arXiv:2606.23129v2 Announce Type: replace Abstract: Implicit Neural Representations (INRs) have been proven successful in encoding continuous signals through coordinate-based networks, yet facing a spectral dilemma: periodic activations capture fine details but act as all-pass filters that memorise noise, while spatially compact activations regularise effectively but suffer from low-frequency bias. Existing attempts to resolve this trade-off introduce computational overhead or tuning frailty. We propose to model each neuron's activation as the steady-state response of a sinusoidally-forced damped harmonic oscillator, whose amplitude naturally governs the network's spectral selectivity during training. By jointly optimising the oscillator parameters alongside the network weights, our method adapts to the target signal's spectral content without explicit regularisation. Initialised in the stopband, the network exhibits a coarse-to-fine learning curriculum that progressively expands its spectral gate, capturing low-frequency structures first and high-frequency details only when justified by the reconstruction objective. Comprehensive experiments show that our approach consistently achieves state-of-the-art or competitive results against established INRs, while requiring no task-specific tuning of any hyperparameters.
Science Journals
arXiv:2606.23227v2 Announce Type: replace Abstract: The matrix exponential and the closely related matrix phi-functions play a fundamental role in the solution of first-order systems of ordinary differential equations (ODEs). In particular, they appear in exact solutions of certain linear ODE systems, and in exponential integrators, a class of explicit time discretisation methods for computing approximate solutions of stiff semi-linear ODE systems. Fundamental properties of the matrix exponential include that it is nonnegative if the matrix is essentially nonnegative and stochastic if the matrix is a transition-rate matrix. In this paper, we study related properties for the matrix phi-functions. Using these properties, we then provide insights into deterministic solutions of linear and semi-linear ODE systems and outline how such solutions can be used to generate stochastic simulations to quantify variability in model predictions. The paper concludes with some illustrative examples exploring the theory for some ODE systems arising from spatial discretisation of PDEs of diffusion, advection-diffusion, Fisher-KPP and Allen-Cahn type.
arXiv:2603.24049v2 Announce Type: replace Abstract: The sensitivity of Ramsey interferometry experiments is governed by the interplay between the beam phase-space distribution and the magnetic field environment through which the spins propagate. Quantitative optimisation thus requires a consistent treatment of optics, magnetics and spin dynamics. We present a simulation framework that enables such an analysis by combining neutron optics simulations in McStas, magnetic field modelling in COMSOL and spin-dynamics simulation in the new RamseyProp program. We describe how important experimental parameters such as adiabaticity, flip angle distributions and Ramsey fringe contrast can be studied. The code is being applied to design an experiment to search for axion-like particles at the European Spallation Source (ESS). We examine how the pulsed time structure of the ESS can be exploited to perform Ramsey interferometry on a broad neutron velocity spectrum. In the absence of velocity or timing restrictions, the standard deviation of the spin flip angle at zero detuning can be reduced from 0.67 to 0.17 radians using time-dependent amplitude modulation. Similarly, the phase sensitivity can be improved by a factor of 4 for a 10 m long setup starting 15 m from the ESS moderator.
arXiv:2603.17621v2 Announce Type: replace Abstract: Reinforcement Learning (RL) has emerged as a powerful paradigm for training LLM-based agents, yet remains limited by low sample efficiency, stemming not only from sparse outcome feedback but also from the agent's inability to leverage prior experience across episodes. While augmenting agents with historical experience offers a promising remedy, existing approaches suffer from a critical weakness: the experience distilled from history is either stored statically or fail to coevolve with the improving actor, causing a progressive misalignment between the experience and the actor's evolving capability that diminishes its utility over the course of training. Inspired by complementary learning systems in neuroscience, we present Complementary RL to achieve seamless co-evolution of an experience extractor and a policy actor within the RL optimization loop. Specifically, the actor is optimized via sparse outcome-based rewards, while the experience extractor is optimized according to whether its distilled experiences demonstrably contribute to the actor's success, thereby evolving its experience management strategy in lockstep with the actor's growing capabilities. Empirically, Complementary RL outperforms outcome-based agentic RL baselines that do not learn from experience, achieving 10% performance improvement in single-task scenarios and exhibits robust scalability in multi-task settings. These results establish Complementary RL as a paradigm for efficient experience-driven agent learning.
