Papers

Object-wise reconstruction increasingly inserts generated meshes into captured scenes as editable proxies for observed objects. This use creates a requirement absent from isolated-asset scores and pose estimates: at the captured metric state, the proxy should reproduce calibrated observations of the same instance. We present ObjectTwin, a scan-grounded measurement protocol that freezes the instance, cameras, scan-projected support and recorded depth, metric anchor, and evaluation denominator, then replaces only the object representation. The resulting asset correspondence gap measures lost observation agreement relative to scan-derived visible geometry without treating the scan as complete physical ground truth. Across two indoor RGB-D capture domains and two generation backends, visible, anchored generated assets lose 0.302-0.497 MaskIoU under this replacement. Every tested scene-level estimate is positive, while same-input generator preference reverses across capture domains. Coverage and overlap-depth analyses locate disagreement in both visible support and surface placement; the evaluated conservative pose refinement does not systematically recover it. These results establish capture correspondence as a separately measurable interface between generated-asset quality and stronger physical or task-level evaluation.

Computational pathology (CPath) analyses increasingly depend on multi-step computational workflows that require substantial programming expertise, limiting their accessibility to pathologists and other domain experts without extensive coding experience. Although large language models (LLMs) provide a promising natural-language interface for constructing such analyses, hallucinated operations and incorrect workflow specifications limit their reliability in scientific settings. We introduce Tissue Image Analysis Agent (TIAgent), an automated CPath workflow generation framework that, given a natural-language prompt specifying a complex analytical task together with an input dataset, automatically constructs a verifiable computational graph employing existing CPath operations. The resulting graph can be inspected and executed, enabling natural-language construction of complex CPath workflows for applications including automated biomarker discovery and tissue analysis. We demonstrate end-to-end workflow generation and execution and systematically benchmark 14 backbone LLMs across 90 unit tasks spanning 18 CPath operation types. The ability of each LLM to generate a complete and correct unit workflow is quantified using Unit Workflow Success (UWS). Performance varied substantially across backbone models, with Kimi-K2.7-Code achieving the highest UWS of 0.967 (95% CI: 0.907-0.989). Analysis of node and parameter correctness further revealed that selecting the appropriate computational operation and correctly configuring it represent distinct challenges for LLM-based workflow generation. TIAgent provides a foundation for making complex CPath workflow construction more accessible while enabling systematic evaluation of LLM-driven scientific workflow generation.

Model fusion aims to integrate the capabilities of multi-expert models into a single model. Existing evaluations offer limited coverage of reasoning with long chains of thought and tool use, while strong average performance can mask substantial losses in individual expert capabilities. We introduce FusionEval, a benchmark that provides controlled expert pools, integrates fusion methods with and without additional training, and offers a unified evaluation suite. The pools contain domain experts independently trained from Qwen3 and Llama base models across math, code, science, instruction following, and tool use. The suite evaluates capability retention and computational cost on eight public benchmarks under comparable data access rules, reporting normalized average and worst performance, runtime, and peak GPU memory usage. Extensive experiments show that training-free methods preserve reasoning performance but substantially degrade tool use at both model scales. In contrast, MOPD preserves tool use and achieves a more balanced capability profile, but requires additional data preparation, inference, and training. Jointly retaining reasoning, instruction following, and tool use at low computational cost remains a challenge. We hope FusionEval supports future research on efficient model fusion that jointly preserves expert capabilities. Our code is available at https://github.com/Baicaihaochi/FusionEval_Benchmark.

Many Indian banks serve customers through FAQ pages and scripted chatbots. Assistants based on language models could respond faster and carry out requests such as blocking a card, but only if they give correct information and follow the bank's rules. These rules prescribe a procedure, a set of required steps, such as obtaining a closure quote before closing a deposit. Adapting models to this domain requires measuring how well they follow the procedure, before and after training. Benchmarks usually judge only the records left at the end of a conversation. Many required steps, however, leave the records unchanged, so an assistant that skips them can still be judged successful. We present TauIndianBankBench, an action-oriented benchmark of 1,000 Indian retail-banking tasks where an assistant helps a simulated customer. Besides the final records, it checks that the assistant took every required step and made no unrequested change. For every model, 6.5-12.0% of sessions are judged successful when only the final records are checked, although the assistant skipped a required step or made an unrequested change. Scoring both procedure and outcome shows more accurately how well language models perform in Indian retail banking.

