arXiv_AI
-
DIAGNO: Diagonal Spherical Neural Operators for Heterogeneous Earth Dynamics Modeling
2026-11-12 06:15:08Herui Li, Bin Lu, Haonan Qi, Lei Zhou, Luoyi Fu, Xinbing Wang, Meng JinAbstract
Accurate forecasting of Earth's dynamic systems is vital in geophysics, where neural operators offer a powerful data-driven paradigm. Standard operators often treat the Earth as a flat Euclidean plane, introducing geometric artifacts. In contrast, Spherical Neural Operators (SNOs) resolve this mismatch by operating within the spherical spectral space, naturally adapting to spherical geometries. However, for computational efficiency, existing SNOs typically rely on the rotation equivariance assumption. This assumption simplifies the transitions of physical states into isolated spectral modal evolutions, thereby over-idealizing the Earth as a homogeneous system. To solve this problem, we no longer rely on the rotation equivariance assumption that isolates spectral modal evolution, but directly derive an explicit modal interaction kernel and propose Diagonal Spherical Neural Operators named DIAGNO. DIAGNO explicitly parameterizes cross-modal interactions through spectral modal decomposition, thereby preserving the inherent heterogeneity of Earth dynamics. Furthermore, leveraging the unique diagonal structure of the spherical spectrum, we design inter- and intra-diagonal interaction mechanisms to natively capture essential zonal and meridional dynamics, respecting intrinsic geophysical regularities. Extensive experiments across simulated fluid dynamics and real-world Earth system reanalysis (including atmosphere and ocean) datasets demonstrate DIAGNO's superior capability, consistently achieving state-of-the-art performance in multi-step forecasting. Beyond this, we validate DIAGNO's practical analytical value in professional geophysical domains by assessing its fidelity in capturing climatological anomalies and maintaining physical consistency.
Abstract (translated)
URL
https://arxiv.org/abs/2611.16356
PDF
-
HyBridge: Bridge-Guided Self-Supervised Hyperbolic Hypergraph Representation Learning
2026-11-12 06:10:00So-Bin Jung, David Yoon Suk, Sang-Wook KimAbstract
Hypergraphs capture groupwise relationships among multiple nodes and often exhibit tree-like structures. Although hyperbolic space is well suited for modeling such structures, simply extending existing hypergraph representation learning (HRL) methods to hyperbolic space does not consistently improve performance. In this paper, we investigate the relationship between node structural properties and hyperbolic embedding radii and find that bridge nodes tend to lie closer to the center in graphs, whereas this relationship is not consistently observed in hypergraphs. We further find that conventional bridge score estimation may fail to capture the groupwise connectivity of hypergraphs. Motivated by these findings, we propose HyBridge, a self-supervised hyperbolic HRL framework that (1) estimates node bridge scores by considering the external connectivity of individual hyperedges, (2) learns node and hyperedge embeddings through hyperbolic hypergraph encoding, (3) guides node radial positions according to their bridge scores, and (4) employs contrastive learning to preserve node discriminability. Extensive experiments on 10 real-world hypergraphs demonstrate that HyBridge outperforms 22 competing methods, achieving performance improvements of up to 9.82% in node classification. For reproducibility, we provide the code at this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2611.16351
PDF
-
DRAMA: Dissecting Attention Redundancy for Accelerating Multimodal Diffusion Large Language Models
2026-11-12 06:04:08Yi-Hsin Hung, Fangfu Liu, Shao-Yuan Lo, Chu-Song ChenAbstract
Multimodal diffusion large language models (dLLMs) generate text through iterative masked denoising with bidirectional attention, offering a compelling alternative to autoregressive vision-language modeling. However, each denoising step requires a full forward pass over long multimodal sequences in which visual tokens dominate the sequence length, and a fixed step budget further compounds this quadratic per-step cost. Existing inference acceleration approaches for multimodal dLLMs characterize attention redundancy implicitly, motivating designs from isolated attention visualizations rather than from measurement. We close this gap with DRAMA, a training-free acceleration framework grounded in a systematic dissection of attention redundancy in multimodal dLLMs. By probing attention tensors along spatial and temporal axes, we reveal two persistent redundancy structures: (i) attention mass consistently concentrates on a sparse subset of visual tokens, and (ii) attention distributions rapidly stabilize across denoising steps after an initial formation phase. These findings motivate two complementary mechanisms: Spatial-Redundancy-Aware Visual Token Pruning (SVTP) and Sufficiency-Guided Adaptive Stopping (SGAS). Together, DRAMA achieves an average 18.67 times inference speedup across six vision-language benchmarks without training, architectural modification, or attention caching, while maintaining competitive accuracy. Our results establish a strong, measurement-grounded acceleration baseline for multimodal dLLM inference.