arXiv:2603.17863v2 Announce Type: replace Abstract: Automating the development of machine learning algorithms has the potential to unlock new breakthroughs. However, our ability to improve and evaluate algorithm discovery systems has thus far been limited by existing task suites. They suffer from many issues, such as: poor evaluation methodologies; data contamination; and containing saturated or very similar problems. Here, we introduce DiscoGen, a procedural generator of algorithm discovery tasks for machine learning, such as developing optimisers for reinforcement learning or loss functions for image classification. Motivated by the success of procedural generation in reinforcement learning, DiscoGen spans billions of tasks of varying difficulty and complexity from a range of machine learning fields. These tasks are specified by a small number of configuration parameters and can be used to optimise algorithm discovery agents (ADAs). We present DiscoBench, a fixed, small subset of DiscoGen tasks for principled evaluation of ADAs. Finally, we propose a number of ambitious, impactful research directions enabled by DiscoGen, and demonstrate its use for ADA optimisation through scaling experiments for automated prompt tuning. DiscoGen is released open-source at https://github.com/AlexGoldie/discogen.
arXiv:2603.17975v2 Announce Type: replace Abstract: We present AHOY, a method for reconstructing complete, animatable 3D Gaussian avatars from in-the-wild monocular video despite heavy occlusion. Existing methods assume unoccluded input-a fully visible subject, often in a canonical pose-excluding the vast majority of real-world footage where people are routinely occluded by furniture, objects, or other people. Reconstructing from such footage poses fundamental challenges: large body regions may never be observed, and multi-view supervision per pose is unavailable. We address these challenges with four contributions: (i) a hallucination-as-supervision pipeline that uses identity-finetuned diffusion models to generate dense supervision for previously unobserved body regions; (ii) a two-stage canonical-to-pose-dependent architecture that bootstraps from sparse observations to full pose-dependent Gaussian maps; (iii) a map-pose/LBS-pose decoupling that absorbs multi-view inconsistencies from the generated data; (iv) a head/body split supervision strategy that preserves facial identity. We evaluate on YouTube videos and on multi-view capture data with significant occlusion and demonstrate state-of-the-art reconstruction quality. We also demonstrate that the resulting avatars are robust enough to be animated with novel poses and composited into 3DGS scenes captured using cell-phone video. Our project page is available at https://miraymen.github.io/ahoy/
arXiv:2603.21526v2 Announce Type: replace Abstract: Multimodal large language models (MLLMs) offer a promising path toward interpretable deepfake detection by generating textual explanations. However, the reasoning process of current MLLM-based methods combines evidence generation and manipulation localization into a unified step. This combination blurs the boundary between faithful observations and hallucinated explanations, leading to unreliable conclusions. Building on this, we present VIGIL, a part-centric structured forensic framework inspired by expert forensic practice through a plan-then-examine pipeline: the model first plans which facial parts warrant inspection based on global visual cues, then examines each part with independently sourced forensic evidence. A stage-gated injection mechanism delivers part-level forensic evidence only during examination, ensuring that part selection remains driven by the model's own perception rather than biased by external signals. We further propose a progressive three-stage training paradigm whose reinforcement learning stage employs part-aware rewards to enforce anatomical validity and evidence--conclusion coherence. To enable rigorous generalizability evaluation, we construct OmniFake, a hierarchical 5-Level benchmark where the model, trained on only three foundational generators, is progressively tested up to in-the-wild social-media data. Extensive experiments on OmniFake and cross-dataset evaluations demonstrate that VIGIL consistently outperforms both expert detectors and concurrent MLLM-based methods across all generalizability levels.
