E-graphs are a powerful algorithmic technique for automated equational reasoning but have resisted a conceptual understanding independent of issues of presentation and implementation. We propose an operational semantics for e-graphs based on the fundamental category-theoretic concepts of diagrams, limits, and initial functors. We illustrate the formalism through numerous example calculations and establish its soundness and completeness. Technical prerequisites are kept to a minimum in the main text; less standard elements of the mathematical theory are developed in two appendices.
Diffusion models have achieved success in high-fidelity data synthesis, yet their capacity for more complex, structured reasoning such as text following remains constrained. While advances in language models have leveraged strategies such as latent reasoning and recursion to enhance text understanding capabilities, extending these to multimodal text-to-image generation tasks is challenging due to the continuous and non-discrete nature of visual tokens. To tackle this problem, we draw inspiration from modular human cognition and propose a Recursive Modular Adaptation (ReMoA) framework integrated into joint attention layers of conventional diffusion models, iteratively refining visual tokens over multiple recursion steps while efficiently sharing parameters via sparse selection of neural modules. Comprehensive evaluations on both class-conditioned ImageNet generation and text-to-image generation using the GenEval and DPG benchmarks demonstrate that recursive visual latent refinement contributes to richer object details and better-aligned outputs. Additionally, visual latent trajectory and routing analyses reveal that neural modules learn to specialize across diffusion timesteps, with each module's trajectories diverging at later, low-noise steps to support context-adaptive refinement. ReMoA increases recall for DiT-XL from 0.57 to 0.62 on ImageNet, and improves the overall GenEval score for SD3-medium from 68.33% to 71.18%, without introducing additional parameters through recursion.
The finite-group isomorphism problem asks whether two groups, given by their Cayley tables, are isomorphic. Despite decades of study, obtaining an $N^{o(\log N)}$-time algorithm for general groups of order $N$ remains a major open challenge. Groups of nilpotency class $2$ have long been a central obstacle to achieving this goal. Recent breakthroughs have established $N^{o(\log N)}$-time algorithms for $p$-groups of class $2$ with odd $p$ (Sun, STOC '23; Ivanyos-Mendoza-Qiao-Sun-Zhang, FOCS '24; Grochow-Qiao-Stange-Sun, STOC '25) and for groups of Frattini class $2$ (Grochow-Ivanyos-Qiao-Sun, FOCS '26). However, the problem remained open for general groups of class $2$. In this paper, we present an algorithm for testing isomorphism of groups of class $2$, with running time $N^{O((\log N)^{25/26} \log \log N)}$. Our algorithm combines a tensor-based restriction of the search space, an encoding of group relations by alternating matrix tuples, and an orbit computation to recover the remaining central power information.
Reasoning models allocate test-time compute across distinct behaviors such as planning, deriving, and verifying. However, models frequently "overthink," continuing to verify and re-derive long after finding a correct answer. This wastes expensive compute and can even degrade performance. Existing interventions, such as token budgets or steering, primarily regulate reasoning duration rather than targeting specific behaviors. In this work, we demonstrate that Mixture-of-Experts (MoE) routers provide a natural, training-free interface for behavioral control. We first split reasoning traces into sentence-level segments and label each with one of six reasoning behaviors (planning, execution, restatement, fact, reflection, and conclusion), so that every token carries the behavior of its segment. Linear probes on router logits predict this label nearly as well as probes on the residual stream, the model's full hidden state and the standard target for probing, while using about half as many features. These signals generalize: experts associated with a reasoning behavior on one dataset predict the same behavior on the others, across six datasets spanning five domains (math, code, science, knowledge, and commonsense). Building on this, we selectively edit expert router scores during inference to curb overthinking. Across four models and four benchmarks, this intervention reduces the number of generated tokens by a median of 30.4% (and up to 63.3%). Crucially, it encourages the model to stop once the problem is solved, rather than skipping essential reasoning steps. Accuracy never drops by more than 0.5 percentage points, and frequently improves. Ultimately, MoE routing offers both an interpretable lens into reasoning states and a lightweight mechanism for optimizing test-time compute.
Merging companies that each run Salesforce must consolidate their customer data into a single organization while both source systems remain in use. This requires reconciling different schemas and picklists, resolving duplicate customers across and within organizations, preserving relationships between records, and keeping the target current until cutover. This article proposes a reference framework for an autonomous extract-transform-load pipeline for this task. It combines partitioned bulk extraction, declarative mapping templates, rule-based and statistical quality gates, master data management with deterministic and blocked fuzzy matching, constrained loading with per-record error capture, and change data capture guarded by a data-health circuit breaker. Each stage is mapped to Jitterbit Integration Studio constructs. The framework was implemented as an open reference prototype and evaluated on synthetic two-organization scenarios with known ground truth at three scales, from 32,040 to 383,973 source records, over five random seeds. At a preset threshold, hybrid matching reached pairwise F1 scores of 0.918 for accounts and 0.885 for contacts; calibrating the threshold per object would raise these to 0.938 and 0.963. The quality gates prevented all load failures and all silent loss of unmapped picklist values without wrongly flagging any clean record, whereas removing them produced about 1,000 load failures per run. Over five simulated days of hourly incremental synchronization, replayed events were ignored, an injected upstream fault was contained in the batch in which it occurred, and the target agreed with a full rebuild on more than 99.3 percent of field values. Validation on a production deployment remains future work.
