Vision-language classifiers can learn new labels from text alone, so the label lists they serve keep growing. We show that this growth silently changes conformal decisions about old labels. Common softmax-based conformal scores compare each label with an average over all labels; adding labels changes that average, and with it the score of every old label, even when the model, the image, and the old calibration data stay the same. Recalibrating after the update keeps the coverage guarantee, yet across eleven settings it changes at least one old decision on 53-90% of new images. Coverage cannot reveal this, because it checks only the true label: in five controlled settings that keep the old calibration, so that only the normalizer changes, old-class coverage moves by at most 0.01 while 60-67% of images change. We propose a lifecycle audit that measures these changes alongside coverage and set size, and a simple fix, PrefixCP: each class keeps the label registry that existed when it was added. Its old decisions then never change, each class keeps exact class-conditional coverage under exchangeability, and one prefix scan computes all historical normalizers at once. PrefixCP recovers 40.9-58.5% of the set-size gap between never updating the reference and recalibrating, which gives smaller sets but changes old decisions.
Large Reasoning Models (LRMs) have achieved impressive performance via explicit Chain-of-Thought (CoT) reasoning, yet this process introduces critical safety risks. We formalize LRM safety objectives as requiring non-malicious reasoning and answers, safety consistency, and explicit intent identification. We argue that existing alignment methods suffer from poor adversarial generalization, a significant ``safety tax'' on reasoning, and unfaithful coarse-grained rewards, primarily due to their neglect of the LRM's intrinsic reasoning capabilities for safety. To address these, we propose \textbf{SAVeR$^2$}, a novel \textbf{S}afety \textbf{A}lignment method via \textbf{R}ewards with \textbf{R}easoning capability. SAVeR$^2$ decomposes the safety objective into four fine-grained verifiable reward signals and utilizes a difficulty-aware data selection strategy to stabilize training. Crucially, we introduce a reasoning-preserving gradient projection mechanism that analytically resolves conflicts between safety and reasoning gradients, effectively eliminating the safety tax. Extensive experiments on models ranging from 1.5B to 671B demonstrate that SAVeR$^2$ achieves superior safety performance and significantly lower over-refusal rates across multiple benchmarks without compromising general reasoning capabilities.
In embedding visualizations, nearby instances are interpreted as semantically related, although neighborhoods derived from input embeddings may not reflect content relationships. Neighborhood-based dimensionality reduction (DR) methods can propagate this limitation by optimizing layouts from embedding-based neighborhood graphs. We propose Knowledge-Guided Neighborhood Graph Construction (KGN), which represents relational knowledge derived from each instance as a set of knowledge graph (KG) triples. It uses the degree of triple sharing as pairwise semantic evidence to adjust embedding distances and reconstruct the top-k neighborhood graph; neither KG extraction nor neighborhood construction uses class labels. We also introduce IDProfile (IDP), which interprets residual mismatches by independently assessing whether the source content supports the assigned class label and the neighborhood-majority class. Across eight text and image dataset-encoder conditions, we characterize KGN over the full guidance-strength sweep and four KG similarity measures. Across this sweep, the maximum NMI and LT_2D each exceed the corresponding kNN baseline in all eight conditions, with the largest gains reaching 0.68 and 1.35, respectively. The gains and the extent of neighborhood reconfiguration vary across conditions and are larger under encoder conditions with lower kNN baseline scores. The IDP analysis shows that the largest reductions are concentrated in Placement Issue candidates, whereas Label Issue and Unresolved Case instances account for a larger share of the remaining review set. These results indicate that KGN can narrow the evaluated gap between embedding-based neighborhoods and content-supported semantic relations, while IDP shows that a residual mismatch should not automatically be interpreted as a projection failure.