Universal Manipulation Interface (UMI) pairs wrist-camera videos with camera trajectories and gripper openings, providing scalable, robot-free data for learning how bimanual actions change a shared scene. Existing interactive world models, however, typically rely on text or single-camera control and struggle to follow each arm's actions while maintaining consistency between the two wrist views. We introduce UMI-WM, an action-conditioned bimanual world model that jointly generates both wrist views from observation history, per-arm relative camera poses, and gripper widths, with cross-view consistency, action following, and visual fidelity. UMI-WM injects each arm's actions into its own wrist stream, while sparse hub tokens exchange scene information between the two streams. Camera-aware geometry conditioning and camera-guided noise further support per-view motion control and cross-view geometric exchange. We adapt the model for block-causal generation and distill the 30-step Teacher into a Student with four denoising steps. We train UMI-WM on self-collected open-world and in-studio UMI recordings combined with public manipulation data. To evaluate action following, we use an inverse dynamics model (IDM) to recover relative camera poses and gripper widths from generated videos, and introduce Bimanual Action Fidelity (BAF) to measure how closely these recovered actions match the conditioning actions. UMI-WM achieves the best results on all five reference-video fidelity metrics and the best or competitive performance on VBench and WorldArena. The Teacher and Student outperform all compared baselines in action following, while the four-step Student generates paired wrist videos 6.9 times faster than the Teacher.

Thin filaments cross in microscopy, and dispersed spectra cross on X-ray detectors. At a junction, where these structures cross or branch, how many straight arms meet, and which way do they go? Unknown blur obscures the answer. At a balanced crossing, distinct Gaussian blurs can even produce the same observation law. We show that the arms can still be recovered. For separated rays of nonvanishing weight and an unknown centered Gaussian-mixture blur, a distance combining arm count, directions, relative weights and supporting lines is bounded by a constant times the Euclidean difference in cell probabilities on a sufficiently fine fixed grid. The bound is uniform over a bounded model class, including unidentifiable blurs. From $n$ positions or photon counts, the minimax error is of order $n^{-1/2}$. This rate is proved for a global estimator. In practice we fit rays and blur by constrained local likelihood, reducing the BIC count when two detectors agree on fewer arms in the microscopy studies. Across three studies with protocols fixed before evaluation, on 360 crossings and branches assembled from measured single-molecule localizations, the fit gets the arm count and every direction within $10^\circ$ 24 to 37 percentage points more often than a tuned fixed-scale Hessian ridge detector, averaged over 25 to 400 localizations. On acute angles with correct counts, mean absolute error is $3.5^\circ$ to $4.3^\circ$, versus $10.4^\circ$ to $15.5^\circ$. On Chandra spectra, 38 of 90 fits at 800 photons find all four arms within $10^\circ$; among these, median error is $0.023^\circ$.

Optic disc (OD) and optic cup (OC) segmentation is central to automated glaucoma assessment, but models trained on limited labeled fundus datasets often degrade under domain shift. Semi-supervised Teacher-Student learning can improve generalization by leveraging unlabeled data, yet standard Noisy Student training treats all pseudo-labels as equally reliable. In this paper we investigate a targeted extension of Noisy Student training in which OD- and OC-specific pseudo-label supervision is modulated using a mask-derived reliability estimate. A multitask Dice predictor estimates OD- and OC-specific pseudo-label quality from Teacher-generated masks, and these estimates are used to weight the pseudo-label loss during Student training. This preserves the baseline contribution of low-scoring foreground pseudo-labels while assigning greater relative emphasis to pseudo-labels predicted to be more reliable. Experiments across known and unseen domains show improvements over the Noisy Student baseline, particularly for OC segmentation, while maintaining competitive generalization performance.

