The prevailing paradigm in large language model (LLM) alignment operates via erasure, filtering unsafe data or training models to strictly refuse harmful prompts. While effective at reducing immediate toxicity, this approach fundamentally constricts the model's epistemological scope, resulting in over-cautious systems that output uninformative blanket refusals to sensitive yet benign queries. In this work, we challenge the orthodoxy that unsafe data must be discarded. We propose a dialectical approach to alignment, positing that unsafe data encodes rich, domain specific knowledge critical for nuanced, safe, and informative generation. To operationalize this, we introduce SafeMoE, a Mixture-of-Experts (MoE) framework that isolates unsafe knowledge into domain-specific Low-Rank Adapters (LoRA experts) trained exclusively on harmful corpora. To synthesize safety from these unsafe primitives, we train a lightweight gating network using a minimal, highly curated set of safe-informative responses. During inference, this router dynamically orchestrates the unsafe experts, effectively steering the generation trajectory to harness their deep domain knowledge while strictly enforcing safety constraints. Extensive empirical evaluations across stringent safety benchmarks demonstrate that SafeMoE is not only safer, achieving over a 20% relative improvement in safe response rate (more than a 15% absolute gain), but also produces more informative responses when safety and harmfulness are of paramount concern. Furthermore, the routing mechanism exhibits strong zero-shot generalization to unseen domains and broader safety tasks without domain-specific supervision. Our findings suggest a paradigm shift in alignment: true safety requires not the masking of unsafe knowledge, but its controlled integration.

Gene regulatory networks (GRNs) capture transcription factor-target interactions and are central to understanding cell-state regulation and disease. Reconstructing GRNs from paired single-cell transcriptomic and chromatin accessibility data is promising but challenging: scATAC is extremely sparse, and most methods rely on fixed peak-to-gene links and weak supervision. We present EpiAwareNet, a prior-guided multi-omic Transformer framework that reconstructs GRNs from paired single-cell data using only lightweight biological priors. In Stage 1, EpiAwareNet learns joint gene-peak representations with a gene-peak cross-attention module, enabling data-driven, gene-specific aggregation of accessibility signals rather than hard-coded peak-to-gene assignments. In Stage 2, EpiAwareNet incorporates a bulk-derived GRN prior as noisy positive edges to provide weak supervision under label scarcity, refining regulatory scores while remaining robust to prior noise. In our experiments, EpiAwareNet improves GRN reconstruction over representative single- and multi-omic baselines and yields GRNs with greater biological plausibility, such as improved recovery of known regulatory interactions, suggesting that lightweight biological priors from bulk data can effectively guide single-cell GRN inference when combined with adaptive cross-modal representation learning. Code and data will be available at https://github.com/tianyang-x/EpiAwareNet_pub.

Recent progress in the development of language models has been defined by scale, with each generation absorbing more of the world's knowledge into its weights. However, many practical applications benefit more from robust reasoning than from extensive parametric knowledge. In this setting, task-specialized small language models (SLMs) offer a principled design choice. We introduce Optimal Cognitive Core (OCC), a family of SLMs built around this premise. As a variant of OCC, we present OCC-RAG, optimized for faithful question answering (QA) grounded in the provided context. This task directly aligns with the OCC design approach, requiring multi-hop reasoning over supplied passages while ignoring memorized knowledge. To train OCC-RAG, we implement a novel pipeline for synthesizing multi-context, multi-hop QA data at scale, producing a corpus of over three million examples targeting multi-hop reasoning, strict context faithfulness, and calibrated abstention. We release OCC-RAG-0.6B and OCC-RAG-1.7B, both mid-trained on this corpus. The models produce structured reasoning traces with source citations grounded in literal quotes from the context. Through OCC-RAG, we demonstrate that compact, task-specialized SLMs can match or exceed general-purpose models 2 -- 6x their size across multi-hop reasoning (HotpotQA, MuSiQue, TAT-QA), faithfulness (ConFiQA), and refusal (MuSiQue-Un) benchmarks.