Reinforcement learning has achieved strong performance in many tasks by letting agents learn from nothing more than a reward signal. However, in more complex tasks, agents still struggle to learn good, generalizable policies, even with carefully designed rewards and abundant samples. We argue that this difficulty stems from the reward decoding burden: designers encode their knowledge of the desired behavior into rewards and penalties, and agents must infer this knowledge from long sequences of noisy scalars. To address this burden, we propose prescriptive policy learning (PPL), which directly teaches agents higher-level knowledge about the tasks, termed principles of behavior, such as how to maintain balance, how to accumulate energy, and how to land smoothly. Through appropriately selected key quantities and references, PPL can directly indicate at every step whether the agent follows the principles and to what extent. Our experiments show that agents can learn good, generalizable policies without any reward. Taught only simple heuristic principles, PPL agents outperform reward-based methods with far fewer samples. Their policies also transfer directly to harder settings and keep performing well, demonstrating their strong generalization ability. In summary, PPL provides a new approach in which rough knowledge of the desired behavior is enough for agents to learn good, generalizable policies, even when the system dynamics cannot be accurately modeled.
Deep search and long-horizon reasoning unfold as branching trajectories that pursue parallel sub-queries, abandon dead-end detours, and cross-check sibling findings. Yet modern serving engines, harnesses, and RL trainers execute episodes as flat, append-only sequences: stretching a context window leaves failed detours that dilute attention, in-place edits break KV caching and recurrent states, and all-layer cache splicing overwrites hybrid recurrent working memory and triggers hijacking. We introduce Octopus, a co-designed serving runtime, memory manager, and RL framework executing episodes as a GPU-resident DAG over radix cursors (89.6% KV hit rate). Agents coordinate via five primitives: fork shares parent KV prefixes in $O(1)$ with isolated 32K budgets, rewind restores recurrent checkpoints, read_thread mounts sibling KV pages in zero copy, and wait and done synchronize results. Octopus splices sibling KV pages only into full-attention layers under monotonic position offsets while recurrent layers skip mounted spans, retaining 92.2% of full-text answer information and cutting hijacking $21\times$ ($57.9\% \to 2.7\%$) with zero new KV pages. During RL, our packed DAG kernel and DAG-GRPO objective deduplicate shared prefixes via loss masks while keeping mounted KV edges differentiable, cutting rollout-training KL divergence $32\times$ and routing reader gradients into producer KV projections. Trained on Qwen3.5-4B with binary rewards, Octopus self-organizes a commander-worker hierarchy and latent KV mesh, lifting BrowseComp-Plus accuracy under BM25 from 26.0% (32K ReAct) and 32.1% (143K GRPO) to 49.6% (+17.5 pts; 51.2% on OpenResearcher), surpassing 9B context-editing RL (42.5%), matching o3 (49.3%), and outperforming dense-retrieval agents up to $250\times$ larger: Tongyi-DeepResearch (44.5%), DeepSeek-R1 (16.4%), Kimi-K2 (35.4%), GPT-4.1 (36.4%), and Claude 4 Opus (36.8%).
Vision-language-action models are typically built on a simple premise: perception is first compressed by visual encoders, and action reasoning begins only afterwards. Yet manipulation often hinges on precisely the visual signals most vulnerable to this abstraction, e.g., local geometry, boundaries, depth, and subtle state changes. This raises a fundamental question: should action policies inherit a fixed visual endpoint, or directly access evolving visual computation? We introduce NEO-VLA, a structurally encoder-free vision-language-action framework. Images are tokenized by lightweight patch convolutional layers and evolve jointly with language throughout the backbone, allowing a flow-matching action expert to draw from instruction-conditioned visual representations at different depths. This simple change reveals a consistent pattern. Native architectures outperform encoder-based counterparts under controlled comparisons; full fine-tuning reduces but does not erase the gap, and additional scale alone cannot recover it. More strikingly, layer-wise probing shows that the action-critical signals often peak well before the final visual representation. Across diverse manipulation settings and distribution shifts, NEO-VLA consequently achieves stronger and more robust control. Together, these results suggest that embodied control benefits from access to visual representations across the computational hierarchy, rather than a fixed encoder output.