Sparse attention accelerates few-step video diffusion transformers by computing a few key blocks exactly and compensating for the rest. However, existing sparse attention still faces key challenges, including identification overhead from finding critical blocks at every denoising step, suboptimal budget allocation that sets sparsity budgets uniformly or by local metrics, and cross-step error accumulation as reused masks or clusters go stale. Observing that the errors left on skipped blocks cancel each other once every key is compensated, we propose DecoupAttn, a new paradigm of decoupled sparse attention: semantic clusters compensate every key, and exact attention corrects their error on a few blocks fixed once offline by position. To support this paradigm, we introduce Distortion Budget, a rate-distortion objective for budget allocation that places exact computation across timesteps, layers, and heads where it removes the most distortion that compensation leaves in the final latent. Built on this budget, we further propose Reassign Cache, in which each step reassigns the drifting keys to cached centroids and refreshes compensation while the budgeted blocks stay fixed. Moreover, DecoupAttn is training-free: a few calibration clips compile one plan offline, and inference runs standard block-sparse kernels with one fused compensation kernel. Our code will be released after submission.
We study the minimum-disagreement cluster editing problem that, from a complete edge-signed graph, where a positive edge represents two vertices agreeing to each other while a negative represents a disagreement, constructs disjoint cliques of upper bounded size such that the total number of negative edges inside cliques and positive edges outside cliques is minimized. When the sizes of the candidate cliques are unbounded, the problem is NP-hard and APX-hard, and many approximation algorithms with improving ratios have been proposed in the literature and the state-of-the-art approximation ratio of $α= 1.3865 + ε$. The capacitated variant, in which the candidate cliques have sizes no greater than a given $k$, is less studied and the best approximation ratio is $5.37$. In this paper, we study the latter problem. We first present a downsizing procedure that achieves an approximation ratio of $α_k = 2 α+ 1$, based on any $α$-approximation algorithm for the uncapacitated problem. Secondly, we propose a hybrid of cluster-LP sampling and pivoting and show that for every fractional capacitated cluster-LP solution of cost $L$, the hybrid using ordinary pivoting has an expected cost at most $\frac 95 L + \frac 1{10} \OPT_k$, and the hybrid using conditional pivoting has an expected cost at most $\frac {306}{203} L + \frac {50}{203} \OPT_k$. A cardinality-constrained adaptation of the approximate dual-separation framework of García-Soriano and Schohn supplies the required fractional solution and yields a randomized $(1.9 + ε)$-approximation algorithm and a randomized $(1.7537 + ε)$-approximation algorithm, respectively.
Multimodal Large Language Models have shown promising capabilities in bridging visual and textual reasoning, yet their reasoning capabilities in Open-Vocabulary Human-Object Interaction (OV-HOI) are limited by cross-modal hallucinations and limited viewpoints of images. To address this, we propose ImagineAgent, an agentic framework that integrates cognitive mapping, tool-augmented reinforcement learning (RL), and generative world modeling for robust OV-HOI understanding. Specifically, we first propose an innovative CoT dataset named hicodet-6K for supervised fine-tuning (SFT), which effectively bridges the perception-to-cognition gap by structuring perceived entities into interaction pairs for comprehensive predictions. Subsequently, we develop a multimodal tool library integrating online retrieval, image cropping, and generative modeling, enabling the agent to dynamically augment reasoning with domain-specific tools to resolve visual-semantic ambiguities and hallucinations during inference. Moreover, we incorporate a generative model to reconstruct alternative viewpoints, enabling the agent to 'imagine' under limited viewpoints. Finally, we propose a composite reward mechanism to jointly optimize prediction accuracy and tool efficiency. Evaluations on both SWIG-HOI and HICO-DET datasets demonstrate that our method achieves state-of-the-art performance while requiring merely 36.7% of the training data compared to existing methods, validating our robustness, empirical effectiveness and efficiency.
