Sample-Level Supervision

Modern vision-language models (VLMs) can directly answer many image-grounded questions, yet they often struggle with complex queries requiring fine-grained visual details or external knowledge. To acquire this missing evidence, agentic VLMs invoke tools such as image cropping, image search, and text search. However, existing training paradigms primarily evaluate tool-use based on final answer correctness, leaving evidence acquisition and utilization insufficiently supervised. This leads to two critical shortcomings: (i) models frequently issue redundant or off-target tool calls that fail to gather necessary evidence, and (ii) even when appropriate tools are called, models often fail to extract the necessary information from the resulting observations. To address these limitations, we introduce the NTEP (Necessary Tool-Evidence Path), a novel annotation scheme that explicitly specifies the essential external evidence and corresponding tool calls for each query. Building upon this, we propose NTEP-R (NTEP Reward), a supervision mechanism ensuring that each tool invocation strictly advances the reasoning process toward the final solution. Specifically, our approach rewards the agent for aligning its pre-call intent with a necessary evidence-seeking goal, and for ensuring the information summarized from the post-call observation aligns with the necessary evidence. Furthermore, we introduce a non-repeated-goal regularizer to penalize redundant calls that revisit satisfied NTEP goals. Extensive evaluations on seven image-grounded benchmarks demonstrate that our 8B-parameter instantiation, NTEP-8B, significantly improves both search-oriented accuracy and tool-use efficiency within a unified three-tool framework. These results highlight the critical value of fine-grained tool-evidence path supervision for training robust agentic VLMs.

Direct Preference Optimization (DPO) simplifies alignment through pairwise comparisons but assumes all observed preferences are reliable. Real data often violates this assumption, leading to reversed, weak, or ambiguous labels that cause harmful policy updates. To address this, we propose Posterior Label Correction DPO (PLC-DPO) to robustly optimize preferences by routing each pair's training signal as a clean, flip, or tie case. The key idea is to use the calibrated policy-reference margin as online evidence to take appropriate correction actions. This reframes noisy preference learning as actively correcting supervision direction and strength rather than merely filtering suspicious examples. Across 57 dataset-model-benchmark cells, PLC-DPO obtains the best mean win rate against DPO (60.5 vs. 55.5 for the next-best method). Injected-noise and tie stress tests, human disagreement analysis, and self-confirmation diagnostics further show that the routing remains stable and distinguishes flipped from weakly directional pairs.

Safety tuning can improve harmful refusal, but models may learn surface-form shortcuts: wrapped harmful prompts bypass safety, while similarly wrapped benign prompts are over-refused. We propose Wrapper-Based Intent-Form Augmentation (WIFA), an automatic intent-group augmentation method that pairs wrapped harmful examples with structurally matched wrapped benign counterexamples, requiring no external teacher or manual per-wrapper intent labels. We use WIFA as a common data layer for two complementary fine-tuning routes: WIFA-Boost, a two-stage high-safety recipe, and Anchored Group-Consistent Refusal Training (A-GCRT), which regularizes refusal/compliance decision scores across same-intent wrappers and anchors harmful and benign groups on opposite sides of a margin. In the Qwen setting, WIFA-Boost reaches the strongest transformed-harmful refusal, while A-GCRT reduces OR-Bench over-refusal from 25.7% for the base model to 17.4%; reproduced baselines do not match these operating points. Llama results and ablations over data structure, two-stage order, and A-GCRT components support this intent-group interpretation without claiming universal below-base over-refusal.

Pairwise constrained clustering typically relies on hard must-link/cannot-link labels, whereas realistic pairwise supervision may be real-valued and entangle intrinsic ambiguity, expert judgment, and stochastic corruption. Existing deep constrained clustering (DCC) methods mainly target hard, expert-agnostic constraints, treating soft labels mostly numerically rather than semantically. We formalize this setting as uncertainty-aware probabilistic constrained clustering (UPCC), defining a canonical aleatoric target through a heterogeneous observation process and analyzing its conditional identifiability. We introduce ProbPair, an angular pairwise objective for probabilistic relations, and build ECI-PP, an estimator--corrector--integrator framework that refines imperfect supervision via belief estimation, correction, and reliability-aware integration. Across challenging probabilistic supervision settings, experiments on diverse benchmarks show that ECI-PP outperforms state-of-the-art DCC methods and remains robust with a shared default configuration.