Sign language generation (SLG) maps written sentences into signing motion while preserving both lexical content and fine-grained articulation. This requires a model to decide what signs should be expressed and to realize them through precise handshape, trajectory, facial expression and body coordination. Existing pose-based systems and reconstruction-only motion tokenizers often entangle these two decisions, leaving the generator without an explicit text-aligned intermediate target for what to sign before refining how to sign it. To address this issue, we introduce Semantic Residual Vector Quantization (SRVQ), a framework for SLG that frames generation as a two-level process moving from semantics to kinematics. The newly introduced SRV Quantizer learns a hierarchy in which the first residual level encodes sign content and subsequent levels refine pose, handshape, trajectory, and body coordination, all within a single quantizer and without separate semantic representations or external planning stages. Through word-aware conditioning for hierarchical code generation, a masked semantic generator predicts semantic codes and a kinematic generator completes the residual codes for full SMPL-X motion synthesis. We evaluate using retrieval and FID in a shared text-motion embedding space, alongside standard geometric metrics. On CSL-Daily and Phoenix-2014T, SRVQ achieves R@1 of 82.3 and 70.3, outperforming the previous SOTA by 30 and 46 points, respectively.
EEG extended-source imaging aims to localize active cortical regions and characterize their spatial extent, yet supervised inverse networks typically depend on synthetic EEG-source pairs and may be sensitive to synthetic-to-real mismatch. We propose a physics-informed self-supervised implicit neural representation (INR) framework for spatial source reconstruction without paired source labels or simulation-based supervised pretraining. The lead field matrix encodes the source-to-sensor volume-conduction model and constrains the INR through agreement between predicted and measured EEG. Cortical current density is modeled as a continuous coordinate-conditioned field. A classical inverse estimate provides warm initialization and spatial constraints on source activity, while patch-integrated projection maps the field to discrete source coefficients for forward modeling. Simulations across varying noise levels, source extents, and source counts showed consistently low spatial dispersion and good localization performance. Quantitative and qualitative evaluations using real EEG recordings from epilepsy patients and intracranial stimulation experiments further supported the feasibility of the framework for source imaging. These findings suggest the potential of physics-informed self-supervised INR for extended-source imaging.
Tool-using large language model (LLM) agents translate model outputs into consequential actions, turning indirect prompt injection into an execution-security problem. Existing defenses constrain tool calls, trajectories, or information flows, but often leave action semantics, state dependencies, and evidence relations implicit, limiting generalization across heterogeneous tools and complicating safe updates. We present SecureGARD, a dynamic ontology-guided governance framework. SecureGARD grounds heterogeneous calls in a shared semantic representation and combines an executable ontology with episode-level task state to derive deterministic ready, waiting, and forbidden decisions. A waiting decision identifies an unmet but recoverable obligation, enabling bounded replanning without expanding the agent's authorization boundary. SecureGARD evolves its ontology offline from verified training executions. An offline LLM proposes instance-independent action-state axioms, which are accepted only if they address ontology errors exposed by training failures, preserve logical consistency and protected entailments, and pass held-out validation before the ontology is frozen. We evaluate SecureGARD on AgentDojo, ASB, and AgentDyn under benign and indirect prompt-injection settings. SecureGARD achieves a 0% attack success rate (ASR) across all three benchmarks while improving task utility under attack compared with existing baselines. These results show that explicit semantic governance can strengthen execution security without indiscriminate blocking.
Temporal point processes (TPPs) are the canonical model class for event sequences in continuous time, with applications ranging from seismology to electronic health records. Recent benchmarks report that neural TPPs substantially outperform their classical counterparts, such as Hawkes processes. However, these comparisons set flexible neural models against Hawkes processes in their most restricted forms and covered only a narrow range of domains. We revisit this comparison under autoregressive next-event prediction, evaluating Poisson and Hawkes processes of increasing flexibility against neural TPPs on five established benchmarks and two large-scale medical tasks derived from UK Biobank disease histories and MIMIC-IV ICU records. First, once Hawkes processes are equipped with flexible, learnable baselines and kernels, the previously reported gap to neural TPPs largely closes. Second, neural TPPs are likewise limited by their intensity parametrization: a Transformer with a spline-based intensity head outperforms previous intensity-based neural TPPs and matches or exceeds the intensity-free model. Decomposing the log-likelihood into time and type terms shows that this gain lies almost entirely in modeling when events occur (log p(t)) rather than which type occurs (log p(d|t)). Third, only on MIMIC-IV do neural TPPs retain an advantage in predicting event types, underscoring the value of complex datasets for evaluating TPPs. We release TorchPP, an open-source PyTorch framework with unified implementations of classical and neural TPPs.