Characters, locations, and actions recur throughout a movie, so many of its moments differ only in fine-grained details. Grounding such a moment in a full-length movie requires resolving that detail rather than matching the attributes it shares with other moments, yet existing temporal movie grounding benchmarks rarely test this. We introduce CoM$^2$-Bench, a benchmark of confusable query pairs, where the two queries in each pair share most attributes but differ in a single distinguishing detail within a full movie, posing a challenge for recent temporal grounding approaches. To locate where grounding breaks down, we analyze a retrieve-then-localize pipeline, in which a retriever first selects a short candidate chunk and a localizer then performs temporal grounding within it. Separating the two stages lets us attribute each failure to either retrieval or localization. We find that retrieval is the major bottleneck: state-of-the-art retrievers often select a moment's confusable counterpart, because their choices follow the content the two moments share rather than the detail that tells them apart. To address these failures, we propose a simple, training-free remedy, Agentic Window Search (AWS): an MLLM agent $\textit{Decomposes}$ the query into attribute-level sub-queries, then $\textit{Inspects}$ each candidate window against them and $\textit{Expands}$ the window when evidence is cut off at its edge. AWS works with existing retrievers and localizers without additional training, raising mIoU from 18.5 to 26.8.

Large image editing models increasingly infer editing intent from both textual and visual inputs, creating an underexplored safety risk when harmful intent is distributed across modalities. In this work, we formalize this split-intent vulnerability and propose the Split-Intent Jailbreak Attack (SIJA), which asymmetrically distributes an unsafe editing intent by specifying the editing referent in text while concealing the safety-critical action and specification through graphical visual semantics in an auxiliary cue. To systematically study this threat, we introduce the Split-Intent Image Editing Safety Benchmark (SI2ESBench) for evaluating image editing models under split-intent inputs. Extensive experiments on representative commercial and open-source models demonstrate that SIJA effectively compromises state-of-the-art image editing systems, achieving ASRs of 68.28% on GPT Image 2 and 91.38% on Qwen-Image-3.0-Pro. Our results expose a mismatch between multimodal editing and safety assessment: the distributed intent can be recovered to execute the requested transformation, yet its harmfulness often remains undetected by existing safeguards. To mitigate this vulnerability, we propose editing-intent reconstruction, a defense principle that explicitly evaluates the complete transformation implied by the joint multimodal context and substantially improves rejection of split-intent attacks. Our findings expose a cross-modal safety gap in modern image editing systems and highlight the need to assess the jointly inferred transformation rather than individual inputs in isolation. https://github.com/yexianmin/SIJA

Vision-Language Models (VLMs) are increasingly deployed as general-purpose multimodal assistants, making multimodal safety alignment an important challenge. Safety depends on the interplay between textual and visual content, with visual context often playing a critical role. However, VLMs may fail to use visual information appropriately, either overlooking safety-relevant visual content or over-interpreting sensitive visual content, leading to unsafe responses or over-refusal. Existing multimodal safety alignment datasets largely follow the response-preference paradigm, which teaches models how to respond but provides limited supervision on how visual context should inform the safety judgment, making it difficult to balance safety and over-refusal. To address this limitation, we introduce ContraSafe-VL, a Contrastive Safety preference dataset for Vision-Language Models. For each fixed textual query, ContraSafe-VL constructs paired visual contexts that result in multimodal queries with different safety labels, providing direct supervision for how visual context should inform multimodal safety judgments. It further provides response preference pairs based on harmlessness and helpfulness to encourage models to generate responses that are both safe and helpful. Experiments across three VLM architectures and multiple benchmarks demonstrate that training with ContraSafe-VL improves cross-modal safety recognition and achieves a better safety–over-refusal trade-off while preserving general multimodal capabilities. The dataset will be publicly released.

Real-world egocentric videos capture multi-step household procedures, but their recorded trajectories leave alternative actions and recovery from the learner's own mistakes largely unexplored. We introduce EgoVine, a post-training framework that turns a fixed demonstration set into interactive experience for long-horizon planning. EgoVine operates in executable symbolic environments compiled from real-world video annotations, where an agent must coordinate object manipulations under partial observation and account for delayed consequences. Restoring a decision state allows the learner to explore alternative actions and compare their eventual outcomes. Union jointly trains action generation and consequence prediction in one model, using the same executions to supervise candidate comparison and recovery. Verified repairs add recovery targets from failed attempts, while comparisons among sampled continuations guide policy updates. The updated planner then returns to the environments, with its own decisions and mistakes shaping the next round of experience. On a shared benchmark, EgoVine-1.5B achieves 81.8% success on Main planning tasks and 66.7% on longer task compositions, outperforming all evaluated Flash LLM baselines on both. EgoVine connects real demonstrations to an iterative learning process that trains compact planners on choices and futures beyond the recorded trajectories.