Low-precision numeric types such as 8-bit integers, bfloat16, and half-precision floating point (fp16) are often used in AI inference workloads to reduce memory bandwidth, and RISC-V's openly specified Vector extension (RVV) is a natural target for compiling such workloads with auto-vectorization. This paper examines whether LLVM's RISC-V auto-vectorizer handles these narrow types equally well, using a multiply-accumulate loop written for fp32, int8, bf16, and fp16. Under a common baseline target (-march=rv64gcv), int8 and fp32 vectorize cleanly, while bf16 and fp16 degrade to increasingly scalar-heavy fallback code, with fp16 falling back to per-element software conversion calls, unless additional target-feature extensions (zvfbfmin/zvfbfwma, zvfhmin) are enabled. Under QEMU user-mode emulation, the baseline bf16 and fp16 kernels execute $17.0\times$ and $75.9\times$ as many dynamic instructions as their extension-enabled counterparts. Compiler diagnostics do report this degradation, but do not distinguish, in their wording, between a limitation that a compiler flag resolves entirely and one reflecting an inherent conversion cost, leaving no direct signal toward the fix. These results are supported by assembly-level instruction analysis, by direct inspection of LLVM's optimization remarks, and by dynamic instruction counts collected under emulation.
Hui Chen, Antoine Didisheim, Mo Pourmohammadi, Luciano Somoza, Hanqing Tian
原文摘要
We use sparse autoencoders (SAEs) to "brain scan" large language models, mapping internal activations to interpretable concepts and enabling targeted "steering" of specific features. We show that prompting models to adopt specific preferences activates many unintended concepts, complicating causal inference. In contrast, SAEs permit smooth, more isolated steering of target concepts. Furthermore, our scans reveal a stark disconnect between the model's internal feature contributions and its self-generated explanations, showing that these explanations are unreliable guides to its reasoning. Finally, SAEs provide a lightweight intervention for existing models, producing interpretable embeddings and allowing researchers to attenuate specific economic biases.
High-fidelity photovoltaic (PV) module mapping requires complete instance recovery and accurate delineation. Vision foundation models offer zero-shot segmentation, but converting incomplete and inconsistent predictions across views into a complete, geometrically accurate map of individual modules remains challenging. We propose the Multi-View PV Module Mapper (MV-PVMapper), a training-free framework coupling vision foundation models with unmanned aerial vehicle (UAV) photogrammetry. First, Segment Anything Model 3 (SAM 3) generates candidates from original images using text prompts. Their contours are then transferred into object space through per-module slope correction to limit distortion from surface-model noise, and verified against representative module dimensions to guide cross-view instance selection. Finally, accepted instances return to all covering images as box prompts for SAM~3, supplementing text prompts with spatial cues for iterative segmentation, verification and fusion. On five-direction surveys of three operational rooftop PV plants at approximately 1 cm nadir ground sampling distance, two feedback iterations recover 7,710 of 7,711 reference modules with no false positives at an intersection-over-union (IoU) matching threshold of 0.5. Macro-averaged area precision and recall reach 99.00 % and 98.52 %, respectively. Averaged across sites, the aggregated Jaccard index reaches 97.55 %, compared with 60.73 % and 41.24 % for orthomosaic segmentation and multi-view pixel-voting baselines, respectively; average precision at the stricter IoU threshold of 0.95 reaches 95.67 %. Ablation studies show that module-size constraints are critical for instance separation, while slope correction and feedback to all covering views improve boundary accuracy. These results are achieved without task-specific image annotation or segmentation-model weight updates.
Unlike conventional imaging modalities, such as magnetic resonance imaging, which are often well described by a linear regression framework, coherent imaging systems follow a significantly more complex model. In these systems, the task is to estimate the unknown image ${\boldsymbol x}_o \in \mathbb{R}^n$ from observations ${\boldsymbol y}_1, \ldots, {\boldsymbol y}_L \in \mathbb{R}^m$ of the form $ {\boldsymbol y}_l = A_l X_o {\boldsymbol w}_l + {\boldsymbol z}_l, l = 1, \ldots, L,$ where $X_o = \mathrm{diag}({\boldsymbol x}_o)$ is an $n \times n$ diagonal matrix, ${\boldsymbol w}_1,\ldots, {\boldsymbol w}_L \stackrel{\text{i.i.d.}}{\sim} \mathcal{N}(0,I_n)$ represent speckle noise, and ${\boldsymbol z}_1, \ldots, {\boldsymbol z}_L \stackrel{\text{i.i.d.}}{\sim} \mathcal{N}(0,σ_z^2 I_m)$ denote additive noise. The matrices $A_1, \ldots, A_L \in \mathbb{R}^{m \times n}$ are known forward operators determined by the imaging system. Here, $m$ can be less than $n$, which makes the linear model ill conditioned, and we collect $L$ measurements of the whole setup and try to quantify the loss. The fundamental limits of conventional imaging systems have been extensively studied through sparse linear regression models. However, the limits of coherent imaging systems remain largely unexplored. The goal of this paper is to close this gap by showing, for the first time in the literature, that the minimax risk of estimating ${\boldsymbol x}_o$ scales as $\frac{\max\{σ_z^4,\, m^2,\, n^2\}\, k \log n}{m^2 n L},$ where the parameter $k$ quantifies the effective complexity of the image class.