Unified Multimodal Models (UMMs) aim to support visual understanding and generation within a single model. Existing post-training methods either use understanding as a fixed teacher to improve generation, or rely on ground-truth trajectories to supervise both understanding and generation. Prism instead derives bidirectional supervision from self-generated interleaved rollouts, constructing training targets according to the reliability of their text and image chains. We propose Prism, a bidirectional self-teaching post-training method for UMMs. Like a prism decomposing mixed light, Prism decouples a coupled interleaved trajectory into a text chain and an image chain, assesses the reliability of each chain, and constructs a corresponding training target from their reliability states. Experiments on both SenseNova-U1 and BAGEL show that Prism improves both understanding and generation. On the generation side, GenEval improves from 85.8 to 89.2 (+3.4), and DPG improves from 86.1 to 87.3 (+1.2) on SenseNova-U1. On the understanding side, the total MME score improves from 2282.6 to 2311.6, and MMVP improves from 58.0 to 62.0. Consistent improvements are observed on BAGEL. These results highlight that Prism enables coordinated improvements in generation and understanding by constructing reliability-guided self-training targets from self-generated trajectories, without relying on an external verifier or externally annotated corrected trajectories during target construction. Code is available at https://github.com/aliyun/qwen-dianjin/tree/master/DianJin-Prism.
Enzyme engineering usually optimises which candidates to assay within a fixed model and workflow; when the bottleneck is a mismatched method or a failing workflow, that route is hard to escape. We present EnzymeRSI, a dual-loop system for recursive self-improvement (RSI). The outer loop evolves harness assets without changing model weights; the inner loop develops prediction and selection algorithms under controlled conditions; a cloud laboratory system is specified for closing the physical loop with quality-controlled feedback. The outer loop raised the mean development-set Spearman correlation on ProteinGym by 0.044 across five assays, reaching 0.335 on CD19 and 0.683 on OPSD (rhodopsin); of the two first places, only CD19 was produced by the outer loop itself. In ThermoMPNN self-improvement the inner loop reached a PCC of 0.849, and on TadA-Bench a Spearman correlation of 0.307 in 1.13 h. We examined the system's experimental decision-making in three enzyme systems. For the luciferase RLuc, 10 of the 32 nominated substitutions not pre-specified reached at least twice wild-type activity in a single round. For the cytidine deaminase A3A, a variant selected by a rule that caps loss of unmodified C raised 5mC conversion in the weaker CT context from 11.4% to 46.5% and reversed the substrate preference in both contexts. For the DNA polymerase KOD-E, 189 single-substitution variants defined a landscape on which different objective functions selected divergent next steps from the same parent; one assayed combination reached 31.1 times the parent on a modified substrate. By treating the organisation of research and the development of algorithms as separately scored objects of evolution, and linking a frozen decision record to a physical measurement, this work offers an evolutionary paradigm for moving beyond the blind spots of a fixed method and for autonomous enzyme engineering.
Conformal Prediction (CP) provides a model-agnostic framework for constructing finite-sample prediction sets with coverage guarantees and has become a tool for reliable decision-making in high-risk settings. Recent advances have extended Split Conformal Prediction to increasingly complex learning problems, including multilabel, high-dimensional, and functional outputs. In this work, we study CP for Distribution-to-Distribution Regression, a setting of growing interest across many application domains. We show that a naive application of Split CP yields prediction sets that, while marginally valid, are non-interpretable, e.g., lacking a well-defined volume or a tractable sampling procedure, and of limited practical value. To overcome this, we propose a new conformal framework that operates in an orthogonal basis space, producing interpretable prediction sets while preserving coverage guarantees. We further develop an adaptive variant with asymptotic conditional coverage under mild assumptions. Empirical results on synthetic and real-world datasets validate the effectiveness of the proposed approach. The code is publicly available at https://github.com/tuanhacs/conformal_d2d.