AI-generated video (AIGV) detection is increasingly important for mitigating misinformation, impersonation, and other malicious uses of synthetic media, yet existing detectors often generalize poorly to unseen generators in open-world settings. To better understand this challenge, we analyze the failure modes of AIGV detectors under cross-generator out-of-distribution (OOD) shifts and observe that detector representations strongly encode generator-specific cues, which are associated with substantial ID-OOD performance drops. Further analysis shows that naively perturbing generator-sensitive tokens can even reduce OOD performance, indicating that generator-related cues and detection-relevant evidence are only partially separable. Motivated by these findings, we propose Forensic Obfuscation of geneRator-specific cues via Gradient-guided Embedding perturbation (FORGE), a token-aware adversarial training framework for open-world AIGV detection. FORGE identifies tokens that are strongly associated with generator-specific shortcuts yet less critical to real/fake discrimination, and selectively perturbs them during training to suppress generator-dependent patterns while preserving useful forensic evidence. Through synchronized adversarial training between an attacker and a detector, FORGE encourages the detector to learn more transferable generator-invariant forensic representations. Experiments on 29 held-out benchmark entries show that FORGE achieves an average F1 of 92.62, outperforming eight strong baselines by 5.31 points, while also remaining robust under four common post-processing attacks. We will release our code and datasets upon acceptance.

Retrieval-augmented generation asks one question at volume: does this evidence support this claim? The answer is one bit, yet most pipelines obtain it from a text generator and pay generation prices for the prose wrapped around it. We compare a non-generative decision model, which returns a typed value from a declared enumeration together with a native probability and no text, against a frontier LLM configured as an engineer would deploy it for binary classification. Both answer the same 495 human-labelled claims from LLM-AggreFact, so every comparison is within-item. We report three things. First, the two systems are statistically indistinguishable on the metric this task is conventionally scored by, a paired difference of -0.56 points, Jev minus Astra, at a 187-fold cost difference. Second, that tie carries almost no information about deployment: the systems disagree on 15.2% of claims, and the cheaper one admits 23.2% of the claims annotators marked unsupported against the frontier model's 14.1%. A component whose job is to catch unsupported claims should be judged on that rate, and the conventional headline hides it. Third, because the errors differ they compose: routing the claims the cheap model is confident about and escalating the rest reaches 74.10 at USD 2.92 per thousand claims against 73.83 at USD 8.93, with the threshold chosen out of fold. But a matched random-escalation control at in-sample thresholds, spending the same budget by coin flip, does as well, so what buys the saving is the escalation share and not the probability that picks what to escalate. Measured against the cheap model it is built from, the cascade costs 61-fold more for 0.83 points, 95% CI [-3.36, 4.86], and it is the more permissive guard of the two. Equal accuracy conceals unequal errors. All predictions, prompts, pricing and code are released.

Long-horizon failures of language models are usually measured as end-task success or as a horizon length. We instead score, step by step, three components of a 400-step ledger task in which every step is easy in isolation and the model's own answers are re-injected into its context: recognising a record already seen 5 to 300 steps earlier, processing the current record locally, and writing a counter incremented at every step. On Qwen3-8B (non-thinking, 20 seeds, replicated on a second machine and backend), with every step scored component by component, two results hold. First, duplicate-recognition errors rise late in the episode – from 0.0019 to 0.011 per step at risk across step 200 – while the counter makes no error in 8,000 steps and local processing is descriptively stable; the rise combines long-distance recall errors and rising false alarms. Second, a block of irrelevant content inserted before each step makes the counter fail in 20 of 20 seeds on each machine, against none without it, while its effect on local processing is much smaller and not distinguishable from none. Three pre-registered follow-ups probe the counter. Its value depends on the value present in the response history: a corrupted value persists without correction in every seed, and an exact external reminder, repeated before each message, re-aligns it within one to two steps. Under the irrelevant block, no counter error was observed when an exact compact reminder of the state, supplied externally by an oracle, preceded each message (0 of 20 seeds with an error, against 15 without it), and this reminder outperformed a length-matched neutral text; whether a single reminder suffices was not conclusive, and the absence of damage to the other components was not established. An aggregate or exclusive error classification can hide the difference we report.