Emerging World-Action Models (WAMs) have demonstrated promising performance in autonomous driving by jointly modeling future driving scene evolution and trajectory planning. However, existing WAMs are typically trained with video data, which is only 2D projections of the underlying 4D driving scene. Consequently, WAMs fail to understand and capture the structure of 4D scenes and thus generate visually plausible yet 4D inconsistent future predictions that mislead downstream planning. To alleviate this issue, we present 4D-WAM, a model that leverages geometric foundation models for training-time supervision to enable 4D consistent world modeling. Specifically, we feed WAM-predicted future frames into a geometric foundation model, and use 4D-aware responses to define a 4D consistency loss. This loss encourages the model to understand, represent, and predict physically consistent 4D scenes during training, without additional inference cost. Moreover, we identify an early-decision phenomenon in WAMs and propose a decision-oriented timestep sampling strategy that emphasizes supervision at early, high-noise stages, where driving decisions are primarily formed. By propagating 4D supervision to this critical decision-formation phase, the proposed strategy further improves trajectory planning. Extensive experiments demonstrate that 4D-WAM effectively models 4D consistent scene evolution and achieves state-of-the-art performance on challenging NAVSIM-v1 and NAVSIM-v2 benchmarks.

Refurbishment-based noisy-label learning mixes an observed label with a model-derived pseudo target, typically using one sample-wise cleanliness score to control both branches. This creates a hidden coupling: reducing trust in the observed label automatically increases trust in the pseudo target. We show that this complementarity can replace one unreliable signal with another because a pseudo target learned from corrupted supervision may reproduce the noise it is meant to correct. Our representation diagnostics provide a consistent account of this mismatch: noisy supervision redirects deeper layers more strongly, whereas shallower relations remain comparatively stable and provide information beyond the loss posterior. We therefore propose TRACE, a Two-Source Reliability Assessment framework for Label Correction and Sample Reweighting. TRACE assesses the observed label using loss fit, shallow relation stability, and prediction agreement, while separately assessing the pseudo target using model confidence. Its source-specific scores control target correction and supervision strength without assuming complementary reliability. Across synthetic and real-world noisy benchmarks, TRACE improves representative refurbishment baselines and yields more reliable pseudo supervision.

In person re-identification, neighbor-based methods have achieved significant success by interacting with neighbor samples to obtain more robust representations. However, existing methods rely only on affinity relations, causing their success to depend heavily on the reliability of selected neighbors. We find that affinity-only interaction often fails in challenging scenarios due to the inevitable presence of noisy neighbors. To enable effective interactions under noisy neighborhoods, we revisit neighbor-based methods under distinct reliability conditions and propose a novel Adaptive Neighbor Feature Interaction (ANFI) method. The core idea of ANFI is to account for negative effects from noisy neighbors, allowing samples to remain distinguishable from false positive neighbors. Unlike existing methods, ANFI models not only affinity relations but also discrepancy relations, and employs sample-wise adaptive weighting for these two types of relations. Given that capturing negative effects from noisy neighbors differs significantly from traditional relation learning, we derive discrepancy relations from a new neighborhood similarity, which provides more information than pairwise similarity. In addition, we propose Noisy Relation Supervision (NRS) to train ANFI, gradually injecting robustness to noisy relations into the model. Extensive experiments conducted under standard, cross-modal, and cross-domain settings, including comparisons with neighbor-based methods and re-ranking methods, demonstrate the superiority of our method across various neighbor distributions.

Video face swapping has no natural paired supervision: no real footage exists of one person's face performing another person's video. The strongest current answer, DreamID-V's SyncID-Pipe, mints pairs by replacing the identity in exactly two frames of a real clip -- the first and the last -- and regenerating the rest from a pose sequence alone. Pose carries no appearance evidence of the swapped-in identity, so over long clips, occlusions, and extreme pose excursions the synthesized identity has a long unanchored span on which to drift; no published ablation examines anchor count or placement. We propose Adaptive Identity Anchoring (AIA): (i) generalize the synthesizer to arbitrary anchor sets, architecturally natural for diffusion-forcing-style transformers where conditioning on a frame is clamping its tokens to zero noise; (ii) place anchors by a closed feedback loop that scores every generated frame against the real reference identity and inserts an image-face-swapped anchor at the worst-scoring frame until the pair passes a threshold or exhausts a budget; (iii) reuse the loop's verdict as an automatic data filter. A second pathology, the beauty-filter look of over-smoothed skin, has the same root cause: micro-texture, like identity, is priced by none of the pipeline's objectives. We therefore pair AIA with Reality-Referenced Texture Restoration: matched re-graining from each real frame's non-face regions, band-split transfer of sub-identity micro-texture from the real footage, and a second, spectral acceptance channel refereed by the footage's own spectrum. Identity-anchor density, we argue, is a controllable quality dial, and we specify falsifiable experiments -- drift-versus-gap curves, uniform-versus-adaptive placement at matched budgets, student training on AIA-minted data, and texture ablations with a human beauty-filter study -- that would validate or refute the proposal.