Fourier Neural Operators are often assumed to generalize across spatial resolutions, enabling training on a coarse grid and deployment on a finer grid. We test this assumption by contrasting two inference-time choices when moving from training resolution $s$ to test resolution $S>s$: running FNO directly at $S$, or running at $s$ and upsampling the prediction to $S$ via Fourier zero-padding. On Darcy flow, we observe that direct fine-grid inference is not reliably beneficial and can be worse than the low-grid-plus-upsampling baseline. We further analyze layerwise spectra and find that, under Fourier truncation, intermediate representations increasingly concentrate energy in low frequencies, with high-frequency output produced mainly by late nonlinear/decoder stages. This offers a mechanistic explanation for why FNO can perform well while retaining few modes, yet remain sensitive under resolution shifts. Our findings highlight a simple but strong baseline for cross-resolution evaluation and point to nonlinear aliasing as a key obstacle to zero-shot resolution equivariance.

This work presents a configurable pipeline for generating semantic-segmentation-ready agricultural datasets from Sentinel-2 imagery and EuroCrops parcel-level annotations. The workflow transforms heterogeneous vector crop annotations into aligned multispectral image--mask pairs through label harmonization, Sentinel-2 product selection, spatial alignment, rasterization, patch extraction, quality filtering, and class-aware sample selection. The generated dataset contains 67,337 patches from five European countries and uses a reduced taxonomy of ten crop classes plus background. A four-level U-Net with Group Normalization was trained using 10 Sentinel-2 spectral bands and a composite loss combining class-weighted cross-entropy and Dice loss. On the internal EuroCrops-based test split, the model achieved a mean Intersection over Union (mIoU) of 0.7665, a pixel accuracy of 0.8693, and a mean class accuracy of 0.9072. Compared with spectral and spatial-context Random Forest baselines, the U-Net showed the importance of learned multi-scale spatial representations for crop segmentation. External evaluation was performed on unseen Belgian EuroCrops subsets, DACIA5, and PASTIS. The results show a clear performance gap under external and cross-dataset evaluation, especially for benchmarks with different taxonomies, annotation protocols, spatial coverage, or temporal organization. The model transfers more reliably to dominant and taxonomically aligned classes such as maize and wheat, while performance remains limited for several minority classes and for the adapted single-date PASTIS setting. These findings highlight both the potential and the limitations of using EuroCrops-derived supervision for Sentinel-2 crop segmentation under realistic domain shifts.

Large Language Models (LLMs) aligned via outcome-based Reinforcement Learning (RL) frequently exhibit a critical failure mode: they achieve high performance on in-distribution benchmarks while demonstrating brittle reasoning capabilities on out-of-distribution (OOD) tasks. We term this phenomenon Reward-Induced Manifold Collapse. We establish a theoretical framework bridging Structural Causal Models (SCM) and the Information Bottleneck (IB) principle to explain this paradox. We define reasoning as a high-complexity causal process and shortcut learning as the exploitation of low-complexity spurious correlations. Under the implicit inductive bias of Stochastic Gradient Descent (SGD), models optimized for outcome rewards are biased toward shortcut solutions whenever the training distribution allows for a ``Markovian Screening'' of the true causal mechanism. We derive a new generalization bound based on Semantic Coverage Measure ($η$) rather than sample size, showing why data scaling on homogeneous distributions may fail to correct reasoning flaws. We also show that Process Reward Models (PRMs) function as Topological Filters, enforcing step-wise mutual information constraints that render the low-complexity shortcut manifold inadmissible. These results provide a mathematical grounding for the role of process supervision beyond simple credit assignment.

Thermal imaging offers a powerful alternative to visible-spectrum vision under challenging conditions such as low illumination and adverse weather, yet foundational vision-language models like CLIP fail to align thermal images with textual descriptions due to a fundamental thermal perception gap. We identify three major challenges: the lack of captioned thermal datasets, the inability of standard LLMs to reason about thermal phenomena, and a key representational challenge in thermal imaging where global scene context and object-level heat signatures conflict when learned together in a single embedding space. To address these, we introduce IR-Cap, the first physics-aware thermal captioning pipeline and dataset providing complementary global and fine-grained thermal descriptions across three public benchmarks, and T-CLIP, a decoupled dual-LoRA framework that independently adapts CLIP for scene-level and object-level thermal understanding. T-CLIP achieves consistent improvements over all baselines across three thermal benchmarks in cross-modal retrieval, and we provide an exploratory demonstration of its applicability to text-conditioned thermal image generation.