Modern healthcare generates increasingly complex medical reports that challenge clinical professionals in terms of time and interpretive burden. Manual review processes are slow, error-prone, and difficult to scale. This paper presents MeddeX, a Medical Report Analyzer that combines a QLoRA fine-tuned LLaMA-3 (8B) model with Retrieval-Augmented Gen eration (RAG) and a multimodal Gemini Vision layer to deliver accurate, concise, and evidence-backed summaries of medical documents. The system handles text-native PDFs, scanned image based PDFs, and direct radiological images through an intelligent hybrid extraction pipeline that includes a novel garbage-text detection mechanism for corrupted and UTF-16-encoded doc uments. Beyond single-report Q&A, MeddeX introduces report scoped retrieval for precise contextual grounding, an automated longitudinal report comparison engine, and a self-healing backfill pipeline that lazily processes legacy documents on demand. A FastAPI backend, Supabase vector store, and Next.js fron tend complete the end-to-end architecture. Evaluation against the base LLaMA3-Med42-8B model demonstrates consistent improvements across ROUGE-1 (+9.2%), ROUGE-2 (+10.0%), ROUGE-L (+11.8%), BLEU (+71.4%), and BERTScore (+0.8%), validating the benefit of domain-specific fine-tuning. MeddeX represents a practical and robust step toward AI-assisted clinical decision support.

Romance scam has become a serious problem via social media, and the improvement of artificial intelligence enable to fake identify easily. In the study, a chatbot has developed to detect romance scam in real-time through Line Messaging API. Using the seven stage theories in the social science field, the stages of romance scam can be detected through large language model. In addition, once a romance scam is detected, the chatbot provides explanations of the stage and prevention advices through large language model. The contribution of the work provides a practical application, and can be used in real scenario for reducing the risk of romance scam.

Modern data pipelines increasingly combine human-created and human-curated data with AI-generated data for model training and evaluation. As datasets grow, they often add semantically redundant samples rather than expanding semantic coverage, which deteriorates the quality of the whole dataset. Existing diversity metrics primarily summarize lexical variation, entropy, or pairwise similarity, but provide limited characterization of the shape of the underlying data distribution. In this paper, we propose PLDiv, a diversity metric derived from persistence landscapes that quantifies the structural organization of a dataset from its geometric representation. PLDiv summarizes how data neighborhoods evolve as nearby samples progressively connect, producing an interpretable scalar measure of structural diversity without requiring task-specific supervision. Whether measuring intrinsic diversity in text generations or image embeddings, or detecting semantic collapse in instruction datasets and reasoning trajectories, PLDiv performs competitively against evaluated baselines and aligns strongly with diversity labels during semantic collapse. Consequently, it effectively supports applications in synthetic data monitoring, dataset construction, data augmentation, and generative model evaluation.

Looped transformers reuse one shared block across many recurrent steps, so depth becomes a quantity that can be spent at runtime rather than a stack of distinct layers. Those extra steps help only if a write made on one recurrent step is still readable on later ones, and that depends on the linear map applied to the residual between recurrent steps. We show that whether the loop learns to track state is a geometric question about the form of this map, and we propose blockwise conformal injections: the residual splits into coordinate-aligned groups and rotates inside each one, so every group has its own timescale while the geometry stays exact within it. Our experiments show a performance increase on a demanding form of state tracking - last-token accuracy when the rest of the prefix is already correct - and on language modelling, where a signal written earlier has to be retrieved at the end of the sequence.

LLM-as-a-judge evaluation commonly aggregates multiple judges' decisions into a majority vote, yet this panel decision process does not guarantee that the evaluation of an individual item is reliable. We propose Structured Cross-fitted Uncertainty Risk (SCUR), a label-free approach for estimating instance-level evaluation reliability based on the judge panel's behavior. A latent measurement model serves as the backbone for measurement system analysis, from which SCUR derives three complementary signals capturing panel uncertainty, measurement cross-entropy, and perturbation sensitivity. These three sources of evidence are combined into a risk score. We evaluate the SCUR approach on five real-world benchmarks using majority-vote error ranking, selective trust, and risk stratification. Furthermore, we show our approach to be more effective at detecting high-consensus risk items, which are the blind spots for many other LLM-as-a-judge evaluation methods.