While autonomous coding agents have significantly advanced automated test generation, they remain fundamentally limited by lazy generation, a phenomenon where agents prematurely terminate tasks and systematically avoid complex programmatic logic, resulting in inadequate code coverage. Currently, mitigating this premature termination requires continuous human-in-the-loop supervision. This heavy reliance on human intuition creates a bottleneck that negates the efficiency gains of automated generation. We propose SCATE, a framework for adaptive, automated supervision of coding agents that replaces human intervention during test generation. By formulating supervision as a contextual bandit problem, SCATE learns to select the most promising testing actions based on the current coverage and class testability metrics, maximizing coverage gains while minimizing wasted generation effort. Our empirical evaluation demonstrates that SCATE integrates seamlessly with different coding agents. When applied to GEMINI-CLI, it achieves 32.3% higher line coverage and 30.9% higher branch coverage than the agent-only baseline. A comparison with CLAUDE CODE confirms the framework dynamically adapts its policy to optimize each agent's unique strengths. SCATE also consistently outperforms state-of-the-art non-agentic approaches across all metrics.

Multimodal Large Language Models (MLLMs) often follow a fixed Think-then-Answer paradigm, which is inefficient in heterogeneous multitask settings because simple inputs may not require explicit reasoning while difficult ones can benefit substantially from it. Learning when to think is also unstable during post-training, where imbalanced rollouts can drive the model toward always-thinking or always-direct behavior. We propose Switch-Reasoner, a GRPO-based framework that learns to adaptively select reasoning modes for MLLMs. It treats thinking as a virtual tool invocation and allows the model to either answer directly or invoke explicit reasoning before answering. To stabilize this decision, we introduce a dual-level regulation mechanism that balances the overall use of Thinking Mode and Direct Mode while providing sample-level supervision based on the relative benefit of the two choices. Experiments on 11 multimodal tasks show that Switch-Reasoner reduces unnecessary reasoning while maintaining strong performance, achieving a better accuracy-efficiency trade-off.

Subjective NLP tasks often exhibit systematic annotator disagreement, requiring models that represent uncertainty rather than collapse it. We introduce Ensemble Diversity Optimization (EDO), a prediction-space framework that jointly optimizes ensemble weights, effective cardinality, and calibration through a unified differentiable objective. EDO learns ensemble composition and size end-to-end via Gumbel-Softmax relaxation and incorporates a signed diversity regularizer, tuned on validation data, to steer optimization toward either preserving or suppressing disagreement. This regularization prevents ensemble collapse and enables controlled navigation of the utility-calibration trade-off. The framework integrates a soft F1 surrogate, class-weighted cross-entropy to address imbalance, and reliability-weighted diversity to regulate intra-ensemble variability. Experiments on four subjective text-classification benchmarks (ArMIS, ConvAbuse, HS-Brexit, MD-Agreement) show that EDO substantially improves probabilistic calibration, reducing cross-entropy (40-78% depending on baseline) and lowering Brier scores relative to Soft-CE, Soft-MD, Top-5 Voting, and WEL, while maintaining competitive F1 and better alignment with annotator distributions. These results demonstrate that jointly optimizing ensemble structure with a signed diversity regularizer provides an efficient, model-agnostic approach for modeling human subjectivity in supervised learning.

Chart-to-code generation is commonly trained through supervised fine-tuning on reference plotting scripts, implicitly treating the gold code as a fully observable target. However, many chart programs contain latent variables that cannot be uniquely recovered from the rendered image. We identify this latent-observation mismatch in four forms across five chart types: aggregation-induced mismatch, where raw samples are reduced to box statistics or histogram bin masses; normalization-induced mismatch, where absolute scale is removed in pie charts; projection-induced mismatch, where 3D information is lost through 2D rendering; and level-set-induced mismatch, where a scalar field is observable only through selected contour lines. These mismatches introduce target ambiguity and require models to generate information unsupported by the image. We propose Observation-Aligned Supervision, which replaces latent variables with visually constrained quantities. We instantiate it using box statistics, bin weights, and wedge proportions, and study 3D scatter and contour charts through controlled experiments. Across multiple VLMs, observation-aligned supervision generally improves observable-value recovery in both-executable evaluations and mostly improves end-to-end recovery, while the contour study reveals a trade off between observation alignment and representational compactness.