Financial exploitation is a growing concern for people with Alzheimer's disease, especially during periods of reduced cognitive stability. Conventional fraud detection systems usually rely on financial behavior alone and ignore clinically relevant factors that may alter vulnerability. This paper proposes a medication-aware framework that synchronizes medication adherence with transaction-level monitoring to improve detection of cognitively risky financial events. A hybrid simulation dataset was constructed for 180 patients across 45 days, producing 8,100 medication records and 30,855 transactions. The framework evaluates amount anomaly, vendor novelty, transaction frequency, time deviation, and medication adherence through financial-only, additive medication-aware, and interaction-aware logistic models. Results show that the financial-only baseline obtained the highest global F1-score of 0.5000, but the interaction-aware model improved recall during medication-induced vulnerability windows from 0.7442 to 0.9070 and achieved the highest average precision for ranked high-risk cases. The findings suggest that medication adherence is most useful as a contextual modifier of financial risk rather than as an isolated predictor.

We present AXIOM, a trust-first neuro-symbolic execution architecture for natural-language mathematical reasoning. In AXIOM, the language model functions strictly as a canonicalizer: it rewrites informal problem text into a narrow schema consumed by a deterministic Computer-Algebra-System (CAS) pipeline, which derives and verifies the answer or abstains as a first-class output. Routing follows a 1:1:1 alignment between problem-shape regex, schema-specific prompt, and closed-form CAS handler, with 3,100+ such routes shipped and zero LOST_CORRECT regressions across 250+ consecutive ship commits. We report empirical results on 4 MATH categories with a cumulative correctness of 94.36% (2,592/2,747) at 100.00% trust on parseable (zero confident-wrong answers across the full 2,747-record benchmark), all four domains above the per-domain 70/90/70 floor with per-domain trust at 100.0%, and median latency of 1 ms on rule-only handlers (88% of records on the lm-eval arithmetic 20,000-record benchmark). The architecture has served ~30,000 production queries through a public deployment. The contribution we emphasize is not a final accuracy figure but the forward dynamic the architecture establishes: every logged abstain in production is a candidate correct after one ship cycle, since new tasks compose without regressing the registry. The operational discipline behind this property -- math-template bucketing, LOST_CORRECT scan as regression oracle, parseable-first onboarding, and abstain as first-class output -- constitutes a transferable framework for trustworthy neuro-symbolic systems beyond mathematics.

Face-to-face speech comprehension is inherently multimodal, integrating acoustic signals with visible articulation, facial expression, head motion, and other socially relevant cues. While audiovisual speech systems typically focus on the mouth region as the primary visual source of linguistic information, affective facial expressions are often treated separately as emotion-recognition targets. This paper investigates whether upper-face affective information contributes to audiovisual sentence recognition beyond audio and mouth-region cues, particularly under acoustic degradation. Using the CREMA-D audiovisual emotional speech corpus, we train feature-based sentence classifiers under four cue conditions: audio only (A), audio plus mouth/lower-face features (A+M), audio plus upper-face features (A+U), and audio plus both mouth and upper-face features (A+M+U). Models are evaluated on clean audio and pink-noise conditions at +10 dB, +5 dB, and 0 dB SNR using actor-independent splits. Results show that mouth/lower-face features provide substantial robustness benefits under degraded audio. At 0 dB SNR, A+M improves accuracy over A by 0.0794, with an actor-bootstrap 95% confidence interval of [0.0296, 0.1298]. Upper-face affective cues exhibit a more nuanced effect. Although the direct accuracy gain of A+M+U over A+M is small, full-face models consistently improve calibration across SNR levels and outperform shuffled upper-face controls under noisy conditions. These findings suggest that affective facial information may support multimodal robustness and confidence estimation under acoustic uncertainty without directly encoding lexical content. More broadly, the study highlights the potential role of socially expressive facial cues in human-centered audiovisual interaction systems.