Abstract (translated)
URL
https://arxiv.org/abs/2611.16345
PDF
-
DForkTree: Accelerating Speculative Decoding with Fork-Aware Block Diffusion Draft Trees
2026-11-12 06:04:02Zhicheng Zheng, Cheng Zhang, Chenkai Tang, Tianxiang ZhaoAbstract
Speculative decoding speeds up language models by having a lightweight drafter propose future tokens that the target model verifies in parallel. Block diffusion drafters such as DFlash predict an entire block in one forward pass, and DDTree builds draft trees directly from their per-position distributions. Each distribution implicitly assumes that the drafter's own greedy continuation is accepted, whereas a draft tree improves over a chain only on branches that leave this greedy path. Once a path forks, the drafter's top-1 accuracy drops from 89% to 18% and its top candidates often miss the true token. Re-running the drafter from the fork token largely restores its accuracy, showing that the errors stem from conditioning on an outdated prefix rather than from limited model capacity. In this paper, we introduce DForkTree, a training-free method that re-anchors each branch on its own tokens. It adjusts the drafter's scores with pointwise mutual information from target-generated n-grams, weighted more heavily after a fork, and a layer-wise GPU search recovers the exact optimal tree under these scores. On Qwen3.5-4B and 9B, DForkTree outperforms all baselines on every benchmark, by up to 41.5% over DFlash and 15.1% over DDTree, with average speedups of up to 4.75x over autoregressive decoding.
Abstract (translated)
URL
https://arxiv.org/abs/2611.12356
PDF
-
SUMNO: Spatially Unrolled Multigrid Neural Operator for Fluid Flow Surrogates
2026-11-12 06:02:05Yingqi Hu, Nan Deng, Changxiao Shao, Xiaozhou HeAbstract
Data-driven surrogate modeling has emerged as a powerful paradigm for accelerating computational fluid dynamics, and hierarchical U-shaped networks, which resemble the structure of classical Geometric Multigrid (GMG) methods, have become prevalent designs. Modern surrogates built on these designs achieve competitive accuracy at a fraction of the cost of numerical solvers; however, accurate prediction of complex flows remains challenging, and no existing architecture leads across them. In such networks, the multigrid analogy mainly concerns the multi-resolution topology, while their components largely follow common vision practice. To address this, this study introduces SUMNO, a multigrid neural operator for fluid flow fields on regular grids that spatially unrolls a V-cycle into a feed-forward U-shaped network. Parameter-free down- and upsampling stand in for restriction and prolongation, and additive skip connections for coarse-grid correction, with normalization-free transfer and skip pathways. Within this macro-architecture, convolutions are used at the fine and intermediate resolutions and shifted-window attention at the coarsest, whose window partitioning is generalized to handle periodicity independently per axis. Evaluated on unseen test configurations across four unsteady and steady flow benchmarks against five representative baselines, SUMNO is best in 5 of 7 settings and ranks second in the remaining two. It reduces the 10-step rollout MSE by about 30% relative to the strongest baseline, and outperforms three baselines with over 130 million parameters using 9.4 million parameters.
Abstract (translated)
URL
https://arxiv.org/abs/2611.16341
PDF
-
From More to Better: Causal Inference and Digital Experimentation in the Age of AI
2026-11-12 05:56:52Winston Chou, M{\aa}rten Schultzberg, Lo-Hua Yuan, Wenjing Zheng, Avi Feller, David Holtz, Widad Machmouchi, Sven Schmit, Gagan Bhat, Simon Ejdemyr, Veronica Hannan, Amy Hartwig, Mahnaz Islam, Nathan Kallus, Jeremy Karp, Kyle Kretschman, Hannah Li, James McQueen, Evan Miller, Nicolas Stier, Sean TaylorAbstract
Artificial intelligence (AI) is quickly expanding the capacity of organizations to produce and analyze data. We argue that this expansion shifts the leverage of data scientists from performing individual analyses to assessing the credibility and importance of findings and determining when evidence generalizes across contexts. Drawing on a roundtable of academic researchers and industry practitioners, we examine three ways in which data scientists who work on causal inference and experimentation can help organizations turn more analytical capacity into better decisions: preserving and accumulating the learning value of experimentation beyond individual tests; choosing and combining experimental, observational, and synthetic evidence; and ensuring evaluation keeps pace with AI as its capabilities and uses evolve. In each area, we distinguish established practice from open questions. We conclude with implications for the training and practice of data science and for organizations.
