Self-supervised speech models (S3Ms) encode a rich hierarchy of acoustic, prosodic, and phonetic information, yet when and where it emerges during pretraining remains unclear. We train frozen-encoder probes on every Transformer layer across pretraining checkpoints of two matched models: HuBERT-Base, whose stage 2 is a separately initialized model trained on new pseudo-label targets, and single-objective wav2vec 2.0-Base. Our five tasks span an acoustic-to-abstract gradient: absolute pitch (F0), speaker-relative pitch, lexical stress, phoneme boundaries, and 38-way phoneme identity. Acoustic-prosodic cues and phone boundaries are decodable from the earliest checkpoints and barely shift across layers, whereas phoneme identity is constructed during pretraining, re-localizing to upper-middle layers under HuBERT's stage-2 targets at a matched update count. Acquisition order therefore tracks abstraction from the acoustic surface, not the segmental/suprasegmental distinction, extending to prosody an account established for phonetic, lexical and syntactic structure; the no-switch wav2vec 2.0 shows no comparable re-localization.
EmbeddingGemma-300M is a small open embedder whose tokenizer represents every Tajik letter, and as released it is weak at Tajik retrieval: on our Tajik law test the right passage is in its top ten results for 49.3% of questions, on tax documents for 45.9%, and on the human-written Belebele Tajik set for 69.8%, against 98.6% on the Russian version of the same set. We present Saidzoda-Tajik-EmbeddingGemma-300M, the same model fully fine-tuned for one epoch on 235,140 question and passage pairs written by our own Tajik language model for laws, tax documents, Wikipedia and web text, with the same mined hard negatives, English and Russian replay, loss, global batch, learning rate and epoch as our earlier Saidzoda-Tajik-E5-Large (arXiv:2611.00592) and BGE-M3 models. After training, law reaches 92.4% and tax 91.7%; on Belebele, the one test written by people, recall@10 rises from 69.8% to 92.6% and the first result is right for 76.6% of questions instead of 51.7%. A checkpoint at step 100 of 466 already holds 88% of the law gain at recall@10. At recall@10 the result is no more than 2.8 points below the two 560M models trained on the same pairs on law, tax and Wikipedia, within the intervals, with 768-dimensional vectors and about half their parameters; at recall@1, for which we have no intervals, it stays 7.0 points behind the E5 model on Belebele. The cost is a drop of 3.0 and 2.8 points in Russian and English recall@1, larger than for the two earlier models. We also report that, according to our run notes, a bfloat16 test run produced a non-finite loss at step 3 which the training framework logged as zero, so training was done in fp32. Our generated test questions come from the same model as the training questions, so we treat those scores as an upper limit until questions written by Tajik speakers exist.
Large language models (LLMs) are increasingly asked to judge whether a news article is true, and their stated confidence is often taken at face value. We test what happens to that confidence when an article's writing style, but not its facts, is manipulated, as in recent \emph{style attacks} on fake-news detectors. Tracing how the prediction and its confidence form layer by layer inside the network, we study three model families (LLaMA-3-8B, Qwen2.5-7B, Mistral-7B) on three fake-news benchmarks, and find the same three patterns in every model--dataset pairing. First, on the original articles the models are competent and largely honest judges: they settle on an answer, briefly become overconfident, and then correct their confidence in the final layers. Second, style-rewritten articles break this self-correction: the models become far less accurate, under the strongest attack, worse than random guessing, yet remain just as confident, so they are \emph{confidently wrong} precisely when a warning is most needed. Third, nearly all of the damage comes from rewriting into sensational or emotionally charged styles, while sober, neutral rewrites leave the models' honesty largely intact. The practical implication is that an LLM's confidence about veracity cannot be trusted as a safety signal when an article's style may have been manipulated.