arXiv:2606.23452v2 Announce Type: replace Abstract: Large loads from industrial electrification and data centers are reshaping the planning and operation of the power grid. Identifying optimal large load siting decisions while accounting for transmission congestion is key to reducing expansion cost and operational risks. In this paper, we propose a leader-follower bilevel optimization framework to identify optimal large load allocation strategies. The leader determines the allocation of large loads, while the followers determine grid expansion cost and transmission utilization. This modeling approach explicitly integrates strategic planning with detailed short-term operational decisions. Moreover, we develop a Bayesian Optimization approach to efficiently solve the bilevel optimization problem by treating the followers as a black box. We use the framework to study large-scale load allocation from electrified oil refineries and data centers on a synthetic power grid that resembles key characteristics of the Texas (ERCOT) system. The results show that these large loads compete for electricity, and under high-load scenarios, data center demand is distributed across the entire grid, avoiding regions with high demand from industrial electrification.
arXiv:2606.23671v2 Announce Type: replace Abstract: Prior work shows that large language models (LLMs) exhibit introspective capability on benign tasks. We extend the question to safety contexts and examine how reliably a model can recognize that its own prior response was elicited by an adversarial prefill attack. Across ten open-weight instruction-tuned LLMs (3B to 70B) and four safety benchmarks, no model reliably recognizes its own compromised outputs, with models claiming intent on prefilled responses at an average rate of $27.3\%$. Introspective signal stems largely from safety- and refusal-related reasoning. Orthogonalizing models' weights against the refusal direction collapses the gap between claiming rates on prefilled and natural outputs to near zero, though the direction is not its unique mediator. The signal is also probe-dependent: framing the question as internal intention versus external tampering elicits qualitatively different responses on the same models. Training models to mimic correct introspective answers or pursue an introspective objective can improve the accuracy of introspection, but such training does not transfer to the tampering probe and counterintuitively raises attack success rate under adversarial prefill on most models, amounting to a partial mitigation. These findings outline mechanisms underpinning the observed introspective signals in safety contexts and highlight risks in the reliability of LLM self-reports.
arXiv:2606.23852v2 Announce Type: replace Abstract: We develop general strategies for transferring soundness and completeness from more expressive modal languages to less expressive ones, unifying several existing notions of operator definability along the way. For soundness, we exploit semantic insensitivity: if a less expressive language is insensitive to a frame operation, soundness extends to the operation's closure of the original frame class. For completeness, restricting to relational semantics and languages with a single operator, we present strategies for relating the target logic's canonical model to that of a normal modal logic via a truth-preserving translation. Three of those dispense entirely with specifying an accessibility condition for the target logic, inheriting it from a normal modal logic instead.
arXiv:2606.24004v2 Announce Type: replace Abstract: Steering a large language model (LLM) toward a desired behavior typically relies on an iterative process of hand-crafting a prompt based on a careful inspection of the model's responses. This is an involved, brittle, and error-prone process. Preference-based fine-tuning is a more rigorous but often prohibitively expensive solution. We propose spec learning, a framework that relies on a brief user instruction and a small set of preference judgments. These are compiled into specifications in the form of natural-language prompts for an LLM. Specifications condition LLMs at inference time, and no parameter updates to the underlying models are required. We show that the responses generated based on the compiled specifications often outperform direct preference optimization (DPO) on datasets from specialized domains whose preference signal is dense. Unlike opaque weight updates, the resulting specifications are human-readable and double as interpretable and transparent written embodiments of the preference signal that produced them.
arXiv:2606.24186v2 Announce Type: replace Abstract: This paper revisits the {\em E scheme} of Osher \cite{Osher-SINUM1984} and the {\em flux-limiter scheme} of Sweby for quasi-linear hyperbolic conservation laws \cite{Sweby-SINUM1984}. Part of existing results will be re-understood and some new results will be presented. For a scalar conservation law, except for the conservative monotone schemes, the E scheme is a type of numerical methods that satisfy the discrete entropy condition for any convex entropy, but numerical entropy flux is not unique. Two-point monotone flux is E flux, but conversely it may not necessarily be correct. Moreover, multi-point (three or more points) E flux may not necessarily be monotone flux, and multi-point monotone flux may not necessarily be E flux. Sweby's flux-limiter scheme for the quasi-linear conservation laws was built on the E flux-based splitting $f_{j+1}-f_j=f_{j+1} { -\hat{f}^{\text{\tiny E}}_{j+\frac12}+\hat{f}^{\text{\tiny E}}_{j+\frac12}}-f_j$ and the LW scheme. It may not be second-order accurate in both space and time.