How many environments and samples does it take to separate $d$ Gaussian sources whose variances change across environments? Each source has a signature, its vector of variances. The Riesz $2$-energy $E_j$ of source $j$ sums the inverse Fisher informations for rotating it towards other sources. With known whitening, we prove a worst-case squared-angle risk lower bound of order $E_j/(n+E_j)$ for orthogonal estimators at every $n$ and $d$. A quadratic pencil given the signatures attains $E_j/n$ up to a constant above a separation-dependent sample threshold, even with estimated whitening. For well-spread signatures at fixed $K$, the energy grows as $d^2$, $d\log d$ and $d$ for $K=2$, $3$ and $K\ge4$, respectively. A design with $K\asymp\log d$ and pairwise Fisher information bounded below by a constant makes $d\log d$ samples sufficient. The eigenvectors of a single fixed linear combination have an average first-order error coefficient of at least order $d^2$. Simulations up to $d=1000$ support the first-order predictions, and on task fMRI the energy helps rank split-half instability.
Reliability management of industrial fleets depends on early fault warnings, yet lifecycle data across similar assets are highly imbalanced. A few assets accumulate run-to-failure histories, while newly installed or rarely failing assets provide only normal-condition data. This paper asks how diagnostic knowledge can be transferred from a data-rich source asset to a data-scarce target asset when no fault data have been observed on the target. We formulate cross-asset early fault detection as a zero-shot domain adaptation problem, in which zero-shot refers to the absence of target fault samples, and propose a framework that requires only normal data from the target. The framework comprises (1) a statistical process control step that labels the source asset run-to-failure data with an exponentially weighted moving average chart, reducing manual annotation to verifying flagged segments, (2) an Adapted Domain-Adversarial Neural Network (ADANN) that combines domain-adversarial training and a Maximum Mean Discrepancy penalty with an Adapted Classification-Domain (ACD) loss, which anchors the classifier normal region to the target own normal data, and (3) an ahead-of-time labeling scheme with a prediction horizon that targets incipient faults. Under an exact-classification assumption, a theoretical analysis characterizes the domain-adversarial term for a normal-only target and the alignment state in which source fault features remain separated from target normal features. On a public run-to-failure bearing dataset, ADANN attains the highest mean F1-score among four transfer baselines, with a nominally significant precision gain over the strongest baseline, and its robustness is examined under synthetic noise.
Confidence estimates guide human review in LLM-built knowledge graphs, but an erroneous shared entity need not be the only reason an incident fact is wrong. We study how this distinction affects budgeted audits, alongside selection headroom and label validity. Modelling non-adaptive auditing as budgeted covering, we characterise the minimal decision-relevant error statistics and bound the regret of scalar-preserving first-order planning by a structural mass $\Delta^{\mathrm{cl}}$. We simulate ideal repair against benchmark-aligned labels for five extraction configurations and two cross-document graphs on Re-DocRED, followed by frozen confirmation on test and replication on 500 hash-selected training documents. An error-rate $\times$ fan-out baseline leaves 3.98-4.74 percentage points more errors than joint planning with the same actions. Independent-block planning also avoids this disadvantage: its held-out point estimates improve on a fact-only scalar baseline by 1.22-1.79 points and are slightly better than the joint model, although the frozen equivalence criterion remains inconclusive. On 300 DWIE training documents, independent and joint plans also outperform fan-out in point estimates, with extractor-dependent gains over fact-only review. For verification, a generic support check improves three of four model-budget configurations on each held-out set. These comparisons motivate evaluating complete-repair probabilities and retaining both fact-only and factorised baselines when choosing an audit model.
Diffusion image generators owe much of their practical success to accurate denoising networks and strong numerical samplers. However, during sampling, the denoiser predicts the next step from flawed inputs that differ substantially from real data. This train-test gap, or exposure bias, caps image generation quality (i.e., blurry images, missing details, or distorted shapes). To mitigate the problem, we propose PEEK, a training-free deterministic sampler plugin. At each eligible denoising step, PEEK first performs a provisional one-step update, evaluates the denoiser once at the look-ahead state, and interpolates the current clean prediction with this one-step-later prediction to form the next noisy state. By taking the additional step, we obtain a generally more image like prediction from the higher-signal noisy state, mitigating exposure bias. To optimize the interpolation positions, PEEK searches for an interpolation schedule with a greedy metric-driven search made efficient by the PEEK prefix and baseline noise-schedule-rebasing caches. Across pixel-space and latent-space diffusion models with unconditional, label-guided, and text-guided settings on CIFAR-10, LSUNBedroom,I mageNet, and MS-COCO,the look-ahead correction consistently improves over baselines matched by number of function evaluations (NFE), with Fr\'echet Inception Distance (FID) reductions up to 49.3%. The one-time per-model calibration, the only extra cost, takes at most 2.2 minutes in our experiments. Project website is available at https://dachengruoque.github.io/PEEK-Video/.