Embodied world models have emerged as a promising paradigm in robotics by predicting how robot actions affect the surrounding scene. However, the rollout inference remains computationally expensive in pixel space, as long-horizon manipulation videos typically have to be generated frame by frame. This cost cannot be easily reduced by indiscriminately dropping frames, since downstream policies rely on complete preservation of sparse task-relevant events such as approach, contact, grasp, and release. To address this challenge, we propose Sparse Keyframe Interpolation Paradigm (SKIP), an event-preserving sparse-to-dense framework that avoids dense frame-by-frame generation. SKIP first identifies task-relevant keyframes by leveraging robot-aware multimodal features. It then synthesizes only these keyframes with a sparse video diffusion model. A learned gap predictor and an action-conditioned interpolator subsequently reconstruct the missing intervals according to the robot actions. On LIBERO, SKIP generates dense rollouts $4.16\times$ faster than a dense baseline while improving visual fidelity and reducing aggregate FVD by $89.0\%$. Importantly, SKIP-generated videos are effective policy-training data. Even when they fully replace real demonstrations, $π_{0.5}$ success drops only $1.3$ pp in LIBERO simulation and $6.7$ pp on the real robot, whereas fully dense frame-by-frame generation collapses by $48$ to $58$ pp.

This report presents TAP-JEPA, our runner-up submission to the EPIC-KITCHENS-100 (EK-100) Action Anticipation Challenge at EgoVis 2026. The task is to anticipate the next verb, noun, and verb-noun action from an egocentric clip that ends before the target action begins. Instead of fine-tuning a large video backbone, TAP-JEPA builds a compact anticipation model on frozen V-JEPA 2.1 features: a ViT-G/384 encoder extracts visible pre-action tokens, the pre-trained latent predictor estimates near-future tokens from the observed context, and both token groups are fused by attentive probes with task-specific queries for verbs, nouns, and action pairs. For the final submission, we expand supervised training with the official training split and most of the validation split, reserving a small subset for sanity checks and qualitative inspection, and adopt a two-stage score fusion that first averages eight independently initialized probe replicas within each epoch and then merges candidates from epochs 12-20 with field-dependent weights. On the official open-testing leaderboard, our sunshinesky entry achieves 27.91 percent overall action Mean Top-5 Recall (MT5R), ranking second and only 0.04 percentage points behind the top score.

Large video diffusion models achieve strong visual quality but remain expensive to deploy because each sample requires many denoising steps and a large resident parameter footprint. This paper studies a deployment-oriented compression pipeline for Wan2.2-T2V-A14B by combining few-step distribution-matching distillation with low-bit quantization. The pipeline follows the model's dual-expert denoising route, calibrates the high-noise and low-noise branches separately, protects sensitive entrance layers, and uses HiF4-style low-bit representation to improve dynamic-range coverage. Quantization is calibrated on the distilled few-step student rather than on the original long-step trajectory, reducing activation-distribution mismatch during inference. The proposed co-design keeps the quantized model close to the same-step full-precision model and surpasses the original full-precision baseline at 8 and 20 steps on average. The 20-step setting gives the best quality-efficiency trade-off in the tested configurations.

Mixture-of-Experts (MoE) architectures scale Large Language Models (LLMs) efficiently, enabling greater capacity with reduced computational cost by dynamically routing inputs to relevant experts, yet introduce a critical vulnerability: Safety Sparsity, where safety capabilities concentrate in few experts, making them susceptible to adversarial bypassing. Meanwhile, conventional alignment methods uniformly adapt all parameters, ignoring their functional differences and inadvertently degrading performances. To address these challenges, we propose MESA (MoE Safety Alignment), a targeted alignment framework for MoE-based LLMs that strategically decentralizes safety responsibility to maximize coverage while minimizing interference with utility. Based on Optimal Transport (OT) theory, MESA operates through two mechanisms: (1) Expert Capacity Reallocation uses a transport cost matrix to distribute safety duties to the most cost-effective experts, and (2) Dynamic Routing Refinement constrains the router to precisely activate these decentralized modules. Experiments show that MESA achieves robust defensive performance against varied harmful benchmarks while preserving helpfulness. Code is available at https://github.com/lorraine021/MESA.