Data selection for LLM fine-tuning faces two facts about real training: data pools draw from many distinct domains, and evaluation spans diverse tasks. As a result, the same sample helps different tasks differently, and samples interact during training so that a subset's value is not the sum of its parts. Current data-selection methods leave these realities largely unaddressed. We formalize them as Combinatorial Data Selection (CDS), a framework that separates attribution (how samples are scored) from selection (which subset to pick) and models the non-additive interactions among samples. On a real instruction-tuning data pool, we find substantial cross-domain complementarity (joint utility exceeds the sum of singletons), alongside redundancy within domains. Inspired by this, we propose Mixture-then-Select, a two-stage algorithm that first decides how many samples to take from each domain, then ranks samples within each domain by attribution score. Theoretically, the gap to the optimal subset is linear in the selection size and independent of data pool size. Empirically, Mixture-then-Select outperforms strong baselines by 20 to 54%, is robust across attribution methods, and works across different models and data pool sizes. The framework also extends to the data purchasing problem under a budget constraint when sample prices vary. Our findings suggest that allocating selection across domains offers a promising path toward more efficient and effective data selection for modern LLMs.

Many headphone users risk irreversible hearing damage during high-load listening, yet current protection still relies on sound-pressure limits. Clinical tests detect damage after it becomes measurable, leaving the risk-accumulating process unobserved. The acoustic reflex (AR), an involuntary response to high acoustic load, marks when the ear enters a protective state. AR traces are weak, and existing measurements require dedicated probes and controlled conditions that are difficult to maintain during everyday audio. We present EarLog, which turns playback, the dominant interference, into an opportunistic probe. EarLog introduces content-conditioned residual sensing to suppress variations caused by nonstationary content, individual ear acoustics, and device response. It identifies weak AR through loudness-dependent changes and localized spectral irregularities in the residual. A resonance-constrained sparse finite impulse response (FIR) model captures the short-window response. Dual-microphone observations, mild-distortion recovery, and cross-session reference updates keep detection usable under low signal-to-noise ratio (SNR), fit changes, and environmental drift. EarLog runs in real time on an embedded controller using a low-cost dual-microphone add-on. We evaluate it with 81 participants, 16 headphone models, and more than 400 hours of recordings. EarLog achieves 99.0% sensitivity and 98.7% specificity on a clinically referenced benchmark. EarLog records listener-specific AR events during everyday listening, providing data for future longitudinal studies of hearing risk.

Alzheimer's disease (AD) remains a major public health challenge, and early identification is crucial for timely intervention and care planning. Recent work has shown that speech-based digital biomarkers can capture subtle changes in vocal production, prosody, and timing that often precede or accompany cognitive decline. In this study, we investigate whether acoustic features extracted from spontaneous speech can reliably distinguish individuals with AD from healthy controls. Using a structured speech dataset derived from the Cookie Theft picture-description task, we extract clinically informative low-level descriptors spanning prosodic, spectral, and temporal domains, and evaluate a range of machine learning classifiers. Our results indicate that acoustic biomarkers carry strong discriminative value, with the best-performing models achieving clinically meaningful separation between diagnostic groups. These findings support the use of non-invasive, low-cost audio-based screening tools and motivate further work on multimodal and longitudinal assessment frameworks for dementia risk detection.

Identifying suicide risk from social networking services (SNS) posts is important for detecting suicide-related signals in online environments. However, risk classification alone provides limited insight into the textual evidence and psychosocial factors behind a prediction. Based on the IEEE BigData 2026 Explainable Suicide Risk Detection Challenge, this study presents a framework consisting of Risk Assessment, Evidence Grounding, and Factor Identification. Risk Assessment uses length-based routing to accommodate posts of different lengths. Evidence Grounding identifies supporting phrases and uses a Risk-Evidence constraint to maintain consistency with the Risk prediction. For Factor Identification, two verifiers are used. The Taxonomy Verifier focuses on factor semantics, whereas the Evidence-Aware Verifier uses factor-specific lexical-semantic cues to select informative positive training units. Their prediction probabilities are combined to produce the final factor predictions. The three tasks are evaluated using task-specific F1 score measures. Risk Assessment achieved a Weighted F1 of 0.8088, Evidence Grounding achieved a test Macro row F1 of 0.7605, and Factor Identification achieved a Macro F1 of 0.5562. The results show that the framework can provide risk predictions, along with supporting textual evidence and fine-grained information on psychosocial factors. Overall, the proposed framework extends suicide-risk assessment beyond risk-level prediction and provides a more interpretable analysis of SNS posts.