Yoon Pyo Lee, Samrendra Roy, Jay Phil Yoo, Kazuma Kobayashi, Souvik Chakraborty, Syed Bahauddin Alam
原文摘要
Planning electricity systems requires repeated decisions that connect regional resources, investment, operation and network constraints. Here we present an autonomous planning framework that carries prepared regional scenarios through proposal, numerical evaluation, feedback interpretation, redesign and final plan selection. Agents propose investment bounds, storage operating conditions and feeder interventions, interpret numerical feedback, and decide how to revise designs and when to stop. The framework automatically executes annual investment-and-dispatch optimization and representative-feeder verification, applies fixed acceptance criteria and retains the lowest-cost accepted candidate. We evaluate forty carbon-constrained trajectories across ten system-year scenarios in five electricity systems and four countries, using two model deployments and single- and multi-agent architectures. All forty retain candidates satisfying the implemented physical and fixed annual operational carbon-budget checks. Twenty-eight of forty matched reference designs exceed the new scenario ceilings. A separate archive of ninety-six trajectories under a common 1,000 Mt annual ceiling provides matched references and broader regional coverage. The results demonstrate autonomous planning under explicit emissions constraints using public regional data and published institutional demand outlooks.
Saliency maps, offered as evidence that mammography classifiers use the cancer, are untested in dense breasts, which can hide cancers. We tested whether they locate cancers worse there, and why. On 1,345 malignant CBIS-DDSM views, we scored three CNNs' Grad-CAM, Integrated Gradients and occlusion maps by whether their peak hits the lesion and a chance-normalized lesion-to-background saliency ratio (bits; +1 doubles it). Unadjusted gaps were large: Grad-CAM peaks hit 44% of dense vs. 58% of non-dense lesions. Case mix, chiefly more calcifications in dense breasts, explained about half or more of each gap. Occlusion kept a significant average of -0.56 bit gap. Controls also implicate the classifiers: unadjusted gaps exceeded an image-only conspicuity map's, and models relied less on dense-breast lesions. We recommend reporting density gaps before and after case-mix adjustment, per classifier, with conspicuity and lesion-reliance controls.
Siqing Huo, Andrew Parry, Debasis Ganguly, Charles L. A. Clarke
原文摘要
Second-stage neural rankers are commonly applied with a fixed re-ranking depth, assuming that retrieval effectiveness saturates as depth increases. Prior research has questioned the assumption that increasing the re-ranking depth yields linear performance gains, further suggesting that optimal re-ranking depth varies considerably from query to query. In the past, studying such phenomena was methodologically difficult given the scale of manual annotation required. With the advent of LLM-based relevance judgments we can now more easily undertake such studies, in this case to pinpoint the optimal re-ranking depth on a per-query basis. Using dense LLM-based relevance judgments over a typical re-ranking pipeline, we show that many queries exhibit a well-defined optimal re-ranking depth, beyond which effectiveness stagnates or degrades. We formulate re-ranking depth as a query-specific property and study whether it can be predicted a priori from first-stage retrieval characteristics. Through a large-scale analysis, we find that most standard query performance prediction (QPP) methods are ineffective for this task. In contrast, a predictor derived from LLM-assessed first-stage ranking quality, which we term IR-DCG@10, can reduce the average re-ranking depth by up to a factor of 3 while preserving overall effectiveness, depending on the specific first- and second-stage rankers used. Under oracle selection of optimal depths, we further show that retrieval effectiveness can improve by more than 7% while reducing the average re-ranking depth by a factor of 5 on the MSMARCO DEV collection. Given our promising preliminary findings, we would encourage the use of automatic judgments to facilitate research otherwise infeasible under manual annotation. Towards this point, we release relevance judgments, our codebase, and experimental artefacts to support reproducibility and further research.