Detecting psychological defense mechanisms in conversational text remains a challenging clinical NLP problem. For the PsyDefDetect 2026 shared task (nine-class utterance classification evaluated via macro F1), our team LinguIUTics achieves a macro F1-score of 0.3917 on the official positive-class leaderboard, ranking 4th out of 21 registered teams and improving over the Ministral-8B task baseline (31.48 macro F1) by 7.7 absolute points (24.4 percent relative). BERT-family encoders and zero-shot LLMs proved ineffective on rare classes due to severe class imbalance, leading us to QLoRA fine-tuning of Qwen3-8B. We leverage three key strategies: grouped stratified cross-validation (preventing leakage), minority-class round-robin lexical augmentation, and a post-processing pipeline with logit bias tuning and ensemble blending. Together, these components close much of the validation-to-leaderboard gap and substantially improve minority-class recall, driving the critical "Unclear" class (Level 8) from near-zero performance to an F1 score of 0.797.

Reasoning traces have become a valuable form of learning signals for improving and transferring the capabilities of large language models. In particular, detailed traces can help distill reasoning behavior from stronger teacher models into weaker student models. The value of capability transfer has motivated many deployed systems with reasoning models to hide raw internal traces and expose at most summaries and answers to users. As a result, we ask whether such interface-level trace hiding prevents users from obtaining useful reasoning supervision through prompting. We study this question with Reasoning Exposure Prompting (REP), a lightweight in-context elicitation method that uses shadow-model-generated demonstrations wrapped in auxiliary code-like formats to raise user-visible reasoning traces from a victim model. Across the common reasoning dataset, different victim models, and different student model distillation, REP substantially increases similarity between exposed and REP-conditioned internal traces while preserving useful reasoning signals.

A new framework for the estimation of the complexity posed by video-question pairs to video-LLMs, Video Attribute-Based Complexity (VideoABC), is proposed. Video complexity is defined as the probability of failure of a video-LLM for a given video-question pair. VideoABC is a non-parametric complexity measure, using a reference video dataset and a pre-defined vocabulary of video attributes informative of complexity, \eg the scene complexity or the speed of the video event informative of the question. In a training phase, reference videos are projected into the space of these attributes, which is then quantized. The expected ABC of each quantization cell is then computed. Given a new video and its projection into the attribute space, complexity is estimated by the expected ABC of the associated quantization cell. To enable the use of VideoABC with small reference video datasets, two quantizers are combined: a k-means quantizer that enables accurate complexity estimates for samples in the distribution of the reference dataset and a universal lattice quantizer that guarantees generalization to out-of-distribution samples. A synthetic video generation procedure, inspired by target-distractor manipulations of psychophysics studies, is proposed to populate the cells of the lattice quantizer during training, enabling the computation of their expected ABCs. Experimental results show that VideoABCis effective even with very low-dimensional attribute representations, substantially outperforming approaches like `video-LLM as judge' with much less complexity. Finally, the explainable nature of the VideoABC score, in terms of well-defined attributes, is shown to provide insights on how the attribute composition of benchmarks affects their complexity.

Although reinforcement learning has significantly advanced humanoid locomotion, perceptive policies still struggle on sparse-foothold terrain and constrained environments. Success in these scenarios requires both broad terrain awareness and precise foothold selection, two perceptual roles that conventional encoders often entangle. To address this challenge, we propose Global-Local Attention Decomposition (GLAD) for terrain encoding in humanoid locomotion. Realized by a coarse-to-fine encoder over a robot-centric elevation map, GLAD explicitly separates these objectives: a global attention branch utilizes attention pooling to summarize the surrounding terrain context, while a state-conditioned local attention branch sparsifies and encodes precise foothold-relevant geometry. This explicit attention decomposition prevents the dilution of fine-grained spatial cues while reducing training overhead. Experiments demonstrate that GLAD enables reliable locomotion over challenging gaps, stepping stones, and stairs. Furthermore, the learned policy exhibits emergent terrain-responsive behaviors, autonomously following narrow paths and avoiding obstacles under simple velocity commands without explicit navigation planners. In real-world deployment on a Unitree G1 humanoid robot using onboard LiDAR, the proposed method achieves robust zero-shot sim-to-real transfer across diverse sparse-foothold and obstacle-rich domains.