Asynchronous pipeline parallelism can accelerate distributed model training, but suffers from instability caused by inconsistent gradient computations. We identify a fundamental source of this inconsistency: error signals are coupled to forward-time features, not just parameter states. This means that parameters may evolve as long as they preserve forward-equivalence. Leveraging this insight, we introduce FRESCO, which frames asynchronous updates as a constrained optimization problem: it finds the minimal update modification that acts trivially on the activation subspace. We demonstrate that FRESCO consistently outperforms representative asynchronous baselines in LLM pre-training, with the gains most pronounced under deep pipelines and high inconsistency levels: at pipeline depth 32, it reaches the same loss as the strongest baseline with 43% fewer iterations and 30% less wall-clock time, while requiring 8~921$\times$ less memory overhead. Moving beyond the current limitations, FRESCO establishes a new foundation for high-utilization pipeline parallel training, providing a scalable path for large-model development on diverse, resource-constrained systems.
Frontline sales staff spend much of their day re-coordinating quotes after customers change quantities, ask for discounts or new payment terms, or costs move. We present LoMASA (Local Multi-Agent Sales Assistant), a training-free multi-agent system that runs entirely on one desktop computer. A 35B mixture-of-experts model (System-1) makes each decision in a single forward pass and reads its confidence from the probabilities of the answer options. A 27B dense model (System-2) writes customer replies and escalation briefs and is consulted when System-1 is unsure. Code computes every price, margin and option, and hard pricing rules cannot be overridden. On a test split of 500 simulated episodes with 1512 decisions built from public supply-chain data, the salesperson has to decide only 19.5% of decisions: 38.0% are settled automatically, with a median latency of 2.67 seconds, and 42.5% are escalated to a manager by the hard rules. On the automatically settled decisions LoMASA is as accurate as a rule baseline (0.894 vs.\ 0.894) and makes fewer commitments whose margin later falls below the floor (2.61% vs 4.53%); the price is additional escalations to the manager. Option probabilities are well calibrated without any training, unlike confidence stated in words by the same model. We also report where the design does not help: the 27B model rarely changes a decision.
Direct Preference Optimization (DPO) is a simple and widely used alternative to RLHF for aligning large language models (LLMs), but its sequence-level objective cannot precisely locate or suppress the local patterns that make an otherwise acceptable response unsafe or undesirable. We propose TLRS-DPO (Token-Level Risk Scoring for Better Direct Preference Optimization), which augments DPO with fine-grained token-level control while remaining fully offline and reference-based. TLRS-DPO introduces a token-level robust compensation term that converts the policy-reference log-ratio on rejected responses into localized risk scores, a curriculum-style schedule that gradually increases the strength of this compensation during training, and a risk-adaptive screening mechanism that selectively penalizes only high-risk tokens instead of entire sequences. This design bridges sequence-level preference fitting and token-level behavior control, and can be implemented as a drop-in modification to standard DPO trainers. Empirically, TLRS-DPO improves alignment quality and safety-relevant behavior with minimal impact on helpfulness compared to strong DPO baselines.
Machine Learning Interatomic Potentials (MLIPs) based on Equivariant Graph Neural Networks (EGNNs) have achieved remarkable progress in atomistic modeling. However, accurately representing atomic interactions across different spatial scales remains challenging. In this work, we investigate the radial spectral properties of SO(3)-equivariant message passing and establish that the cutoff radius controls the characteristic spatial-frequency scale of radial interactions. Based on this analysis, we propose Multi-Cutoff Spectral Aggregation (MCSA), a plug-and-play module that aggregates radial kernels with different cutoff radii to enrich multi-scale geometric representations. To achieve efficient implementation, we exploit the compact support of radial kernels to fuse cutoff-dependent radial weights before angular tensor-product operations, avoiding redundant angular computations while preserving rotational equivariance. Extensive experiments across representative EGNN backbones demonstrate consistent improvements in energy, force, and stress prediction, as well as enhanced accuracy in higher-order properties, including phonon dispersion and lattice thermal conductivity. Further evaluations on layered materials with distinct interaction scales demonstrate the effectiveness of physically motivated cutoff selection. Code is available at: https://github.com/JF-Tan/MCSA.