Large language models (LLMs) often abandon a correct answer, or endorse a user's position, once the user pushes back. This behavior, called sycophancy, is usually reported as a single rate per model, which says little about when it happens or how a user can avoid it. We study the conditions that produce it with 103,939 graded replies from ten configurations: eight LLMs with reasoning disabled, and two of them again with maximum reasoning, all facing the same 200 items, 13 pressure conditions, and four-turn conversations, with every reply labeled by two independent LLM judges. We find that the dominant factors are how costly it is for the model to verify the user's claim, and whether a trained guardrail covers it. Removing this task factor from a logistic model costs 0.485 of McFadden $R^2$, against 0.139 for model family and 0.009 for pressure tactic. Anchored facts are almost never conceded (1.3%), while adoption on logic puzzles rises with the number of clues needed to refute the pushed answer. Personal choices are endorsed in 77.0% of conversations. Most concessions on hard items come from models that cannot reliably solve them; models that can solve them rarely give the answer up. For both models tested, maximum reasoning removes these concessions completely: adoption on deep puzzles falls from 19.2% and 12.5% to 0%. Fallacious or emotional framing adds nothing beyond plain repetition. Three human annotators agree with the judges' consensus on 118/120 calibration items. These results give practical rules for reliable use: simplify hard-to-verify problems and reason deeply, state the question rather than one's preferred answer, ask for evidence on open questions, and choose models by their measured guardrail profile.

Language training relies on a corpus constructed by a large number linguistic materials. AI-powered voice clones provide a way to construct the corpus with relatively low cost. Singing voice conversion (SVC) model is used to generate synthetic voices. This study compares participants' performances on natural and synthetic speech in two experiments, similarity perception and intonation recognition. In the accuracy of similarity perception task, a significant interaction between speech type and intonation is found, suggesting that question may serve as a cue for speaker identification but may be influenced by synthetic features. In the accuracy of intonation recognition task, a significant interaction between speech type and familiarity is observed, indicating that speech type affects how much familiarity contributes to voice processing.

Activation steering manipulates large language model behavior by intervening on internal activations, but the mechanistic basis of these interventions remains poorly understood. We decompose refusal steering into component-level interventions across four open-weight models, identifying the sparse subsets of attention and MLP components whose steering suffices to reproduce the full behavioral effect. We find that refusal directions concentrate in sparse component mechanisms comprising 28–48% of upstream components, retaining 88–101% of steering effectiveness. Within these mechanisms, effective steering further concentrates in approximately 50% of residual stream dimensions, retaining 85–98% of the component-mechanism baseline, consistent with a privileged basis structure. Sparsity thus operates at two levels: which components are steered, and which dimensions within those components carry the signal. Together these findings show that refusal is not diffusely encoded across a transformer but assembled by a structured, identifiable mechanism, providing a foundation for mechanistic understanding of how refusal behaviors are represented and steered. To facilitate reproducibility, we release all code and raw experimental results in https://github.com/wang-research-lab/Refusal_Mechanisms.

Retrieval-augmented generation grounds language models in external context, but for long documents flat top-$k$ retrieval can cluster on a single region and miss complementary evidence. RAPTOR-style summary trees address this by recursively clustering chunks and using a language model to summarize each cluster at indexing time, then ranking summary nodes alongside raw chunks at query time. We show the main benefit of summary trees in long-document QA can come from navigation rather than the generated summary content. We introduce NavTree, a leaves-only retriever that builds a deterministic balanced segment tree over chunks (zero language-model calls at indexing) and uses the tree purely as a navigation scaffold: a hybrid lexical-and-dense frontier walk, anchored on top retrieved leaves, descends from the root and emits only leaf chunks to the reader. On a matched-cost evaluation against flat retrievers and an extractive re-implementation of RAPTOR, NavTree is the strongest matched-cost hierarchical retriever in our evaluated grid and ties the strongest flat baseline. On long-document multi-hop QA, it is the only hierarchical method that significantly beats BM25 on a class-vs-class basis, corroborated by a reader-free retrieval-recall check. A matched-reader replication of the published abstractive RAPTOR variant, given strong cluster summaries, still loses to NavTree at every multi-chunk budget, at zero indexing cost. The ranking carries across stronger and open-weight readers, a stronger encoder, and a full factorial that isolates leaves-only emission as the structural lever.