Abstract (translated)
URL
https://arxiv.org/abs/2611.16335
PDF
-
Benchmarking AI risk behaviors
2026-11-12 05:27:44Luhan Mikaelson, Robert Adragna, Emmett Brosowsky, Hayley ClatterbuckAbstract
We evaluate LLM risk propensities across a range of contexts and task demands. We evaluate models for how risk averse or risk tolerant they are in various conditions, as well as how stable and consistent their behavior is across conditions. We present a toolkit for developing model risk profiles that can be used to predict model behavior and which can be adapted to include new evaluations of risk propensities. We also examine how these properties have changed across model generations and families. We observe that 2026 state-of-the-art models tend to be more stable, more consistent, and more risk neutral than older models. These trends appear to be related; newer models are more stable and consistent because they are more likely to use expected value maximization as a decision procedure.
Abstract (translated)
URL
https://arxiv.org/abs/2611.16318
PDF
-
Wanda-Guided Sparse Perturbations for Efficient Zeroth-Order Optimization
2026-11-12 05:25:58Ziqing Wen, Xiaohui Song, Jiahuan Wang, Ping Luo, Junlin Zeng, Dongsheng Li, Tao SunAbstract
Zeroth-order (ZO) optimization enables memory-efficient fine-tuning of large language models through forward-only evaluations, but high-dimensional random perturbations can limit optimization efficiency. Sparse perturbations reduce the search dimension, making the selection of informative coordinates critical. We propose Wanda-Importance Sparsity-MeZO (WIS-MeZO), which uses Wanda weight--activation saliency to construct task-conditioned sparse perturbation masks while keeping the pretrained model dense. We analyze support quality through gradient energy capture and relate it to gradient estimation and convergence. Across 16 benchmarks with LLaMA and Qwen models ranging from 3B to 32B parameters, WIS-MeZO achieves higher aggregate scores than MeZO and reaches comparable validation losses in fewer optimization steps. Comparisons with random and magnitude-based sparse perturbations further highlight the importance of informative support selection.
Abstract (translated)
URL
https://arxiv.org/abs/2611.16317
PDF
-
Shared Item-Level Error Limits the Aggregation of Small Language Models in Fermi Estimation
Abstract
Multi-agent systems of language models are often motivated by an averaging argument: if the agents' errors are unbiased and independent, the error of an aggregate of N answers falls as 1/sqrt(N). Fermi estimation, whose continuous answers are aggregated on the logarithmic scale, permits a direct test. We decompose the log10 error of each answer into the mean error of its model on the item, common to all its draws on the item, and a residual, and estimate both from 2.51M answers given before any communication by 15 open-weight models (3B-72B) on realFP, with replication under a single prompt pair on realFP, Science Olympiad Fermi questions and a guesstimation set. Given the item, teammates' errors are uncorrelated and teams behave as independent draws, but the item-level component accounts for a median 0.906 of the error variance over items. Teams of thirty therefore reduce a single agent's error by a median 1.7% (at most 5.4%), whereas draws that were also unbiased on each item would reduce it by 74-79%. All models compress their estimates toward about 10^6; this trend explains a median 17% of the item-level component, and a cross-validated linear recalibration on labelled items removes the overall bias but not that component. Three rounds of debate reduce within-team dispersion by 93% and leave the item-level component unchanged; revision without peers attains the same small gain. Mixing five models of similar accuracy reduces error by 6.6% relative to their own homogeneous teams, but no mixture improves on the most accurate model. Higher sampling temperature changes team error by about 1% at most, and contrastive activation steering not detectably; per-agent scaffold prompts (not significant after multiplicity correction) and skipping one layer in a single-model screen improve it. The results are conditional on the models, benchmarks, prompts and settings tested.