Detecting hallucination risk before generation enables abstention, retrieval augmentation, and routing decisions without incurring the cost of decoding. While prior work has shown that such risk can be estimated from a model's internal representations, existing approaches treat this as binary classification over a single decoded output. We instead formulate it as a risk-estimation problem. Under this formulation, we introduce soft-target supervision based on the empirical answer error rate over stochastically sampled outputs - an estimator we prove to be the unique unbiased minimum-variance estimator of the model's per-prompt error probability under its sampling distribution. We further adapt attention probing to the pre-generation setting, enabling the detector to selectively aggregate hallucination-relevant prompt representations. Across three question-answering benchmarks and five models, attention probing outperforms linear probing on short-answer tasks. Replacing binary labels with soft-target supervision further and consistently improves detection quality.

Large vision-language models contain several priors about the world and object interactions, making them useful critics during inference to steer robot policies towards success. However, closed-loop robot manipulation requires judging small visual differences between success and failure, which remains a challenge for current VLMs. We introduce a method to fine-tune critics by constructing pairwise progress supervision using success and failure rollouts obtained from a policy. Our fine-tuned critic excels at fine-grained progress reasoning and subtle failure detection, outperforming prior progress reasoning baselines. Additionally, we use an action-conditioned video model to predict the visual effect of several candidate actions sampled from a policy, and show that our critic can correctly identify successful candidates to execute, improving the average policy success rate by 11% across real-world tasks and 5.9% across simulation tasks.

Emotion annotation is inherently subjective, yet most NLP pipelines still assume "gold" labels, typically produced by majority voting, and treat annotator variation as noise. In this paper, we present a multilabel emotion annotation case study and use it to examine how annotator behavior and aggregation choices affect both agreement estimates and downstream emotion classifiers. Rather than collapsing disagreement into a single label, we represent targets as soft vote-share labels (including an intensity-weighted variant) and evaluate models using both thresholded metrics (macro-/micro-F1) and probabilistic alignment (Bernoulli cross-entropy SoftBCE), alongside data-derived disagreement diagnostics. Across annotation regimes, we show that disagreement is structured and leaves measurable traces in model behavior: hard labels may maximize F1 metrics, while soft supervision yields predictions that better reflect empirical annotator variance and uncertainty. Our results provide practical guidance for designing, aggregating, and evaluating multilabel emotion datasets when multiple interpretations are plausible.

Early failure alerting requires deciding, while a dialog or agent trajectory is still unfolding, whether to flag it as likely to fail. This is challenging because supervision is typically available only as a trajectory-level success/failure label while alerts must be raised from partial interactions. Prior early-classification methods often bridge this gap by assigning the terminal label to every prefix, treating every turn as failure evidence. We hypothesize that this prefix-label assumption is poorly matched to multi-turn language interactions, where evidence of eventual failure is sparse and often delayed. In this paper, we introduce a two-stage approach that learns from this sparse evidence structure and uses the resulting risk estimates for controllable early alerting. Specifically, our attention-based failure predictor learns sparse turn-level failure evidence from trajectory labels and uses it to estimate failure risk from partial histories. We then pair this predictor with $α$-STOP, a single preference-conditioned stopping policy that selects an accuracy-earliness operating point at inference time rather than training a separate trigger for each preference. Across five benchmarks spanning customer support, task-oriented dialog, persuasion, tool use, and planning, we first show that high-relevance failure evidence occupies only 4.7-11.3% of turns and first appears after 59.0-83.6% of trajectories on average. We further show that the attention-based predictor improves Pareto-frontier quality (hypervolume) by 1-10% over naive prefix supervision, and that the full system improves frontier quality by 3-42% over state-of-the-art trigger policies while reducing training cost per operating point by 1-3 orders of magnitude.