Kaiyuan Chen, Shuangyu Xie, Letian Fu, Justin Yu, William Pacini, Sandeep Bajamahal, Hudson Kim, Jaimyn Drake 等 24 位作者
原文摘要
For robots to work reliably in commercial and industrial applications, recent advances in agentic coding systems have potential to combine interpretable robot programming with the open-world adaptability of model-free policies. We focus on "Variational Automation" (VA), a class of tasks that have larger variations in object geometry and pose than fixed automation. Model-free policies often struggle to close the reliability gap for VA tasks, which must be executed persistently and reliably in commercial and industrial applications. Motivated by prior work on Task and Motion Planning (TAMP) and the Robot Operating System (ROS), we introduce Graph-as-Policy (GaP), a multi-agent coding harness that generates directed computation graphs with perception, planning, and control nodes from a Modular Open Robot Skill Library (MORSL). GaP generates an internal simulation environment to rehearse task instances with different graphs in parallel to iteratively refine the graph structure and parameters to improve success rates and throughput. Evaluation with 8 new open VA task benchmarks, 4 in simulation and 4 in real-world, suggests that GaP can achieve success rates that significantly outperform baselines. Details, code, and data: https://graph-robots.github.io/gap
Patient world models are increasingly proposed as simulators of clinical trials, and recent surveys rank their capabilities on a ladder from temporal prediction to counterfactual rollouts. Position on that ladder is currently assigned by reading papers rather than by measurement. We turn the counterfactual rung into a measured quantity. From 30,119 completed phase 2/3 trials with results posted to ClinicalTrials.gov we construct 9,996 audited within-trial contrasts from 7,207 trials, and we score a simulator on how much of its accuracy depends on knowing which intervention was tested. The instrument is an intervention-blind reference: the same model, trained on the same trials, with every intervention-identifying field withheld. Interventional skill is the error the full model removes relative to that reference. On a temporal split (878 test contrasts from 655 trials) a registry-trained gradient-boosting simulator removes a fifth of the error of a trivial baseline, but its interventional skill is -0.12 (95% CI -0.21 to -0.05): withholding the intervention makes it better, and no alternative featurisation, stratum, temporal cut or same-molecule retrieval yields an interval above zero. A frontier language model reading only the registered protocol has interventional skill 0.20 [0.08, 0.31]; smaller and open-weight models sit at the limit of detection. Perturbation tests show the large model uses the intervention as such: swapping the arms reverses the sign of its prediction in 86% of contrasts. The blind reference is itself not drug-blind (a language model recovers the withheld drug class 45% of the time from the blind prompt), so interventional skill is a conservative lower bound. A reliability layer shows that ranking by predicted error lowers it but that calibrated error control does not transfer under temporal shift. Benchmark, code, prompts and per-call model outputs are released.
As sequence lengths continue to grow, distributed attention based on sequence parallel (e.g., Ring-Attention) has become a mainstream solution for large language models (LLMs). Nevertheless, inherent communication dependencies limit computation-communication overlap, reducing GPU utilization. Moreover, in cross-hardware (NUMA or node) architectures, due to imbalanced communication bandwidth, frequent communication synchronization will limit overall performance. Based on these insights, we propose CHARM+, Cross-Hardware Attention with Re-Merge Multistream Mechanism. We reorganize sequential communications into several levels and achieve intra-level parallelism and computation overlap. Furthermore, we decouple cross-hardware communication from intra-hardware communication, eliminating synchronization overhead caused by bandwidth imbalance. Forward and backward algorithms were evaluated on multiple hardware architectures and consistently outperformed state-of-the-art methods. Compared to Ring-Attention, our forward and backward algorithms achieve average speedups of 1.3x and 1.7x on a dual-NUMA PCIe architecture, and 2.8x and 2.9x on a dual-node NVLink architecture, respectively.
Jeremy Collins, Namra Patel, Ayush Agarwal, Shitij Govil, Animesh Garg
原文摘要
Pretrained video generation models distill internet-scale motion priors into robot policies, but full video is a redundant signal for action prediction. We argue that dense 3D motion is the medium through which video pretraining best transfers to robot actions, and that predicting motion rather than video affords better sample-efficiency for action prediction. FloMo is a 5B video-generation model fine-tuned to predict 3D scene-flow and action tokens. Rendering scene flow as RGB and encoding it into the same latent space as video enables a pretrained video generation backbone to efficiently adapt to predict 3D scene flow. Co-trained on action-free human video and action-labeled robot teleoperation data, our model benefits from human-video pretraining and generalizes to unseen objects, motion primitives, and tasks. Replacing motion prediction with a video prediction objective degrades out-of-distribution generalization, while adding a video prediction target alongside motion lowers policy success rather than improving it. When motion, video, and action are predicted jointly, action tokens attend far more to motion than to video, suggesting that predicting video wastes model capacity. Together, these results indicate that motion effectively mediates the transfer from video pretraining to action. More information can be found at flomo-wam.github.io.