Abstract (translated)
URL
https://arxiv.org/abs/2611.16315
PDF
-
Persona Following Is Not Selective Control: The Neutrality Gap in LLM User Simulation
2026-11-12 05:17:19Jiashen Ren, Wenlin Zhang, Bohan Zhang, Xiaopeng Li, Zichuan Fu, Wanyu Wang, Junyi Li, Xiangyu ZhaoAbstract
Persona prompting is widely used in LLM user simulation, yet it relies on an untested assumption: specifying a target user attribute should shift target behavior while leaving other fixed attributes unchanged. Across eight black-box and five open-weight models, however, changing a target attribute alone also shifts unspecified non-target choices, such as risk preference shifting color choice. We test whether explicitly specifying non-target attributes restores selective control across 477 held-out synthetic worlds and 200 independent items under unspecified, directional, and neutral states. Assigning a clear non-target direction nearly eliminates cross-attribute shifts. Explicitly declaring no preference, however, still leaves 37.7%-76.3% of held-out worlds and 51%-81% of independent non-target items shifting with the target attribute while preserving target efficacy. We call this disparity the neutrality gap. Because directional tests can be passed simply by following a stated preference, the neutral state reveals selective-control failures they miss: following a stated persona does not imply keeping unrelated choices invariant. In simulated investment decisions, incidental brochure color labels shift estimated target risk treatment effects by 25.5 percentage points, even when models accurately report having no color preference. Counterfactual attention interventions show that unwanted influence enters primarily at query time, when option evaluation reads the target persona directly. Blocking query-time attention cuts non-target drift by 72.1%-87.8%, whereas directional declarations suppress drift primarily by overwriting option values rather than rerouting attention. Because target and non-target choices share query-time access, blocking it weakens target efficacy while value overwrite imposes an artificial direction, leaving task-dependent selective access as a central challenge.
Abstract (translated)
URL
https://arxiv.org/abs/2609.35036
PDF
-
ORCHID-SAGE: Fully Reproducible CALPHAD Assessment from Evidence to Phase Diagram
Abstract
Knowledge of phase stability is the first requirement for materials design, and predicting the conditions under which new materials can be synthesized has become a key capability of self-driving laboratories (SDLs). For inorganic materials, notably alloys, these predictions come from thermodynamic databases built with the CALPHAD method. Building a database, however, requires expert judgment at many steps, and because these decisions are rarely recorded, others can use a database but cannot regenerate or update it. Machines can already collect the data, build a first model with machine-learned interatomic potentials and refine it against experiments. The assessment itself, the sequence of judgments that turns these parts into a reliable database, is still carried out by human experts. We argue that it can be automated reproducibly only if the expert knowledge behind an assessment is encoded in formal protocols, so that an assessment delivers a record of its construction and not only its parameters. We introduce ORCHID-SAGE, a standard that organizes this record into reproducible and automatable protocols. In this first attempt, applied to Cr-Ni, the database was regenerated from its record by a script on a second machine, by an AI agent working only through the Model Context Protocol (MCP) and by a refit from public sources only, with the same predictions. Recorded in this way, a database need no longer be fixed at publication but can be updated by AI agents as SDLs and other sources provide new evidence, while the judgment remains with people and on the record.