Reinforcement learning with verifiable rewards (RLVR) has greatly advanced large reasoning models (LRMs), but it requires timely training on a huge fully-annotated dataset. To this end, data-efficient RLVR methods have been widely studied from two perspectives: (i) data selection methods identify a small subset of "golden" samples that yield near-full-data performance, but they rely on a pre-existing pool of labeled data. (ii) unsupervised RLVR methods train the model using its own internal supervision signals on large-scale unlabeled data, yet they exhibit suboptimal performance. Accordingly, we investigate the "pick in the dark" setup for RLVR, which aims to select, without prior supervision, unlabeled samples that are most beneficial for training and worthy of annotation. Through systematic analysis, we demonstrate that smart picks hinge on a well-calibrated uncertainty estimator to enable strategic partitioning of data for adaptive training regimes. Building on this insight, we propose PivotTrace, a three-way data triage framework that leverages attention dynamics to trace metacognitive pivots during reasoning. By precisely quantifying uncertainty through pivot density, PivotTrace achieves automated data routing to synergistically maximize both annotation and training efficiency. Empirically, PivotTrace surpasses the fully supervised LRM with only 29.3% annotated samples and 2.75 faster convergence.

Safety detection models require examples of HHH (Helpful, Harmless, Honest)-violating outputs for robust generalization, however such examples are scarce. Activation Steering (AS) has emerged as a data-efficient method for generating target-concept-aligned responses. We investigate whether AS can generate high-quality training datasets for downstream classifiers, a question that remains untested. We present a two-fold study with intrinsic and extrinsic evaluation across $4$ concepts $\times\,2$ models $\times\,4$ steering methods. Intrinsically, beyond the field-standard rubric of steering success (concept alignment) and coherence, we introduce sample- and set-level diversity as a quality axis previously absent from the literature, and find that increasing steering strength reduces response diversity. Extrinsically, we replace HHH-violating examples in the available training data with steered generations and fine-tune detection classifiers. AS-generated data results in a better classifier than the prompting-generated data on $3$ of $4$ concepts. However, only $41$ of $136$ AS configurations outperform prompting, indicating that downstream utility lies in a narrow regime that jointly satisfies success, coherence, and diversity. The harmonic mean of these three axes correlates with downstream AUROC more consistently across concepts than success and coherence alone, providing a practical heuristic target for practitioners tuning AS hyperparameters. Together, our results highlight the potential of AS in synthetic data generation for improving safety detection and identify diversity as a critical, previously overlooked axis for tuning AS.

Large Language Models (LLMs) have achieved remarkable advancements in reasoning capabilities empowered by Reinforcement Learning with Verifiable Rewards (RLVR). Nonetheless, RLVR intrinsically relies on ground-truth labels for reward computation, the acquisition of which is often prohibitively expensive in real-world scenarios. While unsupervised RLVR paradigms attempt to circumvent this by training on pseudo-labels, they are notoriously susceptible to training collapse. Moreover, different samples often exhibit varying annotation values. In this paper, we propose Reinforcement Learning with Active Verifiable Rewards (RLAVR), which actively acquires ground-truth labels for a small set of selected samples and integrates them with pseudo-labels, thereby stabilizing training dynamics and improving performance under limited annotation budgets. To identify valuable samples, we propose the Corrective Advantage Gap (CAG) metric and analyze the sample-level supervision value. Building on this, we introduce Correction-Aware Reliability Estimation for RLAVR (CARE), which translates the oracle CAG criterion into a practical pre-query acquisition policy to substantially improve training stability. Extensive experiments across diverse domains, model families, and model scales demonstrate the effectiveness and generality of our approach. Our code is available at https://github.com/Lumina04/CARE.

Classical noisy-label theory predicts that downstream performance under weak supervision is bounded above by the labeler's accuracy, implying a sharp crossover: once a gold-trained classifier matches the labeler, weak labels stop helping and start hurting. The prediction is theoretical; what is missing is a benchmark calibration that turns it into an instance-level statement for modern foundation-model labelers. We provide such a calibration for BiomedCLIP-generated weak labels on three medical-imaging benchmarks (PCAM, ISIC, NIH-CXR) and six downstream architectures spanning an 11x parameter range. The crossover predicted by theory appears at ng~100 on PCAM, 20-50 on ISIC, and 250-500 on NIH-CXR; weak labels above the crossover degrade AUC by up to -0.10. The location is architecture-invariant for four of five pretrained architectures, and a within-family DenseNet sweep (2.5x parameters, identical pretraining) supports the view that the labeler, not the student, is the dominant constraint. The calibration in turn produces a decision rule operable from 10-20 gold labels: compare gold-only AUC to VLM accuracy on the user's gold set. A structured-vs-random noise sign flip on NIH-CXR shows that the rate-only formulation of the bound is incomplete and identifies a concrete refinement (label-space projection) that future benchmarks can be designed to test.