Xingjian Bai, Jia-Hao Liang, Jayden Teoh, Vincent Sitzmann
原文摘要
We study a reactive optimizer that uses a pretrained language-model agent to inspect training dynamics and rewrite executable optimizer code during a single training run. The agent can change update rules and schedules for individual parameter groups, without optimizer-specific meta-training or memory transferred from other runs. We evaluate this approach, OPAL, on language modeling and image generation when a recipe has not been tuned for the target setting. Across the evaluated settings, it matches or improves on selected fixed recipes. Analysis of its decisions shows changes to learning-rate schedules, optimizer structure, numerical precision, and momentum. Controlled experiments also show adaptation to repeated data corruption and reduced overfitting. We report these behaviors together with the agent's computational cost and implementation requirements.
Hei Yi Mak, Hoang Le, Mehran Taghian Jazi, Yunke Peng, Huan Lin, Zhengzhi Tang, Zhexu Liu, Yao Wang 等 11 位作者
原文摘要
In this work, we study end-to-end FP4 reinforcement learning (RL) post-training for Mixture-of-Experts (MoE) large language models. Recent work on FP4 RL has introduced effective recovery recipes that restore accuracy under aggressive low-precision quantization for dense models, but we find that the same approach yields no improvement over the naive FP4 baseline in the MoE setting. We conduct an in-depth study to understand why. Our analysis identifies the root cause in the routing mechanism unique to MoE: because the router makes discrete expert-selection decisions, training-inference policy alignment becomes acutely sensitive to whether the same recovery technique is applied symmetrically on the training and inference sides. When an asymmetric recipe is used-including the dense-model approach, which corrects only the rollout-training-inference alignment deteriorates sharply, as measured by the Pearson correlation coefficient between training and inference probabilities, and router selection matching rates. Building on this finding, we propose MoEResQ, a quantization-recovery recipe for MoE RL post-training that enforces symmetry between the training- and inference-side corrections. MoEResQ restores training-inference alignment and substantially closes the accuracy gap to full precision, establishing symmetric recovery as the key principle in FP4 RL for MoE models.
KV cache eviction is not merely a token-ranking problem: the workspace retained for generating eviction queries also determines the quality of those queries. Consequently, an early workspace error can distort later query observations and reinforce an incorrect final eviction decision, particularly when evidence is dispersed across a long context. We present EchoKV, which discovers and addresses this failure mode with iterative correction. It initializes a feedback-compatible bootstrap workspace using proposals at future decode positions, then uses early decode-generated queries to recover the workspace before the remaining lookahead steps. It uses the same number of workspace forward passes as LAQ while adding a tensor re-scoring operation for workspace repair. On LongBench (16 tasks, Mistral-7B-Instruct-v0.2 and Qwen2.5-7B-Instruct), EchoKV shows notable improvements over LAQ on several tasks requiring dispersed evidence, including multi-hop cross-document QA and long single-document comprehension. Across budgets and backbones, EchoKV selectively improves evidence-intensive tasks while maintaining comparable performance on the remaining LongBench subsets. These results identify query--workspace coupling as a practical failure mode of query-driven KV eviction.
Youssef Esseddiq Ouatiti, Mohammed Sayagh, Bram Adams, Ahmed E. Hassan
原文摘要
Developers insert logging statements in source code to record its execution, and the resulting logs are essential for detecting and troubleshooting system anomalies. A key part of these statements is the log level (e.g., DEBUG, INFO, WARN, ERROR), which determines whether a message is recorded. Choosing it is a trade-off: levels set too low mask important information in production, while levels set too high create noisy logs and extra I/O and storage overhead. Recent work uses Large Language Models (LLMs) with In-Context Learning (ICL) to predict log levels, but typically augments the prompt with randomly sampled examples. We argue that this overlooks the socio-technical nature of logging, where conventions vary by component and development team. Off-the-shelf LLMs using random examples are reported to achieve median AUCs of only 0.56 to 0.76, leaving room for improvement without fine-tuning. In this paper, we propose a retrieval approach for LLM-powered log level prediction that prioritizes context relevance over random selection and basic similarity. Our method retrieves logging examples relevant along two socio-technical dimensions: (i) functionality (code performing similar operations) and (ii) ownership (code maintained by the same developers). Evaluating four LLMs across ten popular Java projects, we find that this retrieval improves AUC over random baselines by 0.07 (median), and combining both signals yields median AUCs of 0.83 to 0.97. Our ablation studies show that each signal improves on random retrieval, and that combining them outperforms the stronger of the two on all ten projects. Our results also significantly exceed state-of-the-art supervised predictors (e.g., DeepLV) by a median AUC of 0.12 to 0.19. Our findings highlight the importance of socio-technical signals such as code functionality and ownership for LLM-powered log level prediction.