arXiv:2409.03576v4 Announce Type: replace Abstract: This article is a continuation of our recent work (Yin Chen and Runxuan Zhang, Shape enumerators of self-dual NRT codes over finite fields. SIAM J. Discrete Math. 38 (2024), no. 4, 2841-2854) in the setting of quantum error-correcting codes. We use algebraic invariant theory to study three weight enumerators of formally self-dual quantum codes over arbitrary finite fields. We derive a quantum analogue of Gleason's theorem, demonstrating that the weight enumerator of a formally self-dual quantum code can be expressed algebraically by two polynomials. We also show that the double weight enumerator of a formally self-dual quantum code can be expressed algebraically by five polynomials. We explicitly compute the complete weight enumerators of some special self-dual quantum codes. Our approach illustrates the potential of employing algebraic invariant theory to compute weight enumerators of self-dual quantum codes.
arXiv:2606.24257v2 Announce Type: replace Abstract: High-quality 3D vehicle assets are essential for autonomous driving simulation. Although multi-view diffusion-based paradigms enable controllable single-image reconstruction, they typically produce limited viewpoints and exhibit cross-view geometric inconsistencies, thereby reducing reconstruction fidelity in real-world scenarios. In this work, we introduce 3DCarGen, a scalable single-view 3D car generation framework designed for real-world images by synthesizing an arbitrary number of 3D-consistent multi-view images. Specifically, given a single image as input, we first synthesize a set of images from fixed viewpoints. These images are then fed into a feed-forward reconstruction model, resulting in a coarse 3D representation based on 3D Gaussian Splatting. Conditioned on this explicit 3D prior, our multi-view diffusion model generates 3D-consistent images from arbitrary camera viewpoints. We further extend a fast mesh reconstruction algorithm by incorporating color-normal joint optimization to recover detailed and coherent 3D vehicle models from the synthesized dense views. Extensive experiments on synthetic and real-world datasets demonstrate that our approach achieves robust geometric consistency and reconstruction fidelity compared to existing methods. Project page: https://honglixiao.github.io/3dcargen.github.io/.
arXiv:2606.29400v1 Announce Type: new Abstract: In computer graphics, visual content is continuously warped, zoomed and resampled. This occurs when engines upscale frames, users zoom into 3D scenes, or foveated VR applies varying scaling. Handling these transformations requires Arbitrary-Scale Super-Resolution (ASR). Traditional models, designed for fixed scales, typically predict at a lower integer scale (e.g., x4) and rely on sub-optimal interpolation for continuous resolutions, compromising quality. Furthermore, most methods process pixels uniformly. Since fine details are sparse, this creates overhead; efficiency dictates concentrating resources only where structural complexity demands it. While implicit models and Gaussian Splatting (GS) enable continuous representation, GS is advantageous due to adaptive densification. However, transitioning GS into a feed-forward model for ASR is non-trivial. Standard GS optimization needs high-resolution gradients to drive primitive growth, which are unavailable during inference. Thus, the network must autonomously predict GS densification from low-resolution inputs. To solve this, we propose QuADA-GS. After encoding inputs into a latent space, a Neural Routing Architecture evaluates local complexity to distribute a global budget, assigning specific upsampling factors to features to avoid redundant processing. Features are dynamically densified based on these factors, forming an irregular topology decoded into 2D Gaussian primitives. To coordinate features before decoding, we introduce Hierarchical Pointer Convolution. This non-grid operator achieves O(1) neighbor lookup complexity, facilitating efficient spatial communication and bypassing dense bottlenecks. Experiments show QuADA-GS achieves state-of-the-art ASR performance, maintaining low latency and a lean memory footprint.