Foundation video models produce visually impressive results, but their use in embodied AI remains limited because they are primarily trained on natural language rather than low-level control signals. This limitation is especially pronounced for aerial flight, where motion occurs in unconstrained 6-DoF space and small errors in ego-motion can produce large trajectory drift. Generating aerial videos that follow fine-grained inertial actions can support scalable training and evaluation of aerial agents by providing a controllable proxy for real-world or expensive simulation data. To address this problem, we propose \textbf{Aero-World}, a method for converting a pretrained image-to-video diffusion model into a controllable aerial video generator. Aero-World injects sequences of translational acceleration and angular velocity into a pretrained latent diffusion transformer through an action-token stream. A frozen latent-space Physics Probe, trained independently on real video--IMU pairs, provides differentiable inertial-consistency supervision during LoRA finetuning while avoiding computationally expensive video decoding. We further propose \textbf{AeroBench}, a benchmark for evaluating whether generated drone videos adhere to low-level action signals. AeroBench uses Action Alignment Score (AAS) to measure agreement with commanded inertial actions and Physical Consistency Rate (PCR) to measure temporal motion stability. On AeroBench, Aero-World improves mean AAS from 57.7 to 63.6 over action-only finetuning and gives a stronger quality-control trade-off than AirScape, with lower FVD (596.5 vs. 1058.6), higher SSIM (0.595 vs. 0.505), and higher Flow-IMU correlation (0.44 vs. 0.20). These results suggest that frozen Physics Probe supervision is a practical mechanism for adapting pretrained video generators toward more action-aligned aerial motion.

The evolution of Large Language Model (LLM) reasoning is bottlenecked by the scarcity of high-quality process data. While self-alignment via endogenous rewards offers a solution, mining valid supervision faces three challenges: (1) Label Noise via Mimetic Bias, where rewards prioritize statistical likelihood over logical truth, creating a "correctness illusion" that masks compounding errors; (2) Coarse-Grained Supervision, where sparse global outcomes (e.g., in GRPO) fail to provide granular guidance, treating reasoning chains as monolithic; and (3) Distributional Collapse, where signals fail to generalize without amplifying pre-training biases. To address these, we introduce LC-ERD (Logic-Consistent Endogenous Reward Decomposition), a framework framing self-alignment as latent structure mining. We derive a Variational Logic Potential by aggregating consensus from the model's Latent Logic Expertise (LLE) to denoise the reasoning manifold, and introduce a Multi-Agent Value Decomposition protocol based on the IGM principle to quantify individual step utility. Experiments show LC-ERD delivers a robust self-evolution path, uncovering trade-offs between logic consistency and accuracy while identifying high-value reasoning patterns missed by standard rewards. Our code is available at https://github.com/LC-ERD-repo/LC-ERD.

Weakly labeled datasets such as AudioSet have driven recent progress in audio tagging. However, annotation quality varies across sound classes. Labels may be incomplete, ambiguous, or unreliable, which introduces class-dependent supervision bias during optimisation. The issue becomes harder as real and generated audio are increasingly mixed in training, and generated samples do not always match their intended semantic labels. Prior work mainly addressed unreliable supervision from missing-positive labels, while this paper targets three other sources of unreliable supervision: spurious additions, misassignments between similar classes, and weakened label evidence. These effects introduce class-dependent optimisation bias that is not explicitly modeled by most existing methods. To bridge this gap, the paper proposes a Class-wise Supervision Unreliability (CSU) framework that controls supervision strength at the class level during training. CSU learns a separate unreliability parameter for each class and down-weights less reliable supervision without changing the model architecture or inference process. To support evaluations, this paper also introduces ESC-FreeGen50, a manually verified benchmark of 50 sound classes that combines real and generated audio. Experiments on controlled benchmarks and AudioSet show that CSU improves robustness across different architectures and different sources of supervision unreliability. The results indicate that explicit class-wise modeling of supervision unreliability is an effective and practical strategy for robust audio tagging under large-scale weakly labeled training. Code and data are available at: https://github.com/Yuanbo2020/CSU