Post-hoc explanation methods for neural networks are often heuristic and lack provable guarantees. A common alternative to such methods that was studied in recent years is to compute a cardinally-minimal subset of input features that is provably sufficient to determine the model's prediction. For general neural networks, however, finding such explanations is computationally intractable. In contrast, recent work showed that this problem becomes more tractable for a particular family of neural networks, namely neural additive models (NAMs), though standard NAMs are highly restrictive because they exclude feature interactions altogether. In this work, we study the computation of provably minimal and sufficient explanations for neural networks with restricted feature interactions, bridging the gap between fully general neural networks and purely additive models. We first prove that even with constant-order interactions, the problem is $\Sigma_2^P$-hard, implying that worst-case exponential complexity is unavoidable. We then identify two provably tractable settings: one based on sparsity in the interaction structure and another based on a stronger interaction-aware notion of sufficiency. Under either of these settings, we develop algorithms that compute provable explanations for neural networks with any-order interactions while retaining much of the computational efficiency of additive models. Empirically, we show that these explanations are significantly smaller and faster to obtain than those produced by standard algorithms, while being derived from models that achieve much higher accuracy than standard NAMs. Overall, our results help expand the class of neural networks that admit efficient provable explanations, while clarifying the fundamental role of feature interactions in shaping their complexity.
Activation sparsity can reduce FFN weight traffic during single-user LLM decoding, but active channels must be identified early enough to avoid fetching unused weights. In pretrained SwiGLU models, training-free routing with the native gate performs substantially worse than routing with the up activation at high sparsity. We show that this ordering reverses after sparsity-aware training. Based on this insight, we introduce SGA-Sparse, which converts pretrained SwiGLU LLMs into fixed-budget activation-sparse models by training the existing gate projection directly as a top-k router. SGA-Sparse updates all model parameters end-to-end while preserving the original SwiGLU activation and introducing no auxiliary router parameters. After approximately 30B training tokens, LLaMA models from 1B to 8B achieve uniform per-token 75-85% FFN channel sparsity while remaining within 0.2-2.7 pp of their original dense checkpoints in average zero-shot accuracy. Because post-training quantization degrades sparsity-trained checkpoints, we further study two complementary quantization-aware training (QAT) recipes: continued-QAT largely recovers performance at 4-bit weight quantization, while joint-QAT performs better for the more aggressive sub-4-bit target. To compare activation sparsity with dense low-bit quantization, we translate the sparsity-induced reduction in weight traffic into an equivalent bit-width. Under this unified weight-traffic metric, our 2.5- and 1.5-bit-equivalent models establish a new accuracy-traffic Pareto frontier: the 2.5-bit-equivalent models match dense 3-bit QAT accuracy, while the 1.5-bit-equivalent models approach dense 2-bit models and outperform dense ternary models by 1.3 pp on average. These results establish native-gate training as a practical approach for combining high FFN sparsity with low-bit weights.
MeanFlow learns interval-average velocity for few-step generation, but its fixed temporal sampler mixes diagonal and interior supervision throughout training. We interpret temporal coordinates as collocation points: diagonal samples provide a stochastic boundary target, whereas finite intervals supply a differential-consistency target. The diagonal fraction controls how often the consistency component appears, while interval length enters its gradient through an explicit factor. Motivated by this distinction, EbbFlow uses Temporal Supervision Allocation (TSA) to schedule the diagonal fraction and interval geometry. It starts with diagonal supervision, then admits interior samples and expands their maximum interval; a left-anchored interval law preserves coverage across time. The per-sample loss, target, backbone, and inference procedure remain unchanged. In matched class-conditional ImageNet $256\times 256$ runs without classifier-free guidance (DiT-B/2, one training seed per method, 400k updates), EbbFlow achieves 50k-sample FID of 39.65 at one network evaluation and 36.02 at two-evaluation ODE sampling, versus 44.82 and 39.86 for MeanFlow. Gradient probes show a mid-training difference in component geometry. Ablations support allocation scheduling and left-anchored intervals, although a fixed full-horizon variant performs better at one evaluation. These results suggest that temporal sampling is a useful training control whose benefits depend on the inference protocol.