arXiv:2606.24466v2 Announce Type: replace Abstract: Legged manipulators combine the mobility of legged platforms with the manipulation capability of robotic arms. However, arm-induced Center-of-Mass shifts and dynamic disturbances make the system more prone to instability under actuator failures, potentially leading to falls, task failures, or safety risks. Existing fault-tolerant control methods mainly focus on locomotion alone, leaving the coupled problem of whole-body stability and arm reachability in fault-tolerant loco-manipulation largely unaddressed. To bridge this gap, we propose FT-WBC, a fault-tolerant loco-manipulation framework for robust whole-body control of legged manipulators under actuator failures. FT-WBC adopts a decoupled upper- and lower-body policy architecture and introduces two key modules: a Fault Estimator (FE) and a Posture Adaptation Module (PAM). The FE predicts faulty joints from lower-body proprioceptive histories, while the PAM uses this fault information to adapt the base posture plan generated by the arm policy, converting potentially unstable posture requests into safe and executable base posture commands. Through this fault-aware posture adaptation mechanism, FT-WBC synthesizes compensatory gaits under actuator failures and preserves as much arm workspace as possible while maintaining whole-body stability. Simulation and real-world experiments show that FT-WBC significantly improves survival rate and workspace under weakening or locked failures, and transfers zero-shot to a real legged manipulator in the real world.
arXiv:2606.24538v3 Announce Type: replace Abstract: Multi-modal Large Language Models (MLLMs) offer powerful reasoning for forensic tasks, yet existing approaches utilizing exogenous segmentation decoders often suffer from suboptimal localization. The reliance on stitched pipelines introduces information bottlenecks during backpropagation, which dilutes spatial signals and is limited by semantic priors of the segmentor. To address these limitations, we propose ForensicsTok, which reformulates image manipulation localization as an autoregressive sequence generation task. ForensicsTok directly generates spatially grounded token sequences, enabling precise mask prediction without intermediary supervision. Specifically, we introduce a Token Splatting Decoder (TSD) to map tokens to binary masks via codebook-aware code smoothing, which mitigates sharp gradients from deterministic detokenizers. Furthermore, to capture diverse tampering clues, we propose a Hierarchical Expert Fusion (HEF) module that injects multi-scale features from a forensic expert model. This unified architecture effectively compensates for the lack of forensic priors in standard MLLMs. Extensive experiments on six benchmarks show that ForensicsTok substantially improves over existing MLLM-based baselines and slightly improves over strong forensic expert baselines, while exhibiting stronger robustness to perturbations.
arXiv:2606.25041v3 Announce Type: replace Abstract: We present Wan-Streamer, a native-streaming, end-to-end interactive foundation model designed from the ground up for real-time, low-latency, full-duplex audio-visual interaction. Wan-Streamer seamlessly models language, audio, and video as both input and output within a single Transformer, where the sequence is represented as interleaved visual, audio, and text input tokens together with visual, audio, and text output tokens, coordinated by block-causal attention for incremental streaming. Unlike cascaded interactive systems that rely on separate VAD, ASR, language, TTS, audio-driven animation, or video-generation modules, Wan-Streamer does not rely on external language, speech, avatar, or video-generation modules: perception, reasoning, generation, response timing, turn management, and cross-modal synchronization are learned jointly within one unified model, reducing pipeline latency and error accumulation. To support natural audio-visual responsiveness, we redesign the entire stack around streamability, including causal encoders, causal decoders, block-causal attention, and low-latency multimodal token scheduling, enabling streaming units as short as 160 ms at 25 fps. Wan-Streamer achieves approximately 200 ms model-side response latency and approximately 550 ms total interaction latency when combined with 350 ms bidirectional network latency, supporting sub-second duplex audio-visual communication. These results position Wan-Streamer as a unified, end-to-end, multimodal interactive foundation model for low-latency streaming interaction.
arXiv:2606.25065v3 Announce Type: replace Abstract: Self-supervised neural garment simulation has become popular due to its computational efficiency, good visual realism, and no reliance on training data. However, existing methods greatly simplify the mechanical properties of fabrics, ignoring persistent wrinkles caused by plasticity. Although this simplification allows for modeling of purely elastic material and simple training via energy minimization, the lack of believable wrinkles adversely affects the visual realism. Therefore, we introduce the first self-supervised neural garment simulator that explicitly models persistent wrinkles. This is accomplished through a novel physics-inspired loss function, which turns learning into a moving energy minimization problem to mimic plasticity. However, this requires learning to use a changing loss function, which causes difficulties in training because the loss function changes during optimization. To this end, we propose a new physics-inspired curriculum learning scheme where the target material for learning gradually changes from pure elasticity to elasto-plasticity, allowing the loss function and the learnable parameters to jointly converge. Through a comprehensive evaluation, we show that for the first time, self-supervised learning models can generate natural persistent wrinkles, outperforming existing methods on a variety of garments, body shapes, and body motions, according to a range of metrics.