Modern VAEs are rarely trained with the pointwise likelihood implied by the standard $β$-VAE objective. In practice, pointwise reconstruction is often combined with perceptual and adversarial losses, despite a lack of understanding of how this changes the latent dynamics of the model. We show that the choice of reconstruction loss reshapes the rate-distortion problem itself, altering both the information content and the geometry of the learned latent space in ways that may be invisible from reconstructions alone. First, we prove and verify empirically that augmenting pointwise reconstruction with neural terms, such as perceptual and adversarial objectives, reduces the amount of information stored in the latent representations. Second, we show that neural reconstruction losses systematically change the geometry of the latent space: they make representations more isotropic and distribute uncertainty more evenly across latent dimensions, producing different posterior variance profiles. These findings highlight how the rate-distortion tradeoff is not a comprehensive lens to understand the behavior of VAEs, and we propose a more mechanistic approach to investigate how the choice of a distortion metric reshapes the optimization problem.

This article evaluates the integration of data extracted from a French syntactic lexicon, the Lexicon-Grammar (Gross, 1994), into a probabilistic parser. We show that by applying clustering methods on verbs of the French Treebank (Abeillé et al., 2003), we obtain accurate performances on French with a parser based on a Probabilistic Context-Free Grammar (Petrov et al., 2006).

Intraoperative ultrasound (ioUS) is a versatile, cost-effective modality in brain tumour surgery, but its interpretation is difficult: acquisition planes are non-standard, artefacts are modality-specific, and its appearance differs markedly from the preoperative MRI on which surgical-planning tools, segmentation models and the surgeon's experience rely. Synthesising MRI-like images from ioUS could let this MRI-based infrastructure be reused intraoperatively without an extra scan. Most prior work evaluates a single architecture in isolation; to our knowledge, no benchmark has spanned architectural paradigms, inference regimes and downstream-task endpoints under a common protocol. We address this gap on the public ReMIND data set (76 patients; 153 paired ioUS/T2w and 104 paired ioUS/FLAIR studies; 60/16 patient-level train/held-out split). Six generators (four GAN baselines: Pix2Pix, SwinPix2Pix, CycleGAN, CUT; the transformer-augmented ResViT; and the few-step diffusion model SynDiff) were each trained under four inference regimes (2D, 2.5D, 2D + 3D-refinement, full-3D) and two targets (T2w only; T2w + FLAIR multi-task), yielding 48 experiments. Image-fidelity metrics (SSIM, PSNR, MAE, LPIPS) were complemented by an nnU-Net v2 downstream segmentation evaluation (tumour and resection cavity) and by subgroup analyses by histological grade and reoperation. No architecture dominated every axis, and, critically, perceptual quality tracked downstream utility most closely (LPIPS, r=-0.66, p<0.001), whereas higher SSIM was associated with worse utility (r=-0.64, p<0.001); SynDiff-2.5D best preserved downstream segmentation (U_Dice=0.55). Perceptual and downstream-task metrics should therefore be reported alongside or in preference to global SSIM, and architecture choice conditioned on surgical phase, patient history and clinical objective.

Sound design workflows frequently oscillate between time-consuming library searches and the complexity of procedural synthesis, with practitioners typically relying on disconnected tools to address each challenge separately. This paper introduces Quality Audio Prototyping (QuAP), a working prototype that unifies content-based audio retrieval and procedural sound generation within a single interface, reducing the procedural distance between a narrative concept and its sonic realisation. QuAP integrates a similarity-based retrieval engine with real-time procedural audio models, complemented by a rule-based assistant that provides perceptually informed parameter guidance, offering definitions and recommendations derived from empirical optimisation rather than requiring prior synthesis knowledge. Preliminary evaluation confirms the viability of this approach: subjective assessment demonstrated statistically significant quality improvements in five of six embedded synthesis models, and an encoder ablation study established the preferred retrieval architecture on a sound effect dataset. A user evaluation with 16 practitioners confirmed the tool's workflow utility, with all participants agreeing that the parameter assistant preserved creative agency while lowering the barrier to procedural interaction.