Personalized retrieval-augmented generation requires extracting and applying user preferences from extended interaction histories. This task is challenging because preference evidence is scattered across past conversations and remains bound to specific temporal and situational contexts. Consequently, isolated fragments rarely provide complete support for a response, while preferences expressed under different conditions are easily conflated. To address these limitations, we propose Weaver, a framework that incrementally weaves memory-conditioned retrieval trajectories with compositional preference reasoning for personalized responses. Weaver operates on a Dual-Layer Retrieval and Reasoning Graph that couples an upper Intent Dependency Graph with a lower Memory-Conditioned Retrieval Trajectory Graph. Dependency Information Gap Tracking represents the information needed to answer the current request and its prerequisite dependencies in the upper graph. Memory-Conditioned Incremental Retrieval uses established premises and clues from retrieved conversations to locate missing evidence, recording the resulting retrieval trajectories in the lower graph. Scope-Aware Compositional Reasoning jointly interprets the accumulated evidence to determine which preferences apply to the request. It identifies complementary and alternative support and distinguishes conflicting claims within the same temporal and situational context. These conclusions iteratively update task premises and remaining information gaps, guiding subsequent retrieval to establish a coherent contextual basis for the final response. Experiments on personalized RAG benchmarks, particularly PersonaMem, demonstrate that Weaver outperforms leading baselines in both evidence completeness and answer quality.
Speculative decoding accelerates rollout generation, which dominates the cost of reinforcement learning (RL) post-training. Online co-training keeps the draft aligned with the evolving policy, further increasing acceptance length and speedup. However, training large models on long contexts requires context parallelism (CP) and pipeline parallelism (PP), and co-training drafts under this layout poses two obstacles: (1) the draft's branch attention is unsupported by causal CP, and (2) the target features that the draft consumes are spread across PP stages. In this paper, we present the design and implementation of online draft co-training in NeMo-RL. For CP, we extend packed, load-balanced zigzag ring attention by merging rank-local branch attention with causal main-sequence attention. For PP, TapChannel transports intermediate target features across stages via a separate path, leaving the pipeline schedule unaffected. Experiments show that co-training keeps the policy's reward and accuracy close to the no-draft baseline while delivering $1.48$--$2.25\times$ rollout speedups and $1.19$--$1.88\times$ end-to-end training speedups across model scales up to 122B. Our CP design achieves strong scaling at 256K sequence length with significant memory savings over prior work, and our PP transport incurs modest overhead. Code can be found at https://github.com/NVIDIA-NeMo/RL/issues/3698.
Protein structure prediction has largely focused on native conformations, leaving the continuum of unfolded states, folding intermediates, and native structures difficult to access. We present FoldScape, a sequence-conditioned deep learning framework for generating protein conformational ensembles at specified folding fractions $Q$. Trained on approximately one million conformations from 10,000 protein sequences, FoldScape combines coarse-grained folding and all-atom mechanical-unfolding trajectories. It reconstructs three-dimensional $C_α$ chains through hierarchical sliding-window regressions of inter-residue distances, chain-level assembly, and a rectified-flow decoder. By conditioning on folding progress, FoldScape generates conformations from unfolded states through intermediates to native-state structures without protein-specific molecular dynamics at inference. In matched-sample comparisons, FoldScape achieves median conformational-space coverage of 0.989, 0.722, and 0.877 along radius of gyration, fraction of native contacts $Q$, and secondary-structure fraction, respectively, versus 0.05, 0.05, and 0.10 for seed-varied AlphaFold 3 ensembles. Coverage advantages are consistent across the evaluated proteins and robust to binning choices. Reconstruction accuracy remains approximately stable across the evaluated stages of a folding trajectory, and stride-one crop-and-superpose inference extends predictions to chains several times longer than the single-pass limit. Analyses of mechanical-unfolding trajectories show residue-resolved pathways, reproducible extension-conditioned observables across pulling velocities, and geometrically coherent segments in partially unfolded chains. FoldScape extends sequence-based structure prediction beyond a single folded endpoint to controllable conformational-ensemble generation across protein folding.
Neuro-symbolic systems typically combine neural perception with discrete symbolic solvers, limiting end-to-end optimization. We introduce Attention-Based Soft Answer Sets (AS$^2$), a fully differentiable neuro-symbolic architecture that replaces discrete solving with a continuous approximation of the Answer Set Programming (ASP) immediate-consequence operator $T_P$. AS$^2$ maintains probability distributions over finite symbol domains and learns by minimizing the fixed-point residual of a probabilistic version of $T_P$, enabling differentiation through constraint checking without external solvers during training or inference. For spatial constraint-satisfaction tasks, AS$^2$ replaces positional embeddings with constraint-group membership embeddings derived directly from declarative ASP specifications, making the model independent of arbitrary position indexing. On Visual Sudoku, AS$^2$ achieves 99.89% raw cell accuracy and 100% constraint satisfaction after greedy decoding, without external solvers. On CLEVR-Hans, AS2 achieves 99.94% on CLEVR-Hans3 and 87.77% on CLEVR-Hans7. These results demonstrate that differentiable fixed-point reasoning, constraint-aware attention, and declarative constraint specifications can match or outperform pipeline-based and solver-based neuro-symbolic approaches while preserving end-to-end differentiability.