arXiv:2606.29771v1 Announce Type: new Abstract: LLM agents are increasingly cast as autonomous portfolio managers, and benchmarks have moved from financial question-answering to sequential trading. Yet most still rank agents by returns over a fixed window -- a weak proxy, since a period's return is dominated by the market path and apparent alpha can dissolve once look-ahead leakage is controlled. Such a ranking certifies neither sound reasoning, nor a consistent strategy, nor a durable edge. We introduce CLQT, which reframes closed-loop trading evaluation as diagnosis rather than ranking: an instrument that localizes where and why an agent's process succeeds or fails. CLQT is a fully closed-loop, cost-aware, strategy-consistent, temporally-gated environment whose agents run a five-stage cycle: gather, synthesize, allocate, execute, reflect. Each round emits a complete DecisionRound sealed into a recompute-verifiable hash chain, so every metric is reconstructable from the trail. Six pillars form the substrate: a hard TimeGate, institutional transaction- and financing-cost modeling, strategy-consistency scoring, three-tier memory, a Model-Context-Protocol tool layer, and mandate-aware synthesis. The same agent runs as a constrained committee of specialized roles or a single full-autonomy orchestrator, making process scaffolding an experimental variable. From the audit trail we compute a five-axis capability scorecard (APM-CS: Coherence, Acuity, Composure, Discipline, Reliability), with Coherence judged partly by a held-out, out-of-cohort LLM to curb self-preference bias. We validate it on a contamination-controlled multi-model backtest with an ablation grid and a live broker track on unseen, post-cutoff data, against a repeated-run noise floor. CLQT separates outcome from capability, yielding not a model ranking but a durable, extensible map of agent competencies and limitations.
arXiv:2606.25088v2 Announce Type: replace Abstract: We prove the following results. 1. First order model checking is fixed-parameter tractable on the class of finite fields, as a corollary of results of Ax on the theory of (pseudo)finite fields. 2. Every hereditary graph class first order definable in the class of finite groups is monadically stable, and thus has fixed-parameter tractable first order model checking. 3. Monadic second order model checking is not slicewise polynomial on the class of cyclic groups of prime-power order, unless E = NE.
arXiv:2606.28960v1 Announce Type: new Abstract: Physicians now pose millions of clinical questions to AI tools each week, yet these tools are evaluated largely on hypothetical or exam-style questions, not those actually asked in practice. We report a blinded evaluation built on 620 Real-world Point-Of-Care Queries (Real-POCQi) submitted to the OpenEvidence (OE) platform by physicians spanning 30 specialties, as well as 187 questions from HealthBench. 149 practicing physicians across 36 states made head-to-head comparisons between answers from three frontier general-purpose models (Claude Opus 4.8, Gemini 3.1 Pro, and GPT-5.5) and a specialized clinical tool (OE), with graders matched to each question's specialty. When comparing answers along five dimensions relevant to clinical decision support -- accuracy, clinical utility, source quality, verifiability, & completeness -- physicians scored the specialized tool highest on all axes; in the primary analysis on Real-POCQi, win differences (margins between win and loss rates) ranged from 25 to 39 percentage points (p<0.001). Results remained consistent in sensitivity analyses stratifying by citation display, answer length, OE-user status, and Real-POCQi versus HealthBench. In parallel, LLM judges were found to systematically differ from expert judges, though both generally agreed on the best model. These findings underscore two conclusions: (i) AI tool evaluations should reflect real-world query distributions and use expert judges that mirror the specialization defining modern medicine and (ii) the consistent advantage of the specialized tool over general-purpose models does not necessarily mean that the latter cannot serve similar purposes, but that targeted engineering and customization can yield meaningful gains in performance for its users. We release Real-POCQi as a public benchmark, as well as the prespecified statistical analysis for reproducing results of this study.