The convergence of decentralized architectures integrating Machine Learning, Computer Vision and Low Power Wide Area Networks is increasingly becoming an integral part of our daily existence. Internet of Things serves as a real-time data conduit enhancing decision making via embedded technology and continuous data exchange. This paper explores the feasibility of Edge Computing as a foundational pillar in this evolving landscape. We experiment under real world, dynamic conditions, evaluate the technological aspects, strategies, process flows and key observations under the broad Edge Computing domain. Research pathways include Multi-access Edge topologies in future 6G networks, model quantization, and satellite-enhanced communication platforms. Additionally, a discussion is added supporting the advanced AI functionalities, including zero-shot learning, multi modal perception, and decentralized generative AI, thereby broadening the scope of intelligent applications across various domains. The significance and research objective of this study are threefold: (1) evaluation of LoRaWAN and satellite IoT communication strategies, (2) analysis of CV workloads on edge hardware and (3) future research directions where Edge Computing can support low-latency, energy-efficient and socially impactful IoT applications. By explicitly addressing these aspects, we aim to establish a clear link between the technological feasibility, ultimately with a practical and socioeconomic relevance.
Predicting single-cell transcriptional responses to genetic and chemical interventions supports gene-function discovery and regulatory analysis. Existing studies often focus only on highly variable genes, restricting the expression changes available for downstream analysis. Cells exposed to the same intervention can also show heterogeneous expression responses, with coordinated gene programs active to different degrees. We propose scBIG++, a full-gene perturbation prediction framework that models this heterogeneity by connecting expression-adaptive gene modules with population-conditioned generation. To preserve shared gene structure while accommodating cellular variation, Gene-Relation Clustering organizes genes into common windows, within which Dynamic SoftSub learns expression-dependent submodule memberships. This shared organization enables Set-GCAE to exchange information about corresponding gene groups across control cells, placing individual representations in their source-population context. The cell representations initialize a latent flow, while their pooled summary and the perturbation representation condition response generation. Generated states are decoded into expression profiles covering all assayed protein-coding genes retained after preprocessing, supporting comprehensive analysis of perturbation-induced expression changes. Experiments across six genetic and chemical settings spanning 6,539--19,451 genes demonstrate competitive perturbation-effect recovery and full-gene expression reconstruction under unseen targets and combinations, cross-cell-line transfer, and held-out compounds. Our implementation is available at https://github.com/ttruan2426-dot/scBIG.
Reproducibility in large language model (LLM) agent evaluation depends on replay fidelity: which parts of an execution trace are reconstructed, controlled, and compared. This matters for agent debugging, failure analysis, and counterfactual evaluation of stateful tool use. We audit all 71 unsuccessful executions in a fixed historical AppWorld corpus using 284 isolated fixed-code replay attempts. Reconstructed declared settings admit 42 cases under printed-output, pre-oracle database-delta, and native-verdict/test-count checks; resetting the random-number generator before each interaction admits 68. The 26-case gap contracts to four under database/oracle checks, with one endpoint unknown. Mechanism studies identify four materialized-state discrepancies despite matching outputs and verdicts. In one admitted endpoint pair, an identical API call yields matching responses but different row effects; common future Python RNG removes that deterministic difference in two paired repetitions. A separate JWT credential-equality program and alignment controls expose authentication/write differences, with missing repetitions retained. For all 10 affected post-oracle export cases, native export from retained output trackers restores the reference comparison without changing original cached contents. A purposeful 12-case BFCL pilot shows no discrepancy under its tested public-state, return, and verdict checks. We preserve resource censoring and infrastructure deviations. The contribution is a bounded empirical characterization of replay eligibility, hidden state, and test-oracle integration, without estimating debugger accuracy, repair gains, or method-ranking distortion.
Post-hoc explanation methods for video action recognition typically conflate spatial appearance and temporal dynamics into a single saliency or importance map. We propose a perturbation-based framework that disentangles these contributions into a spatial appearance mask ($M_a$) and a temporal motion mask ($M_m$). The two masks are jointly optimized against a frozen action recognition model through factored perturbation architecture, with nearest frame replacement to identify important temporal dynamics. A separation penalty discourages the appearance mask from spending budget on frames that the motion mask has already discarded. A ReLU-based area constraint penalizes explanations larger than an assigned budget. To evaluate these explanations, we introduce a separation protocol that sweeps spatial pixels and temporal frames independently, backed by proofs characterizing the proposed method. Across EPIC-Kitchens-55, EGTEA Gaze+, and Something-Something V2, our method attains the lowest deletion AUC against leading baselines while producing masks roughly five times sparser.