Shaikh Akib Shahriyar, Zinedine Sehili, Adam Arthur, Christopher Schwartz
原文摘要
While generated chest radiographs can serve as valuable data augmentation when explicitly declared, the same pixels become unsupported evidence when submitted as records of real clinical examinations. We study this dataset-intake vulnerability by disentangling three questions: (1) whether an image passes an automated technical quality check, (2) whether an origin classifier flags it as synthetic, and (3) whether a trusted provenance record supports the claimed clinical acquisition. In this study, a submitter uses Qwen3.8-27B to select up to 20 candidate radiographs from pools of 1,000 Cheff-identified or SynthCheX images, while a receiver evaluates image quality via MedGemma 1.5 and predicts origin using a linear probe over frozen DINOv3 features. Although quality selection increases MedGemma quality passes from 12/20 to 19/20 for Cheff-identified images and from 8/20 to 17/20 for SynthCheX, we find no evidence that candidate selection increases passage through both checks simultaneously: zero selected Cheff-identified images pass both gates, and uniform, quality-guided, and surrogate-guided selection yield 2, 1, and 2 SynthCheX passes out of 20, respectively. On unprocessed test cohorts, the origin classifier rejects 4,677 of 5,000 SynthCheX images at a nominal 5% false-alarm rate (flagging 110 of 2,297 real NIH radiographs), yet leaves 152 SynthCheX images that pass both pixel-level checks. Moreover, pixel-level screening is fragile to routine distribution shifts: JPEG compression reduces SynthCheX-versus-NIH AUROC from 0.987 to 0.841, and applying the NIH-calibrated threshold to an external genuine cohort (Open-I) triggers false alarms on 568 of 957 authentic radiographs (59.4%). Because pixel-level checks can filter submissions but cannot verify clinical acquisition claims, we argue that dataset intake pipelines must record and verify source provenance independently of image pixels.
An ordered sequence of end-effector poses can give a redundant robot arm multiple valid joint configurations at each waypoint. Inverse kinematics (IK) computes these configurations. Choosing a configuration only for the next waypoint can leave the arm poorly placed for later waypoints. This may increase planned motion time or cause a later planning failure. This paper presents Future-Window Inverse Kinematics (FWIK), an online method that selects IK solutions by considering a small window of future waypoints. FWIK generates several IK candidates, connects them in a layered joint-space graph, and ranks candidate paths with a simple motion-time estimate. A learned model removes paths whose transitions are likely to cause self-collision. A standard MoveIt 2 planner then checks the next-waypoint IK goals from the remaining paths in ranked order, and the first goal that yields a valid trajectory is selected for execution. We evaluate FWIK in simulation on two 7-DoF manipulators, the Franka Emika Panda and the KUKA LBR iiwa. Each method-planner setting contains five repetitions of 100 sequences with 50 waypoints. Across Pilz, RRT-Connect, CHOMP, and STOMP, FWIK completes 96.8-100% of sequences and gives the shortest planned trajectory duration in all eight robot-planner settings. Its mean duration is 12.8-13.1% lower than STAMPEDE, 22.5-25.6% lower than IKLink, and 27.5-37.2% lower than retry-based TRAC-IK. The results show that a small future window can improve sequence completion and reduce planned motion time while leaving final trajectory validation to the motion planner.
Steering methods influence Large Language Model behavior by identifying semantic directions in hidden representations, but are typically realized through inference-time activation interventions that apply a fixed, global modification to the model's internal states. While effective, such interventions often induce unfavorable attribute-utility trade-offs under strong control, as they ignore the fact that many behaviors are governed by a small and heterogeneous subset of model components. We propose Steer2Edit, a theoretically grounded, training-free framework that transforms steering vectors from inference-time control signals into diagnostic signals for component-level rank-1 weight editing. Instead of uniformly injecting a steering direction during generation, Steer2Edit selectively redistributes behavioral influence across individual attention heads and MLP neurons, yielding interpretable edits that preserve the standard forward pass and remain compatible with optimized parallel inference. Across safety alignment, hallucination mitigation, and reasoning efficiency, Steer2Edit consistently achieves more favorable attribute-utility trade-offs: at matched downstream performance, it improves safety by up to 17.2%, increases truthfulness by 9.8%, and reduces reasoning length by 12.2% on average. Overall, Steer2Edit provides a principled bridge between representation steering and weight editing by translating steering signals into interpretable, training-free parameter updates. Our code is available at https://github.com/Trustworthy-ML-Lab/Steer2Edit