Reinforcement Learning with Verifiable Rewards (RLVR) is widely used to improve reasoning in large language models, but rewards only final-answer correctness with no supervision over intermediate steps. Rubric-based methods such as Rubrics as Rewards (RaR) introduce finer-grained supervision by scoring rollouts against structured criteria, yet the rubric scores are still aggregated into a single scalar applied to the entire response, causing three weaknesses: loss of multi-criterion structure, uniform supervision of correct and incorrect steps, and reward hacking through unbounded self-correction. On 1,000 problems, we find 18.2% of steps in correct-answer responses are wrong yet positively rewarded, while 49.9% of steps in incorrect-answer responses are correct yet penalized. We introduce Step-wise Rubrics as Rewards (SRaR), an RLVR framework that (i) uses an LLM judge to attribute each rubric item to a specific reasoning step, (ii) normalizes per-step rubric scores across rollouts so only steps whose quality varies produce a learning signal, and (iii) combines the per-step reward with the outcome reward through a decoupled advantage estimator that keeps the outcome baseline stable. We further build a 16K-problem rubric dataset by contrastively distilling rubric items from correct and flawed reasoning paths sampled from a strong model. Across six mathematical reasoning benchmarks, SRaR improves average accuracy over RaR by 3.57 points on Qwen3-8B and 2.75 points on Qwen3-32B, raises the Faithful Reasoning Rate on AIME 2025 from 34.5% to 46.7%, and reduces self-correction looping from 48.1% to 26.5%.

In this report, we present HoloMotion-1, a humanoid motion foundation model for zero-shot whole-body motion tracking. A key innovation of HoloMotion-1 is to scale control-policy training with a large-scale hybrid motion corpus, where video-reconstructed motions from in-the-wild videos provide the dominant source of motion diversity, while curated motion-capture and in-house motion data provide higher-fidelity supervision and deployment-oriented coverage. This data regime enables HoloMotion-1 to move beyond conventional MoCap-only training and exposes the policy to substantially broader behaviors, capture conditions, and motion styles. Learning from such heterogeneous data introduces new challenges, including reconstruction noise, source-domain mismatch, uneven motion quality, and the need for temporal modeling under large behavioral variation. To address these challenges, HoloMotion-1 integrates large-capacity temporal modeling, a sparsely activated Mixture-of-Experts Transformer with KV-cache inference for real-time control, and a sequence-level training strategy that improves learning efficiency on extended motion sequences. Extensive experiments on multiple unseen motion benchmarks show that HoloMotion-1 generalizes robustly across diverse motion types and capture conditions, significantly improves tracking accuracy over prior methods, and transfers directly to a real humanoid robot without task-specific fine-tuning.

Exemplar-based image editing applies a transformation defined by a source-target image pair to a new query image. Existing methods rely on a pair-of-pairs supervision paradigm, requiring two image pairs sharing the same edit semantics to learn the target transformation. This constraint makes training data difficult to curate at scale and limits generalization across diverse edit types. We propose Delta-Adapter, a method that learns transferable editing semantics under single-pair supervision, requiring no textual guidance. Rather than directly exposing the exemplar pair to the model, we leverage a pre-trained vision encoder to extract a semantic delta that encodes the visual transformation between the two images. This semantic delta is injected into a pre-trained image editing model via a Perceiver-based adapter. Since the target image is never directly visible to the model, it can serve as the prediction target, enabling single-pair supervision without requiring additional exemplar pairs. This formulation allows us to leverage existing large-scale editing datasets for training. To further promote faithful transformation transfer, we introduce a semantic delta consistency loss that aligns the semantic change of the generated output with the ground-truth semantic delta extracted from the exemplar pair. Extensive experiments demonstrate that Delta-Adapter consistently improves both editing accuracy and content consistency over four strong baselines on seen editing tasks, while also generalizing more effectively to unseen editing tasks. Code will be available at https://delta-adapter.github.io.

Label Distribution Learning (LDL) models supervision as an instance-wise probability distribution, enabling fine-grained learning under inherent ambiguity, but its success relies on high-fidelity label distributions that are costly to obtain and thus often noisy. Motivated by privacy-sensitive applications, we study Federated Label Distribution Learning (Fed-LDL), where data isolation further induces heterogeneous annotation quality across clients, making local updates unevenly reliable and breaking sample-size-based aggregation (e.g., FedAvg). To address this trust dilemma, we propose FedQual, a quality-aware Fed-LDL framework with two coupled mechanisms: (i) quality-adaptive client training guided by a global semantic anchor that calibrates low-quality clients while preserving high-quality autonomy, and (ii) reliability-aware server aggregation that reweights client contributions by effective reliable information rather than raw sample size. To enable rigorous evaluation, we construct four new Fed-LDL benchmarks (FER-LDL, FI-LDL, PIPAL-LDL, and KADID-LDL) with controlled annotation quality disparity. We further provide a theoretical guarantee showing that under heterogeneous supervision quality, client-specific calibration is strictly better than any uniform calibration. Extensive experiments on the proposed benchmarks demonstrate the effectiveness of FedQual.