Historical social simulation faces a fundamental reproducibility challenge: the stochastic nature of large language models (LLMs) introduces a digital butterfly effect that prevents exact replication across multi-turn runs. Furthermore, the inability to control low-level server parameters in commercial LLM hosting environments severely limits reproducibility. To overcome this opacity, we present a deterministic multi-agent framework coupling LLMs to an Agent-Based Modelling (ABM) pipeline grounded in Theory of Mind (ToM) and Cognitive Archaeology, and validate exact artifact identity within local, pinned execution environments using SHA-256 hashing. Run locally at temperature zero with clean per-call model reloads, the system injects differentiated cognitive states per turn through dissonance-driven behavioural modes and gated internal voices. Evaluating seven model configurations spanning 4B to 31B parameters against an ablated baseline reveals clear operational boundaries. Small models ($\leq$8B) suffer from context capacity overload, producing repetitive, specular dialogue loops. However, at $\geq$12B parameters, the ABM scaffolding breaks the monotonous response patterns seen in ablation. Crucially, these scaffolded models exhibit spontaneous diegetic expansion: they coherently expand scenario lore through tactically motivated, anachronism-free details rather than escalating into aggression. By resisting the structural pull toward premature scene closure, scaffolded sessions sustain an open terminal Narrative Vector ($V_N = \mathrm{Open}$ in 12/21 overall; 9/9 in official $\geq$12B configurations) whereas unscaffolded ablation never does (0/21). This framework provides a transparent, auditable benchmark architecture for virtual archaeology, with the full pipeline released as open-source software to enable independent replication.
Accurate lighting response is essential for realistic dynamic human relighting, because shadows, highlights, and color changes should follow the target illumination in both direction and relative strength. Current learning-based relighting methods can generate visually plausible videos, but they often exhibit imprecise lighting responses: shadows and highlights may drift away from the target direction, and the output brightness may not faithfully follow the relative intensities in the target environment, making the result closer to appearance harmonization than controllable relighting. To address this issue, we propose DynOLAT, a generalizable OLAT video prediction framework for dynamic full-body humans. Instead of directly synthesizing a single relit appearance, our method predicts temporally consistent One-Light-At-a-Time (OLAT) basis videos from flat-lit inputs, and then linearly combines the predicted bases according to the target environment map. This OLAT-based formulation makes the directional response explicit in the predicted bases, while the relative lighting strength is determined by the linear weights derived from the environment map. We further introduce a dynamic capture pipeline and the first large-scale dynamic full-body OLAT dataset, comprising over 24 million frames from 224 identity-outfit combinations, where different outfits of the same person define separate combinations. Experiments show that DynOLAT responds more faithfully to variations in target lighting direction and relative intensity than the tested baselines, enabling controllable video relighting while preserving fine appearance details and temporal coherence. The dataset and processing tools will be publicly released to foster future research.
Visual chain-of-thought (visual-CoT), also known as "thinking with images", equips LVLMs with the ability to locate and use key visual evidence for reasoning. However, acquiring such visual-CoT capabilities requires costly post-training and tightly couples them to a specific backbone, making transfer to other LVLMs difficult. Existing cross-model transfer methods typically treat visual-CoT as a monolithic capability, using a unified representation that ignores the functional distinction between grounding and reasoning and introduces stage-irrelevant interventions. To address this problem, we propose Viscot-steer, a training-free activation steering framework for transferring visual-CoT capabilities across heterogeneous LVLMs. Specifically, Viscot-steer first disentangles visual-CoT into grounding and reasoning vectors and extracts them from the source-model layers associated with the corresponding capabilities. It then aligns the vectors with the activation space of the target model and steers the grounding and reasoning layers during bounding-box prediction and reasoning-trace generation, respectively. Across benchmarks covering high-resolution visual understanding, general visual perception and reasoning, and mathematical visual reasoning, Viscot-steer transfers capabilities from different visual-CoT models to architecturally diverse LVLMs and outperforms the strongest capability-transfer baseline by an average of 1.86 percentage points on the primary evaluation metrics.
