow-GRPO-based methods in alignment performance.

Direct Preference Optimization (DPO) aligns language models by optimizing over sequence-level sums of token-wise implicit reward differences. However, we identify a pervasive pathology in this formulation: a disproportionately small subset of high-frequency token types dominates cumulative sequence scores while appearing symmetrically across both preferred and dispreferred responses. Specifically, under canonical Qwen tokenization on Anthropic HH-RLHF, merely 69 token types account for $55.1%$ of all response tokens and $85.9%$ of within-pair shared token mass, exhibiting substantially lower preference-side specificity than the remaining vocabulary. This symmetric ubiquity induces gradient entanglement and dilutes the discriminative preference signal propagated through the objective. To resolve this issue, we introduce Anisotropic DPO (\textsf{ADPO}) and its canonical realization, Frequency-Hard DPO. Using a fixed, label-agnostic vocabulary mask, our method zeroes the implicit reward contribution of high-frequency response tokens while assigning unit weight to informative positions, thereby suppressing gradient interference without modifying preference pairs, discarding context, or introducing learned parameters. Here, anisotropy designates non-uniform token-level objective weighting rather than representational geometry. Extensive empirical evaluations on AlpacaEval, MT-Bench, and Arena-Hard demonstrate that Frequency-Hard DPO consistently outperforms standard DPO across Qwen-2.5-7B-Instruct and Llama-3-8B-Instruct, establishing that selectively masking shared high-frequency tokens offers an effective, zero-overhead mechanism for robust preference alignment.

Domain experts trained from a shared checkpoint can transfer their specialized capabilities to a single student through on-policy distillation (OPD). Existing research primarily focuses on improving this merging process, while the algorithms used to train the experts have received limited systematic comparison. We investigate which training algorithm produces teachers better suited to OPD through controlled single-teacher comparisons of supervised fine-tuning (SFT) and reinforcement learning (RL) across Agentic, Reasoning, and Perception. Teachers and students share the same Qwen3.5-9B initialization, and the two teacher types are compared at similar task performance. Our experiments show that RL teachers yield stronger students and higher recovery of teacher performance gains across all three domains. At their best checkpoints, RL-guided students outperform SFT-guided students by 4.27, 1.50, and 0.86 percentage points, respectively. In Agentic, the best SFT-guided student recovers only 44.44% of its teacher's performance gain over the base model, whereas the best RL-guided student recovers 115.00%, surpassing its teacher. Our further analysis shows that RL teachers undergo smaller parameter displacements from the shared initialization than SFT teachers. These findings support the hypothesis that RL teachers' smaller departures from the student's starting point facilitate learning through OPD, resulting in stronger students.

Distillation enables student language models to acquire new capabilities from expert teachers. However, integrating knowledge from multi-teacher, multi-domain demonstrations into a single student remains challenging. We study supervised fine-tuning (SFT) in this setting, where students must acquire diverse capabilities while maintaining generalization beyond the training tasks. Our experiments reveal varying trade-offs between in-distribution learning and out-of-distribution generalization across SFT methods, motivating more explicit control over this balance. To this end, we propose soft-target fine-tuning (SoFT) to balance learning from teacher demonstrations with retaining the Base model's existing capabilities. SoFT sets a minimum target probability for each demonstrated token while making the smallest KL change to the Base distribution. The resulting objective couples learning from demonstrations with adaptively weighted regularization toward the Base model. We further use domain-specific gradient budgets to control this balance and determine a probability threshold for each trajectory. Experiments on mixed-domain reasoning and agentic tasks show that SoFT achieves the best overall performance among the compared methods, with improvements in both in-distribution capability acquisition and out-of-distribution generalization.

Amortized neural likelihoods enable computationally expensive inference for models with analytically intractable or unspecified likelihoods, but their black-box nature limits interpretability. We introduce a symbolic distillation pipeline that converts trained neural likelihoods into explicit, interpretable expressions optimized for efficient parameter estimation. Our approach uses a recovery-directed objective to guide symbolic regression toward expressions that preserve parameter-recovery accuracy rather than merely approximating the likelihood function. Candidate expressions are evaluated on held-out datasets and selected using a criterion that jointly accounts for expression complexity, parameter-recovery performance, and distributional distance from the learned likelihood. We evaluate the pipeline on the diffusion decision model, a classical cognitive model, whose analytically tractable likelihood provides ground truth for controlled evaluation. The proposed recovery-directed objective improves parameter recovery over standard symbolic-regression objectives. The resulting symbolic likelihoods enable over 100 times faster parameter evaluation than both neural likelihoods and, when available, the exact likelihood, while maintaining a manageable loss in precision. We further demonstrate these computational benefits in Bayesian hierarchical inference on empirical data. Our pipeline provides a lightweight interface for integrating symbolic distillation with existing neural-likelihood estimation methods and can be adapted to a range of simulation-based inference settings.

Flow matching and diffusion can be trained to predict different quantities, most commonly the data $x_1$, the source noise $x_0$, or the velocity~$v$. Although theoretically equivalent, these can lead to substantially different performances. We identify two main drivers for these differences: the source--data signal-to-noise ratio, and the information bottleneck induced by the neural architecture. We show that, beyond intrinsic data dimension, the factor affecting the optimal parametrization the most is a certain signal-to-noise ratio in each data covariance direction. From this analysis, we introduce new spectral hybrid parameterizations that adapt across time and data covariance directions; we show that these are optimal for Gaussian data. We also show that architecture-induced compression changes which parameterization is easier to learn, with $v$-prediction being more sensitive to discarded directions than $x_1$-prediction. Experiments across architectures and source scales show that our spectral parameterizations are robust across regimes, can substantially accelerate optimization, while incurring essentially no additional training cost compared with standard parameterizations.