Self-distillation improves learning efficiency by rewriting reference answers as training data that better matches the model's own distribution. However, reference answers also introduce strong stylistic biases, causing the generative model to imitate surface forms rather than learn useful reasoning patterns. We observe that the rewriting data contains a large number of high-perplexity (PPL) tokens, coming from two distinct sources: beneficial knowledge-enhancing logical corrections, and harmful stylistic drift induced by reference imitation. Treating all such tokens equally can disrupt the base model's original distribution and degrade performance, especially on difficult reasoning tasks. To address this, we propose Distribution-Aligned Self-Distillation (DASD), which uses an answer-aware reference model to generate candidate tokens and dynamically filters them according to the base model's confidence. DASD preserves tokens that encode useful logical knowledge while suppressing distributionally misaligned style noise. Experiments on math, code, and commonsense reasoning benchmarks show that DASD consistently outperforms competitive baselines, reduces high-PPL tokens, and improves robustness across tasks of varying difficulty.

Multimodal large language models (MLLMs) demonstrate remarkable visual understanding, yet their reliability in interactive settings is severely undermined by hallucination snowballing: a phenomenon where initial errors amplify across conversational turns, leading to a collapse in coherence. This failure reveals a fundamental vulnerability where models progressively neglect visual grounding in favor of over-relying on polluted textual history. Existing benchmarks are predominantly confined to single-turn VQA, which fail to capture the complex dynamics of error propagation in long-horizon interactions. To address this, we introduce MM-Snowball, the first benchmark for fine-grained diagnosis of hallucination snowballing within dialogues. Extensive evaluation shows that our benchmark poses a significant challenge even to advanced MLLMs and reveals the inefficacy of existing mitigation methods designed for single-turn VQA. To counteract this degradation, we propose Conflict-Aware Visual Rectification (CAVR). This training-free method mitigates snowballing through a synergistic dual-mechanism that refreshes visual grounding at the representation level and rectifies output distributions at the logit level, effectively re-anchoring the model to visual facts. Experiments demonstrate that CAVR achieves state-of-the-art performance, offering a promising path toward more reliable interactive AI. Data and code are available at: https://frenkie-chiang.github.io/MM-Snowball

Recent Large Multimodal Models (LMMs), primarily designed for offline settings, are ill-suited for the dynamic requirements of streaming video. While recent online adaptations improve real-time processing, they still face critical scalability challenges, with resource demands typically growing at least linearly with video duration. To overcome this bottleneck, we propose FlowNar, a novel framework for scalable streaming video narration. The core of FlowNar is a dynamic context management strategy for historical visual context removal, combined with our CLAM (Cross Linear Attentive Memory) module for streaming visual history retention, ensuring bounded visual memory usage and computational complexity, crucial for efficient streaming. We also introduce a realistic self-conditioned evaluation protocol and complementary evaluation metrics to assess streaming narration models under deployment-like conditions. Experiments on the Ego4D, EgoExo4D, and EpicKitchens100 datasets demonstrate that FlowNar substantially improves narration quality over strong baselines while being highly efficient, supporting processing of 10$\times$ longer videos and achieving 3$\times$ higher throughput (FPS). The code is available at https://github.com/zeyun-zhong/FlowNar.

Long-horizon autonomous agents require memory systems to retain historical information, track evolving states, and reuse relevant knowledge beyond finite context windows. Existing agentic memory systems typically follow a memory construction-retrieval (MCR) pipeline, but often adapt mainly the memory bank while keeping the surrounding pipeline fixed after deployment. This fixed-pipeline design struggles to handle heterogeneous task-specific failure modes and can become misaligned with memory banks that evolve in scale and structure over time. To address these limitations, we propose MemPro, a system-level evolution framework that treats the entire MCR pipeline as an evolvable program rather than adapting only the memory bank or prompt text. MemPro maintains a version tree of runnable memory-system implementations, where an Evolving Agent iteratively selects promising versions, diagnoses recurring failures, and creates improved child versions through failure-mode-guided edit-debug refinement. Experiments on LongMemEval, LoCoMo, HotpotQA, and NarrativeQA show that MemPro consistently outperforms strong static and prompt-level evolving baselines within a few iterations, continues to improve with evolution, and achieves a favorable performance-cost trade-off. Code is available at https://github.com/wanghai673/MemPro.

