# Research Papers — 2026-10 · bundle page 15 of 156

> Full-text markdown of papers #701–750 of 7799 published in 2026-10, newest first. AI-readable mirrors of the full 7799-paper 2026-10 archive (50 papers per file).

Canonical page: https://aipod.dev/papers/archive/2026-10/page/15
[← Previous bundle](https://aipod.dev/papers/archive/2026-10/page/14.md)
[Next bundle →](https://aipod.dev/papers/archive/2026-10/page/16.md)

### A Self-Improving Agent that Grows its Own DSL: Oracle-Free Operator Induction for Exact Question Answering

Authors: Yun-Shiuan Chuang, Yirou Ge, Yixi Li, Yiming Bian, Zijie Pan, Jenna L. Ballard, Uma Kona, Linsey Pang, Prakhar Mehrotra
Published: 2026-10-14
Categories: cs.AI, cs.CL
arXiv: https://arxiv.org/abs/2610.20809v1

A self-improving agent has to decide which new skills to keep, and in practice it keeps whatever earns a high reward, convinces a judge model, or compresses its library, even when the skill is wrong. That is fine until the answers have to be exactly right: over a table of transactions, a total is either right or wrong. We ask a harder question: can an agent grow what it can compute with no reward, no judge, and no answer key, and still be exactly right? A single frozen model runs two nested loops over a small, exact domain-specific language (DSL). The inner loop answers new questions only when two independently written expressions agree on every record. The outer loop discovers its own missing capability and authors an operator for it, admitted only when an independent second implementation agrees and the result obeys a conservation law, all without ground truth. The inner loop roughly triples the answerable-question set, and the outer loop self-extends on disjoint starting sets, inventing operators outside the starting closure; a no-induction control answers none of the new questions. The admission rule is empirically sound: it rejects 138 of 138 buggy operator variants under an independent reference, even one written by a different model family, where the admit-if-it-runs, self-consistency, and compression signals the field relies on admit most of them. Unmodified, the same mechanism self-extends on two further domains with unrelated schemas, business-ledger and clinical-encounter question answering, growing their answerable sets more than tenfold and nearly twentyfold. The loop transfers across model families too: run with further proposers, the gate stays sound and independent families self-extend comparably, with only the induction yield tracking each proposer's DSL fluency.

---

### PRISM: Transfer and Limits of Activation Interventions in Tool-Calling Language Models

Authors: Giovana Kerche Bonás, Thales Sales Almeida, Marcelo da Silva Reis, Rodrigo Nogueira, Marcos Meideiros Raimundo
Published: 2026-10-14
Categories: cs.CL
arXiv: https://arxiv.org/abs/2610.20808v1

Recent literature shows that unsafe tool calls can often accompany textual refusals in language models. To investigate this hidden vulnerability, we introduce PRISM, a framework for evaluating the transfer of activation interventions across precision, languages, and scales. Our findings demonstrate that leveraging lower-resource settings (specifically quantized NF4 labels and cross-lingual transfer) offers a robust pathway to expose and improve model alignment. Notably, using behavioral labels from an NF4 copy enables stable intervention estimation even with very few scenarios. Furthermore, these alignment improvements transfer effectively across languages: five-language NF4-supervised directions successfully transfer to held-out Arabic, reducing unsafe violations in Llama-3.1-8B by 21.95 percentage points without target-language tuning. By mapping paired benign activations, we also successfully transfer these safety interventions to much larger models (up to 70B parameters). Ultimately, PRISM highlights how auxiliary supervision from compressed models and multilingual data can uncover and mitigate hazardous tool-calling behaviors across different model families. We release all experimental code and supporting artifacts at https://github.com/GiKerche/prism, alongside an interactive explorer of recorded model responses and experimental results at https://gikerche.github.io/prism/.