Data-driven model predictive control (MPC) combines learned world models with online trajectory optimization, achieving strong performance in continuous control. However, the per-step cost of sampling and evaluating hundreds of candidate trajectories restricts deployment to control frequencies well below what real-time robotics demands. Motivated by the dual-process theory of human cognition, which distinguishes between fast, intuitive processing (System 1) and slower, deliberative reasoning (System 2), we ask whether every decision requires the same degree of computational deliberation. We propose Fast-TD-MPC, a lightweight framework that adaptively routes between fast policy execution and test-time planning, reserving costly deliberation for states where it is most needed. Fast-TD-MPC delivers competitive task performance across 103 continuous control tasks while achieving up to ~4x faster inference. Under external disturbances, Fast-TD-MPC selectively falls back to planning, maintaining robustness comparable to the original planner.

Code-as-Policy agents accomplish long-horizon embodied tasks by generating and executing code, yet continually improving them with teachers that are stronger but not globally reliable remains a key challenge. Existing distillation methods typically treat the teacher's complete behavior as the supervision target and thus misassign training credit on states where the teacher fails. We propose RE-0, a recursively verified policy improvement framework: rather than assuming that the teacher globally outperforms the student, RE-0 requests local corrections from the teacher on the student's own failure histories and checks in the environment whether each correction is genuinely beneficial; verified corrections yield immediate improvement. Building on this, we propose RE-OPD, which turns verified interventions into supervision for on-policy distillation. Only counterfactually verified teacher interventions provide distribution-level supervision, weighted by their measured local benefit, and the improvement they induce is projected back into the standalone student, so both where supervision is applied and how much credit the teacher receives co-evolve with the student policy. We further prove that the student's per-round gain is lower-bounded by its verified intervention gain up to verification and projection error terms. Experiments across multiple Code-as-Policy embodied tasks show that RE-0 improves both teacher-assisted execution and the standalone student, and generalizes to novel robots and scenes.

On-Policy Context Distillation (OPCD) has recently emerged as a powerful technique for transferring context to student models and for self-improvement. In OPCD, the teacher is conditioned on privileged information, and the goal is to minimize the Kullback-Leibler (KL) divergence between the privileged teacher and the student, evaluated on student-generated tokens. Many existing studies show that using instance-specific gold answers or gold demonstrations as the default privilege can hurt training performance, especially out-of-distribution (OOD). In this work, we instead design general instructions that target common student mistakes observed on the training samples, and show that such simple instructions can outperform gold as the OPCD privilege. In autoformalization tasks, using a matched formatting instruction as the privilege could outperform gold in OOD accuracy by a large margin. In 7 out of 8 experiments using ProverQA, ProofWriter, and ProntoQA as datasets, and Qwen3-Thinking and Olmo3-Thinking families as models, matched instruction privileges outperform gold in OOD by 4 to 17 points, while remaining on par with gold in-domain. Each instruction is only a few sentences (and thus contains much less information compared to all instance-specific gold) and is applied uniformly to every training sample. These results indicate that a general instruction, which applies equally to source and target domain examples, can be substantially more transferable than instance-specific gold in OPCD while maintaining in-domain performance.

On-policy distillation (OPD) is a promising approach to language-model adaptation, aligning teacher supervision with the student's own generated trajectories. When adaptation prompts are distributed across clients, can this process benefit from federated collaboration? We study federated OPD and find that substantial collaboration gains can be obscured by learning-rate sensitivity: FedAvg can perform no better than independent local training at a small learning rate, yet recover a clear advantage at a larger rate. We explain this phenomenon through the student's dual role as learner and generator of future training data. An aggregation-induced optimization lag can delay access to useful teacher supervision, which in turn slows subsequent learning. Our theory establishes this aggregation--rollout feedback in a solvable model with a common optimum and stable updates, and identifies two coupled roles of learning rate: learning from current supervision and reaching future supervision. Guided by this analysis, we propose FedTOPS (Federated Teacher-guided On-Policy Scaling), which reuses teacher feedback on current trajectories to adapt the FedAvg update magnitude under clientwise predictive-change constraints. Across six mathematical reasoning benchmarks, FedTOPS improves macro Avg@8 over FedAvg by 4.56--14.57 percentage points across the evaluated student models and local learning rates.

On-policy self-distillation (OPSD) trains a question-only student with token-level feedback from a teacher given training-only privileged information (PI). OPSD therefore provides dense, on-policy supervision, and is free of a larger external teacher, but its effectiveness rests on how PI is designed and utilized. Our preliminary diagnostics suggest a significant gap between teacher utility and student learnability, where a small fraction of high-disagreement tokens dominate the distillation signal, and short teacher continuations at these positions further expose more explicit PI leakage than transferable correction cues, indicating a strong intent on injecting PI-conditioned shortcuts. We propose Adaptive On-Policy Self-Distillation (AOPSD), which adapts what information the teacher receives and how strongly its feedback influences learning. AOPSD encodes each solution as a reasoning DAG, orders problems by the student's evolving capability, and reveals only the affordable subgraph and its next frontier as PI. For high-disagreement tokens, AOPSD utilizes short teacher continuations as probes to encourage useful guidance while mitigating PI-conditioned shortcuts among teacher supervisions. On HMMT25, AIME24, AIME25, and BRUMo25, AOPSD achieves 72.5% Pass@8, which is 6.7 percentage points above OPSD and 4.2 above the strongest competing baseline while reducing 15 percentage points of training time at lower cost.