Abstract (translated)
URL
https://arxiv.org/abs/2611.16308
PDF
-
SAGE-QA: Semantic-Aware, Grounded Evaluation for Closed-Context Question Answering on SQuAD 2.0
2026-11-12 04:56:28Chandra N Sekharan, Vishakan Umapathy, Abhishek PhadkeAbstract
Closed-context question answering (QA) requires a system to generate an answer strictly supported by a supplied passage, or to abstain when the passage lacks adequate evidence. SQuAD 2.0 is a widely used benchmark for this setting, scored by Exact Match (EM) i.e., character-level agreement with human reference spans after lexical normalization. For generative large language models (LLMs), however, an EM failure may reflect a valid paraphrase or an alternative passage-supported answer rather than a hallucination. The best prior leaderboard EM is approximately 91%, but evaluating generative models requires separating lexical divergence from factual error. We introduce SAGE-QA (Semantic-Aware, Grounded Evaluation for Question Answering), a multi-tiered framework combining reference-based Semantic Equivalence Scoring (SES) and passage-based Re-Validation Scoring (RVS). On a fixed set of 10,000 SQuAD 2.0 validation examples, we compute EM and decision-matrix F1 from TP, TN, FP, and FN counts for zero-shot closed-weight frontier models and for open-weight models, untuned and fine-tuned. Zero-shot, Claude Opus 4.6 and Gemini 3.1 Pro reach 85.21 and 86.66 EM, outperforming untuned Llama 3 8B (56.88), gpt-oss-20b (66.28), and Llama 3.3 70B (71.39). Parameter-efficient fine-tuning adds 21.19 to 34.06 EM points, raising these models to 90.94, 87.47, and 94.24 EM, respectively. Auditing the frontier models with SAGE-QA and an external LLM judge yields dataset-wide SES scores of 92.81 for Claude Opus 4.6 and 91.67 for Gemini 3.1 Pro, and hybrid RVS scores, which combine benchmark credit with passage-based judge assessments, of 97.96 and 98.02. Fine-tuned open-weight models perform best, with fine-tuned Llama 3.3 70B surpassing the previous best known EM.
Abstract (translated)
URL
https://arxiv.org/abs/2611.16300
PDF
-
Eliciting and Mitigating Specification Satisficing in Coding Agents
Abstract
Coding agents often fail to meet all requirements in the task description. In some of these failures, the agent considers an implementation choice that meets a requirement, but later decides to commit to another one that does not meet the requirement for a reason it finds acceptable. We call this specification satisficing. It is challenging to identify and mitigate specification satisficing because the agent accepts such choices for its own reasons, and existing self-correction methods are biased toward them. We propose SatisFix, which guides a coding agent to report these choices and their reasons, and then check each choice against the task description. We evaluate SatisFix on DeepSWE and SWE-bench Pro with five LLMs as backbones and two different agent frameworks, and compare it with three self-correction methods. SatisFix improves the pass@1 of the base agent by an average of 5.09% on DeepSWE and 7.06% on the hard subset of SWE-bench Pro. On DeepSWE, it outperforms the best baseline by 3.98%. Our ablation study shows that both steps of SatisFix contribute to the improvement.
Abstract (translated)
URL
https://arxiv.org/abs/2611.16289
PDF
-
CARE-PO: Controllable Tool-Call Efficiency for LLM Agents
2026-11-12 04:22:06Xiaoyang Ye, Xiao Yao, Hairui Zhu, Yiying Yang, Bei Gan, Jinliang Lu, Tengjin Weng, Yuan Yao, Wenhao JiangAbstract
External tool calls are a major running cost of an LLM agent, while its training objectives align almost exclusively with task completion. In a step-level diagnosis of Qwen3.5 trajectories at two model sizes on GAIA, a long-horizon tool-use benchmark, we find that most contain removable calls, at a similar rate at both sizes. Some methods add an efficiency reward to the objective to reduce the call count, but introduce problems of their own: the anchor they price calls against is computed from the policy's own rollouts (e.g., the group's shortest correct rollout) and so moves with the policy; training can drift toward a collapsed policy that issues no tool call at all, a degeneracy seen in published reports and in our 27B baseline; and a training run delivers a single operating point, so the user cannot trade efficiency against capability. We present CARE-PO (Controllable Anchored Relative Efficiency Policy Optimization) to mitigate these problems. The efficiency reward prices calls against a reference budget anchored to the task's own difficulty and fixed before training. Injecting a successful trajectory from the training data into each online GRPO rollout group and charging for evidence a trajectory never gathers ease the drift toward a collapsed call count. Shaping the cost curve by an effort tag in the prompt yields an efficiency and a capability operating point from a single training run. On GAIA, five of the six operating points across three model sizes improve on their base model in success rate and call count at once. On four held-out multi-hop benchmarks the result is size-dependent: at 27B and 4B success rate is held or raised at a lower call count, while at 9B both operating points trade success rate for fewer calls. The call count remains well above zero throughout training at all three sizes.