---

### When Reasoning Traces Conflict with Evidence: How Answer Configuration Shapes Evidence Following

Authors: Shaoang Li, Yanhang Shi, Jian Li
Published: 2026-10-14
Categories: cs.AI, cs.CL
arXiv: https://arxiv.org/abs/2610.20805v1

A reasoning trace retained at answer time plays two roles: it can raise accuracy, and it can compete with the current evidence when the two disagree. We disentangle these roles by crossing verified evidence versions with the traces generated from each and answering from the same traces and evidence under two configurations: continuation of the model's own thinking and a separate answer call. Across three model families (gpt-oss, Gemma4, Qwen3) on HotpotQA and MuSiQue, the separate call raises evidence-following over continuation by 29.9 to 60.6 points. On the same pairs, removing the trace lowers accuracy by similar amounts under both configurations. Yet continuation repeats the stale conclusion in most conflicts, and the separate call follows the current evidence far more often. Benefiting from a trace and deferring to it are therefore distinct properties. Further analyses show that the final line of a trace carries most of its value, gold evidence helps most before the trace is written, distractors weaken correction under a separate call, and renewed reasoning, unlike a verification instruction, follows the current evidence in most conflicts. Because trace deletion alone cannot reveal which source an answer follows, evaluations should report both properties together with the answer configuration.

---

### Optimal Explicit and Implicit Regularization Strengths in High-Dimensional Continual Linear Regression

Authors: Gilad Karpel, Daniel Soudry
Published: 2026-10-14
Categories: cs.LG
arXiv: https://arxiv.org/abs/2610.20804v1

We study continual linear regression in a high-dimensional regime, where a learner observes a sequence of tasks and updates its parameters sequentially. In this setting, we derive an explicit closed-form expression for the generalization loss of a model trained with gradient descent for a fixed number of steps per task, which holds for arbitrary teachers. We prove that the optimal fixed step size decays nearly inversely with the number of tasks, specifically as $\ln T / T$, for a single teacher and, more generally, i.i.d. teachers. Under this scaling, we show that the resulting generalization loss is asymptotically equivalent to that achieved by optimally tuned L2 (isotropic) regularization. Furthermore, we extend our analysis to generalized ridge regression with diagonal regularization matrices, obtaining a closed-form characterization of the generalization loss and identifying necessary and sufficient conditions under which the optimal regularizer is isotropic. Finally, we identify a regime where the optimal step size remains finite even as the number of tasks tends to infinity, thus revealing a non-trivial tradeoff between task variability and implicit regularization, in contrast to the i.i.d. teacher scaling. Our theoretical predictions are corroborated by experiments on both linear models and neural networks, providing practical guidance for the design of continual learning systems.

---

### Privacy-Preserving Text Inference for Health-Misinformation Classification and Semantic Retrieval

Authors: Jasleen Kaur, Chloé Cachet, Koray Karabina, Jean Belo Klamti, Philippe Lamontagne
Published: 2026-10-14
Categories: cs.CR, cs.LG
arXiv: https://arxiv.org/abs/2610.20796v1

Health-misinformation analysis presents a dual challenge: misleading claims can affect public well-being, while systems designed to detect them may expose sensitive user-submitted text. This motivates privacy-preserving inference for public-health applications in which user queries may contain confidential medical information and trained models and reference databases are maintained on organizational servers. We design a privacy-preserving text inference framework based on CKKS homomorphic encryption for health-misinformation classification and semantic retrieval. The classification pipeline fuses BioBERT embeddings, PCA-reduced MPNet embeddings, and SVD-reduced TF-IDF features, while the retrieval pipeline combines MPNet embeddings with ISCC-SCT semantic fingerprints through late fusion. We evaluate predictive performance and benchmark plaintext versus ciphertext-based execution. The hybrid classifier achieved an accuracy of 0.971 and a macro-F1 score of 0.950, with identical reported metrics under plaintext and ciphertext-based inference. The hybrid late-fusion similarity model achieved a Pearson correlation of 0.8691 with human sentence-pair similarity scores on STS-B. For ciphertext-based retrieval, we evaluated naive and optimized implementations. The optimized implementation uses CKKS single instruction, multiple data (SIMD) packing to compute and return multiple similarity scores together, reducing single-query latency over 7,128 indexed items from 222.028 to 5.168 seconds and communication overhead from 6.11 GB to 2.13 MB. These findings demonstrate the technical feasibility of privacy-preserving text inference for health-misinformation applications and characterize its predictive utility and efficiency trade-offs.