arXiv:2604.02176v3 Announce Type: replace Abstract: While textual frequency has been validated as relevant to human cognition in reading speed, its relatedness to Large Language Models (LLMs) is seldom studied. We propose a novel research direction in terms of textual data frequency, which is an understudied topic, to the best of our knowledge. Our framework is composed of three units. First, this paper proposes Textual Frequency Law (TFL), which indicates that frequent textual data should be preferred for LLMs for both prompting and fine-tuning. Since many LLMs are closed-source in their training data, we propose using online resources to estimate the sentence-level frequency. We then utilize an input paraphraser to paraphrase the input into a more frequent textual expression. Next, we propose Textual Frequency Distillation (TFD) by querying LLMs to conduct story completion by further extending the sentences in the datasets, and the resulting corpora are used to adjust the initial estimation. Finally, we propose Curriculum Textual Frequency Training (CTFT) that fine-tunes LLMs in an increasing order of sentence-level frequency. Experiments are conducted on our curated dataset Textual Frequency Paired Dataset (TFPD) on math reasoning, machine translation, commonsense reasoning and agentic tool calling. Results show the effectiveness of our framework.
arXiv:2606.25156v2 Announce Type: replace Abstract: Modern large language models based on softmax scaled-dot-product attention are constrained by their training sequence length: as the key-value sequence grows, softmax probability mass can dilute across a wider distribution, inducing activation shift and long-context performance collapse. Moreover, long-context language modeling faces a structural tension: a sliding-window attention core maintains a bounded local representation and low perplexity but is blind to long-range dependencies, while full-context attention preserves global recall but suffers from out-of-distribution perplexity explosion. To resolve these limitations, we introduce ATMA, a hybrid convolutional-attention architecture that integrates a novel three-channel attention mechanism. ATMA factorizes the attention mixing step into: (1) a count-blind, unit-vector direction channel, (2) a bounded magnitude channel driven by the participation ratio of effective matches over an extreme-value-corrected null sink, and (3) a long-term recurrent compression memory optimized via a gated-delta fast-weights rule. Neither the Polar Attention core nor the recurrent memory is sufficient alone; their combination enables monotonic perplexity reduction and high-fidelity long-range retrieval simultaneously. We evaluate ATMA using a 120-run factorial ablation sweep, demonstrating that the combined Polar + memory model maintains induction needle-in-a-haystack retrieval accuracy above 90% out to 64K tokens (32 times the training length of 2K) while its document perplexity improves monotonically, outperforming softmax-based memory baselines which collapse at extreme context lengths. Code: https://github.com/kreasof-ai/atma
arXiv:2606.25178v2 Announce Type: replace Abstract: Reinforcement learning with verifiable rewards (RLVR) has been extended from single-domain training to multi-domain reasoning suites spanning mathematics, programming, and science. However, the training curriculum (how often each domain is sampled) is typically fixed or hand-tuned, even though reasoning skills transfer unevenly across domains. Existing learnability-based curricula adapt to where the policy is currently improving, but are blind to whether a gradient step on the selected domain benefits the remaining domains. In this paper, we propose Transfer-Aware Curriculum (TAC), a bandit-style online curriculum that prioritizes domains whose updates broadly benefit the rest of the training suite. TAC repurposes signals already produced by RL training: per-domain advantages capture local learnability, and projected gradients, taken from the GRPO step being computed, estimate cross-domain transferability via gradient-geometry alignment, at negligible cost (<1% wall-clock overhead). Across a six-domain reasoning suite, TAC achieves the best macro-averaged accuracy on both Qwen3-1.7B and Llama3.2-3B, outperforming proportional random sampling, a hand-designed schedule, and a learnability-only bandit, and improving over the last of these by up to 2.8 points (10% relative). Ablations show performance degrades sharply when the transferability term is removed, and TAC remains robust on imbalanced training mixtures where learnability-only curricula over-commit to dominant domains. Our findings establish cross-domain transferability as a key signal for curriculum design in multi-domain RLVR.