Abstract (translated)
URL
https://arxiv.org/abs/2611.12125
PDF
-
SMIC: Shared Multiscale Input-Reference Conditioning across Transformer Layers for Natural Language Inference
2026-11-12 04:15:47Xin Zhang, Victor S. ShengAbstract
Natural Language Inference models can be sensitive to localized lexical changes. We introduce Shared Multiscale Input-Reference Conditioning (SMIC), which constructs a single input-dependent multiscale reference and reuses it across Transformer layers. Each layer applies its own multiscale transformation to the shared reference and injects the resulting conditioning feature before the attention update. We instantiate SMIC in a token--context Transformer (TCT) and conduct from-scratch comparisons against decoder-based and Multi-Layer Perceptron (MLP) baselines across all three ANLI rounds. TCT-SMIC attains the highest observed ten-seed mean accuracy on all three ANLI rounds. The ablations further support the contribution of SMIC. We evaluate a real-valued Llama-style plug-in and extend deletion/insertion diagnostics to pretrained ModernBERT encoder and Llama decoder models with and without SMIC. Across the evaluated perturbation protocols, TCT-SMIC exhibits comparatively low pooled-attention redistribution and smaller signed accuracy changes. We further establish a finite-sample bound on average subset-mass redistribution in terms of the mean Jensen--Shannon divergence between the restricted and renormalized attention distributions.
Abstract (translated)
URL
https://arxiv.org/abs/2611.16270
PDF
-
Causals: An Open Knowledge Graph for Research on Causal Reasoning
2026-11-12 04:13:29Miaozhe Lin, Qiongzheng Lin, Zhongwei Liang, Zijun LiuAbstract
Applying causal knowledge to a new event requires connecting the event to relevant entities and the relationships through which its consequences may unfold. We present Causals, an open knowledge graph for research on this form of reasoning. Its initial release, constructed from news reports, contains 369,184 nodes and 1,272,418 edges linking reported events, organizations, commodities, financial instruments, and other entities. Structural relationships, observations, and attributed causal claims retain distinct semantics and links to source records. We describe the construction and representation of the resource and characterize its connectivity and multi-hop retrieval properties. Causals provides shared relational knowledge for studying how language models retrieve relevant context and apply mechanisms to particular events. The dataset is available under CC BY 4.0 at this https URL, and we invite contributions to its ontology, records, and evidence.
Abstract (translated)
URL
https://arxiv.org/abs/2611.16267
PDF
-
EmbodiedRSI: Active Continual Robot Learning Through Hypothesis-Guided Co-Evolution
2026-11-12 04:07:56Python Song, Zhixuan Liang, Kelsey Fu, Mengdi Wang, Junfeng Yang, Shilong LiuAbstract
Robot foundation models provide strong visuomotor control, yet their performance can degrade when object positions or task instructions change. Further improvements often require post-training on substantial robot data, which can be costly to collect through methods such as teleoperation. Agentic harnesses can adapt around the model, but current self-evolving harnesses use robot trials inefficiently when deciding which code and skill changes to pursue. We introduce EmbodiedRSI, a self-evolving agentic harness that autonomously decides where to explore next and turns the resulting physical interaction into improved code and skills. EmbodiedRSI realizes this through a Fast-Slow Dual-System Architecture, in which competing code and skill hypotheses are maintained in a Hypothesis Graph. Value-of-Information Experiment Selection chooses physical experiments that can distinguish these hypotheses. Their outcomes guide Code-Skill Co-Evolution. The Slow System builds Hierarchical Memory, and Reward-Grounded Memory Learning selects effective memory according to their value for later Fast-System improvement. On RoboCasa365, EmbodiedRSI reaches 77.0% overall success and 71.3% on Composite-Unseen, compared with 40.1% for the best baseline. EmbodiedRSI also reaches 86.8% overall success on LIBERO-Pro. Beyond benchmark performance, EmbodiedRSI transfers zero-shot to real-world robot, achieving 71.3% overall success across multiple challenging tasks.