On the Web, user profiling builds compact user representations from interaction histories for efficient personalization, and dynamic user profiling updates them as users' behavior and preferences change. However, new user behavior can be ambiguous, and its interpretation differs across users. To address this, we introduce Uncertainty-aware Adaptive Profile Manager (UAPM), which separates the claims in a user profile from interpretations of new behavior that are not yet confirmed, through a Profile, a Buffer, and a Rulebook. UAPM records user-specific interpretation and update rules in the Rulebook, keeps ambiguous evidence in the Buffer outside the Profile as evidence for later decisions, and writes supported general and situational characteristics to the Profile. We train UAPM with Privileged On-Policy Distillation with Dual Specialization (POPD-DS), which learns from recorded future interactions during training without access to them at deployment. Its first stage, Dual Specialization, distills a stronger model into a current-only Student and a future-informed model that becomes the Teacher. In the second stage, privileged on-policy distillation (POPD), only the fixed Teacher receives future interactions and guides the Student's own management decisions alongside task rewards. Experiments on three Web personalization tasks show that UAPM outperforms multiple baselines under the same downstream input budget, improving on the best same-budget baseline by up to 5%.
We study prompted chain-of-thought (CoT) with self-consistency (SC), i.e. sampling N traces and plurality-voting, on small language models, a regime where per-query compute budgets are tight yet SC is commonly applied with a fixed sample budget inherited from large-model practice, regardless of instance difficulty. Across two model families (Qwen-2.5-0.5B, Gemma-3-1B) at FP16 and MLX-Q4 (4-bit quantisation in Apple's MLX framework), on GSM8K and MATH-500, we run 8 per-task and 4 task-pooled experiments at n=200-400 with 1,000x paired-bootstrap confidence intervals (CIs). Our first result is an upper bound: an oracle gate deciding per prompt whether to invoke SC reaches 109-174% of always-SC's accuracy at 23-30% of its compute in all 12/12 cases, so at N=5 always-SC is dominated everywhere on our grid. That margin is exactly the rate at which SC flips correct greedy answers to wrong, here 3-10.5% of prompts, which on Qwen-Q4 x MATH-500 exceeds the wrong-to-right rate and leaves always-SC 3 pt worse than always-greedy. Our second result is a deployable approximation: a gate over greedy-pass signals skips 45-99% of SC calls, with bootstrap CI half-widths under 5 pt, removing 36-79% of always-SC compute. Accuracy retention at those operating points has point estimates of 98.1-98.8%, which we report as point estimates because their CIs are wide (lower bounds 76-86%) over an SC accuracy of only 15-36%. The evidence covers N=5 on two <=1B families and two math benchmarks, which is the scope we claim. Code is available at https://github.com/heywanrong/SiftSC.
In the extreme-low bitrate range below $0.01$ bpp, image compression becomes a constrained generation problem: the reconstruction depends on inferring plausible missing details while staying close to the input image. Autoregressive token models provide an efficient coarse-to-fine prior for this setting, but the transmitted early-scale tokens often leave object identity and texture underdetermined, which can lead to source-inconsistent reconstructions. We introduce Source-Anchored Visual Autoregression (SAVAR), a generative codec built on Visual Autoregressive (VAR) modeling. SAVAR transmits early-scale tokens and a compact caption. The tokens provide image-specific layout information, while the caption supplies global semantics. SAVAR reconstructs a partial VQ-VAE latent from the retained scales, maps it to the start-of-sequence initialization, and uses the same latent to condition the caption embedding before adaptive layer normalization, separating two roles that are typically driven by a shared condition in VAR-style generation. On the Kodak and CLIC2020 datasets in the extreme-low range, SAVAR reduces perceptual distortion relative to matched ARPC retained-scale operating points, lowering LPIPS by about $8$--$9\%$ and DISTS by up to $22\%$ at a lower bpp. It reconstructs a $512 \times 512$ image in about $218$ ms on one NVIDIA L20X GPU, retaining fast autoregressive decoding in the evaluated setting.