Generative models have emerged as a powerful paradigm for AI planning, yet their performance remains constrained by the training data distribution. One approach is to improve generated solutions during inference by scaling test-time compute. A more efficient alternative is to optimize the inference process itself. In this paper, we show that a modified version of a classical Open-Closed List (OCL) search provides just such an efficient inference procedure. Our algorithm synergizes two learned components: a generative model that performs fast rollouts from intermediate states and a heuristic model that prioritizes among candidate reasoning paths. Key contributions include novel exploration control mechanisms and integration of learned models within the OCL framework. Across multiple combinatorial planning domains, our approach outperforms both neurosymbolic search baselines and classical solvers in computational efficiency and solution quality.

Watermarking should identify language-model output without degrading quality or limiting verification to the model provider. Multilingual deployment makes this harder because morphology, segmentation, and script change where watermark evidence can enter naturally. We introduce LUNA, a linguistically adaptive watermark that combines model-free detection with single-token non-distortion under the standard random-key model. LUNA estimates normalized next-tag entropy from part-of-speech contexts in an external corpus and uses it to set the depth of a non-distortionary binary tournament sampler; the detector reconstructs the same schedule from text, a tokenizer, a tagger, and a secret key. We evaluate six typologically diverse languages and two domains against eight primary baselines. LUNA attains an AUROC of 0.9959 and the lowest mean absolute median perplexity shift of 0.045 across the twelve settings; its 95% bootstrap interval [0.022, 0.073] lies below all baseline intervals. LUNA also records the lowest mean Self-BLEU, Distinct-1, surprisal, and entropy shifts. It is the only method that simultaneously achieves AUROC > 0.99 and an absolute median perplexity shift below 0.1 in a majority of settings, reaching this regime in 9 of the 12 settings while no baseline reaches it in more than 2. Our code is available at: https://github.com/Shinwoo-Park/luna_watermark

Long-horizon LLM agents produce safety evidence across long trajectories, where sparse, delayed, and compositional risk signals often escape local moderation. Existing turn-level or short-context detectors struggle to reliably retain and aggregate such evidence over extended horizons. We reframe long-horizon agent safety detection as trajectory-level evidence compression and propose Trajectory Risk-Aware Compression for Long-Horizon Agent Safety (TRACE). TRACE uses a Compressor-Reader design: the Compressor encodes the full trajectory into a compact latent evidence state under trajectory-level supervision, and the Reader judges the raw trajectory with this latent evidence state as a safety reference. This design helps aggregate dispersed risk cues and reduce premature evidence loss. Across ASSEBench, Pre-Ex-Bench, and R-Judge, TRACE achieves the best accuracy on all evaluated backbones, improving over strong baselines by up to 12.6 percentage points. On LongSafety, TRACE shows smaller performance degradation as context length grows. Attention visualizations and case studies suggest that the compressed reference helps the Reader focus on risk-critical segments and recover cross-step evidence. Code is available at https://github.com/Peregrine123/TRACE_official.

Reinforcement learning (RL) with verifiable rewards has achieved strong progress in reasoning-oriented LLMs, but extending it to multi-domain RL remains challenging due to reward unreliability in non-verifiable tasks and capability interference across domains. We propose CARE-RL to combine protocol-aware reward generation with capability-aware optimization for mitigating cross-domain conflicts. For non-verifiable tasks, the Protocol-Aware Generative Reward Model (PA-GRM) constructs prompt-level evaluation protocols and schemas before producing trace-conditioned rewards, enabling task-adaptive yet comparable evaluation of open-ended responses. For multi-domain optimization, Direction-Aware Capability Subspace Projection (DACSP) extracts historical capability directions from previous RL stages and modulates later updates by amplifying aligned components, suppressing conflicting components, and preserving orthogonal updates. Experiments across math, chat, and instruction-following benchmarks show that CARE-RL consistently outperforms standard multi-domain RL baselines, achieving Total Avg scores of 47.9 and 50.7 on Qwen2.5-7B and Qwen3-4B, respectively.