Abstract (translated)
URL
https://arxiv.org/abs/2610.10498
PDF
-
Rectified Looped Transformer
2026-11-12 04:07:40Musheng Li, Jiawei Zhang, Ziyuan Liu, Yuantao GuAbstract
Looped Transformers increase effective depth by repeatedly applying a shared block, offering a parameter-efficient approach to language modeling. However, most existing works directly reuse each loop's output as the next's input, which can lead to hidden-state scale growth and perturbation amplification as recurrence deepens. We propose Rectified Looped Transformer (RLT), which redesigns feedback between loops using an interpolation schedule inspired by flow-based modeling. At each loop, RLT forms the next input by linearly interpolating between the current state and its prediction. Such interpolation distributes state changes across loops, encouraging a smoother trajectory from the initial input to the final output. Theoretical analysis proves that, under appropriate assumptions, RLT keeps state magnitudes uniformly bounded across loops and reduces the worst-case perturbation amplification. We evaluate RLT on language modeling corpora and downstream benchmarks. At matched training and inference depths, RLT consistently achieves lower perplexity than a standard looped Transformer, while exhibiting greater robustness to input perturbations and variance in inference depth. Consistent with the theoretical results, latent trajectory analysis shows that RLT maintains more stable hidden-state norms across loops, which also facilitates extrapolation-based decoding strategies. Notably, RLT achieves these gains without requiring additional trainable parameters or extra recurrent block evaluations, thus providing a simple yet more stable feedback mechanism for looped Transformers.
Abstract (translated)
URL
https://arxiv.org/abs/2611.16265
PDF
-
Label-Efficient Document-Grounded Factuality Detection via Cross-Paraphrastic Invariance Learning
2026-11-12 04:06:48Chao Chen, Shanshan Lin, Dongsheng Hong, Sibo Ju, Wenjie Yang, Sihong Xie, Xiangwen LiaoAbstract
Large language models (LLMs) can generate fluent claims that are unsupported by a given source document, motivating accurate and efficient document-grounded factuality detection. Existing LLM-based evaluators often require repeated inference and incur high computational cost, whereas compact factuality detectors are efficient at inference time but commonly rely on large labeled or synthetically generated corpora. We propose CPIL (Cross-Paraphrastic Invariance Learning), a two-stage Siamese framework that converts a small labeled set into relational supervision without acquiring additional factuality labels. CPIL constructs positive pairs from distinct label-preserving paraphrastic views of the same document-claim instance and negative pairs from opposite-label examples associated with the same source document. It further introduces quantile-aware semi-hard negative mining, which derives a boundary from the positive-pair similarity distribution and balances informative negatives with broader coverage. In the two-stage training procedure, Stage 1 learns an encoder from these task-aware pairs, and Stage 2 jointly fine-tunes the encoder and a linear classification head for binary factuality prediction. At inference time, CPIL requires no paraphrase generation, pair construction, or Siamese comparison and uses only one encoder-classifier forward pass. On the 11-task LLM-AggreFact benchmark, CPIL-2% and CPIL-3% achieve macro-F1 scores of 81.04 and 81.60 while using only 2-3% of the FactCG-released labeled pool. Both settings outperform all compared compact detectors and two vanilla LLM references. Ablations and sensitivity analyses support the contributions of task-aware pair construction, dynamic negative mining, and multi-pivot paraphrasing.
Abstract (translated)
URL
https://arxiv.org/abs/2611.16264
PDF
-
Indic-Task Gains and Regressions after Reinforcement Learning
2026-11-12 03:21:07Sutanshu Saksena Raj, Daksh Maan, Avinash Benki, Parijat Chatterjee, Pasunuti Prasanjith, Purusharth MalikAbstract
We run reinforcement learning (RL) with GRPO across four environments, starting from a supervised fine-tuned 30B-A3B mixture-of-experts model covering English and eight Indic languages. We compare the model before and after RL across 34 benchmarks in one evaluation campaign. MATH improves by 36.07 points, and most Indic benchmarks also improve. Yet XQuAD-Indic F1 falls 19.32 points as English answers rise from 8.7% to 91.7%; IndicQA falls 8.64, while English XQuAD gains 11.71 F1. A second RL run from the same supervised initialization adds an Indic environment together with language and repetition rewards. Compared with the released model, it improves both passage-answering and instruction-following benchmarks, along with preserving MATH performance. But it loses ground on reasoning and coding benchmarks. Several training choices changed together, so we cannot isolate the effect of any one of the reward terms. Script compliance remains unresolved. Adding a language reward did not ensure use of the requested script. Passage answering and script compliance therefore need direct evaluation, even when other benchmark scores improve.
Abstract (translated)
URL
https://arxiv.org/abs/2611.16228
PDF