For a stochastic program $\min_{z\in Z} c^T z + h(z)$ with a random cost vector $c$, regret, or the expected value of perfect information, is usually computed by re-solving the problem scenario by scenario. We show that regret is determined by the covariance between the cost vector and the optimal decision $\pi^s(c)$. The identity is exact and holds without regularity assumptions once the covariance is integrated along the path from the mean to each scenario. With $c_s=E[c]+s(c-E[c])$, $Reg=-\int_0^1 s^{-1} Cov(c_s,\pi^s(c_s))\,ds$. Three consequences follow. First, the plain covariance is always an upper bound: $0\le Reg\le-Cov(c,\pi^s(c))$. Second, when the optimal decision rule is affine, as in unconstrained and equality-constrained (e.g., budget-constrained Markowitz) quadratic programs, regret equals exactly one half of the covariance. For purely linear objectives, an affine rule forces zero regret. Third, the ratio of regret to covariance lies in $[0,1]$ and is governed by the shape of the covariance profile along the path. The noise level at which this profile first becomes positive is a decision-stability margin that tightens the bound. The identity also holds scenario by scenario for any predicted cost, so it applies to linear and combinatorial programs with discontinuous decision rules. We give dimension-free concentration bounds and a central limit theorem for the plug-in estimator. Across 7 problem classes, 8 cost distributions, and the shortest-path benchmark of Elmachtoub and Grigas (2022) with $10^6$ draws per configuration, the identities hold to within quadrature error, and the covariance bound holds in every scenario.
The value of memory lies not only in recall but in using it to provide appropriate assistance. Recent benchmarks test memory use in personalisation and agentic tasks, but seldom evaluate assistance without fixed answers or integration across multiple participants. Group chats make both challenges more acute: an assistant must integrate information from multiple members, track revisions and conflicts over time, and produce one response for the group. Yet existing group memory benchmarks rely on synthetic conversations and inherit constrained evaluation from general memory benchmarks, focusing on recall against fixed answers or decisions over a closed set. We introduce GroupAssistBench, the first memory benchmark from real-world group conversations with a deployed group assistant, covering 80,299 messages in 50 groups and costing over $53K. Over the same histories, GroupAssistBench evaluates memory recall (L1) with reference answers and memory use (L2) with service requests without fixed answers across single, distributed, and conflicting memories (M1--M3). Requests are verified to require group history, memories are anchored to their sources, and L2 uses tailored rubrics with penalties derived from real user feedback. Across evaluations of frontier models with full context, retrieval methods, and memory systems, the strongest memory system rivals full context on recall but falls far behind on use, especially for distributed and conflicting memories. Together, two additions that relate retrieved memories and bind remembered facts as response constraints bring consistent gains, showing that group memory systems must improve use, not recall alone.
Real-world time series exhibit complex dynamics across multiple temporal resolutions, yet effectively incorporating multi-scale information into generative modeling remains challenging without increasing generative complexity. We propose UniFlow, a unified framework that integrates multi-scale representation learning with efficient non-autoregressive generation through a shared-semantic discrete interface. Specifically, a multi-scale tokenizer employs shared-index voting to consolidate complementary temporal information into a single public-ID sequence, while scale-specific codebooks preserve the information required for reconstruction. Building on this unified representation, we develop geometry-aware categorical flow matching, which initializes transport from learned low-rank anchors and generates public identities non-autoregressively. This design captures multi-scale temporal structure without requiring multiple generative token streams or sequential token decoding. Experiments across four benchmarks and both short- and long-term settings demonstrate improved generative fidelity, strong cross-scale consistency, and substantial inference speedups over representative autoregressive and diffusion-based baselines. Code is available at https://github.com/kkking-kk/UniFlow