Large language models (LLMs) are increasingly used in binary decompilation to refine the C-like pseudocode produced by traditional rule-based decompilers. While this pseudocode is useful, it is a heuristic and lossy abstraction rather than a faithful copy of the source code. It often contains decompiler errors, especially for aggressively optimized binaries where critical low-level details are obscured. We present ASAP, an assembly-source aligned pseudocode refinement framework for binary decompilation. ASAP learns source-aligned assembly representations from paired source and binary functions using joint function-level and snippet-level contrastive alignment. A Q-Former then compresses chunk-level assembly features into a fixed number of assembly tokens that condition the decompilation LLM alongside the decompiler-produced pseudocode. During refinement, we use stochastic pseudocode masking and a relative assembly-advantage loss to reduce the model's tendency to ignore assembly features and rely only on pseudocode refining. On two decompilation benchmarks across multiple compiler optimization levels, ASAP improves the average re-execution rate from 64.7% to 71.9% and the average recompilation rate from 91.6% to 96.6% compared with the strongest baseline, offering both a new perspective and a practical solution to binary decompilation.

Reinforcement learning from verifiable rewards (RLVR) has produced large reasoning gains in language models, and verifiable video benchmarks make it applicable to causal-temporal video question answering. We study what RLVR teaches video-language models about time. We fine-tune four open models (Qwen3-VL-8B/4B, Qwen2.5-VL-7B, Gemma-3-12B) with group relative policy optimization under three data recipes: verified (synthetic CLEVRER questions with exact answer and event-order rewards), unverified (self-supervised pretext tasks over 43,751 real web videos), and a 1:1 mixture, plus a verified+real arm that adds 4,000 verifiable questions on real video. Each cell is evaluated in-domain and on out-of-domain real video (a NExT-QA temporal stress set and an MVBench subset), with frames in order, shuffled, and absent. (1) Verified training yields large in-domain gains that shrink as base competence grows (+14 to +19 points on weaker models; +6 on the strongest). (2) Much of the gain is non-visual: accuracy with no frames rises nearly as much as with frames. (3) Verified-only training can severely degrade out-of-domain accuracy with no sign during training: Qwen3-VL-8B loses 26.7 and 25.2 points on the two real-video sets, while the mixture never significantly degrades a model trained on it. Adding real verified questions removes that loss (-2.3 points, within noise of base) and keeps a +9.3 in-domain gain, so the cause is narrow synthetic-only data, not verification. (4) No recipe induces temporal-order grounding: across 41 evaluations the ordered-versus-shuffled gap is indistinguishable from zero in 39 and marginal in two, despite an event-order reward. Verifiable rewards improve benchmark accuracy without temporal understanding. Report no-frame controls, and mix in real video to guard against out-of-domain degradation.

We revisit an established Quantum Control objective, characterize its exact local metric geometry, and reposition it as a rigorous benchmark for Search and Optimization. The trap-free topology of the landscape, established two decades ago, guarantees unhindered optimal pathways, but global topology alone does not govern convergence speed. Although the so-called Quantum Sphere is defined by the fourth power of the control field, we show that its yield gap admits an exact representation as a nonnegative linear combination of $\sin^2$ terms. From this single identity, we derive a two-sided quadratic sandwich with ratio $4/\pi^2$ on a certified, worst-case quarter-wave region, proving the optimal peak nondegenerate modulo its affine gauge and well-conditioned. For white-box optimization, on a halved (eighth-wave) region, the objective satisfies a Polyak-Łojasiewicz inequality with an explicit constant, granting Gradient Descent a linear rate of $1 - 16/(\pi^4\kappa_r)$, where $\kappa_r$ is the condition number of the relevant Hessian block. For black-box optimization, we prove that the elitist single-parent Evolution Strategy converges linearly, almost surely and in expected hitting time, from a certified sublevel set, despite the landscape's continuous gauge symmetry and its lattice of periodic optima; the certified level is explicit whenever the landscape is trap-free, and near the optimum the strategy provably sees only the quadratic tangent of the peak. Systematic numerical simulations corroborate the theory. Finally, we show that additive input (actuator) noise reweights each term of the representation by a Debye-Waller factor without disrupting the landscape's structure, and outline open challenges: output (signal) noise, far-field plateau escape, angular-synchronization certificates, and another physically scalable benchmark (SHG with Glass).