Weak-to-strong alignment offers a promising route to scalable supervision, but it can fail when a strong model becomes confidently wrong on examples that lie in the weak teacher's blind spots. Understanding such failures requires going beyond aggregate accuracy, since weak-to-strong errors depend not only on whether the strong model disagrees with its teacher, but also on how confidence and uncertainty are distributed across examples. In this work, we analyze weak-to-strong alignment through a bias-variance-covariance lens that connects misfit theory to practical post-training pipelines. We derive a misfit-based upper bound on weak-to-strong population risk and study its empirical components using continuous confidence scores. We evaluate four weak-to-strong pipelines spanning supervised fine-tuning (SFT), reinforcement learning from human feedback (RLHF), and reinforcement learning from AI feedback (RLAIF) on the PKU-SafeRLHF and HH-RLHF datasets. Using a blind-spot deception metric that isolates cases where the strong model is confidently wrong while the weak model is uncertain, we find that strong-model variance is the strongest empirical predictor of deception across our settings. Covariance provides additional but weaker information, indicating that weak-strong dependence matters, but does not by itself explain the observed failures. These results suggest that strong-model variance can serve as an early-warning signal for weak-to-strong deception, while blind-spot evaluation helps distinguish whether failures are inherited from weak supervision or arise in regions of weak-model uncertainty.

Recent works proposed test-time alignment methods that rely on a small aligned model as a proxy that guides the generation of a larger base (unaligned) model. The implicit reward approach skews the large model distribution, whereas the nudging approach defers the generation of the next token to the small aligned model when the large base one is unconfident about its outcome. In this work, we first show that both approaches can be reduced to sampling from similar graphical models, where they differ only in the definition of a rejection criterion (or distribution). Moreover, we argue that the confidence criterion is ill-motivated due to linguistic phenomena like ambiguous phrasing. We propose a novel rejection criterion based on a conservative confidence bet. Experimentally, our novel approach outperforms previous work on several datasets.

Temporal action segmentation (TAS) in untrimmed videos requires dense temporal supervision. However, most of the annotation cost is spent identifying action transitions where segmentation errors concentrate and small temporal shifts can disproportionately degrade segment-level metrics. We introduce B-ACT, a clip-budgeted active learning framework that explicitly allocates supervision to these error-prone boundary regions. B-ACT operates in a hierarchical two-stage loop: (i) it ranks and queries unlabeled videos using predictive uncertainty, and (ii) within each selected video, it detects candidate transitions from the current model predictions and selects the top-$K$ boundaries via a novel boundary score. The boundary score fuses neighborhood uncertainty, class ambiguity, and temporal prediction dynamics to reveal the underlying importance of each frame. Importantly, our annotation protocol requests labels only at the boundary frames while still training on boundary-centered clips to exploit temporal context through the model's receptive field. Extensive experiments on GTEA, 50Salads, and Breakfast demonstrate that boundary-centric supervision delivers strong label efficiency and consistently surpasses representative TAS active learning baselines and prior state of the art under sparse budgets. Gains are largest on datasets where performance is highly sensitive to boundary placement, as measured by edit and overlap-based F1 metrics.

Reliable models should not only predict correctly, but also justify decisions with acceptable evidence. Yet conventional supervised learning typically provides only class-level labels, allowing models to achieve high accuracy through shortcut correlations rather than the intended evidence. Human priors can help constrain such behavior, but aligning models to these priors remains challenging because learned representations often diverge from human perception. To address this challenge, we propose an attribution-based human prior alignment method. We encode human priors as input regions that the model is expected to rely on (e.g., bounding boxes), and leverage a highly faithful subset-selection-based attribution approach to expose the model's decision evidence during training. When the attribution region deviates substantially from the prior regions, we penalize reliance on off-prior evidence, encouraging the model to shift its attribution toward the intended regions. This is achieved through a training objective that imposes attribution constraints induced by the human prior. We validate our method on both image classification and click decision tasks in MLLM-based GUI agent models. Across conventional classification and autoregressive generation settings, human prior alignment consistently improves task accuracy while also enhancing the model's decision reasonability.