---

### EEG-Based Prediction of Music Enjoyment Using Deep Learning with Explainable Temporal Analysis

Authors: Parshav Singla, Simranjit Kaur, Maanya Walia, Niharika
Published: 2026-10-14
Categories: cs.LG, cs.HC, cs.SD
arXiv: https://arxiv.org/abs/2610.20795v1

Understanding why listeners enjoy music is difficult because the experience is subjective and changes as a piece unfolds. Electroencephalography (EEG) offers a non-invasive way to study neural responses during listening, yet earlier work has mostly treated music emotion as a categorical label or relied on black-box models. We formulate enjoyment prediction as a regression problem on the public MUSIN-G dataset, which contains 128-channel EEG from 20 listeners together with song-level enjoyment ratings. Six deep learning models were compared under a participant-wise protocol, in which 15 listeners were used for training and validation and 5 were held out for testing. DeepConvNet, CNN-LSTM and GNN-CNN reached the lowest mean absolute error (0.97 to 0.99 on the 1 to 5 scale), while the temporal convolutional network (TCN) had the highest Pearson correlation with the ratings (r = 0.21). Gradient-based and occlusion-based explanations, averaged over all held-out windows, identify the time segments within each two-second window that most influence the TCN predictions. The study provides a reproducible cross-listener benchmark for EEG-based enjoyment regression and shows how explainability analysis complements standard metrics.

---

### Blame Errant Words, Not the Canvas: Defect-Aware Spatial Credit Assignment for Visual Text Generation

Authors: Sujie Hu, Huaisong Zhang, Pujing Yang, Qian Wang, Xiaomin Li, Xinhang Leng, Chen Li, Jing Lyu, Xiu Li, Ling Pan
Published: 2026-10-14
Categories: cs.CV
arXiv: https://arxiv.org/abs/2610.20794v1

Recent reinforcement learning approaches have improved visual text generation, yet generated images still suffer from incorrect wording and malformed glyphs. We identify two coupled limitations: recognition-based rewards often overlook visible glyph defects, while whole-image scalar feedback obscures the uneven quality of individual text regions. These limitations motivate our central principle: $\textit{Blame errant words, not the canvas}$, to effectively preserve local quality distinctions. First, we introduce $\mathbf{TextDART}$, a defect-aware region-targeted reward model that localizes individual text regions and independently evaluates their semantic correctness and glyph quality, providing fine-grained, region-precise feedback. Building on TextDART, we propose $\mathbf{SCANNER}$, which turns these regional judgments into spatially differentiated policy feedback. It normalizes rewards for each text target across sampled images and casts the resulting regional advantages onto the corresponding latent tokens. Extensive experiments on existing benchmarks and our newly established $\textbf{PolyText-Bench}$ demonstrate that our approach achieves state-of-the-art performance in text accuracy and glyph fidelity. Project page: https://sujiehu.github.io/textdart-scanner

---

### Sliced Wasserstein Sketches with Error Certificates

Authors: Elaheh Akbari, James Liu, Keaton Hamm, Soheil Kolouri
Published: 2026-10-14
Categories: cs.LG
arXiv: https://arxiv.org/abs/2610.20788v1

Compact representations that support geometrically meaningful comparisons are valuable for learning from probability distributions. We study optimal, free-weight $M$-atom quantization of one-dimensional probability measures in quadratic Wasserstein distance and extend it to higher dimensions via slicing. Using $L$ shared projections yields a sketch with $O(LM)$ comparison time. Our main result is an exact decomposition of squared distance error into residual and partition-mismatch terms, yielding computable certificates for arbitrary centroid sketches. For optimal sketches, a finite $(2+δ)$ moment for some $δ>0$ guarantees $O(M^{-1})$ error in the finite-direction sliced distance, while equal-mass sketches can converge more slowly. Each slice preserves its mean, while storing residual energies enables covariance recovery from sufficiently many generic directions. Experiments on document, point-cloud, and flow-cytometry benchmarks show that optimal sketches substantially reduce storage while preserving the classification performance of uncompressed sliced distances. They also improve generalization from few labelled subjects and support end-to-end learning as compact pooling layers.

---

### EvolvingBrain: An Adaptive Skill System for Multi-modal Agents Self-Evolution in Open Worlds

Authors: Runxuan Liu, Jingqi Shi, Baoxin Wang, Honghong Zhao, Haoyuan Xu, Yueqing Sun, Jiafeng Liang, Tao He, Dayong Wu, Qi Gu, Shijin Wang, Ming Liu et al.
Published: 2026-10-14
Categories: cs.AI
arXiv: https://arxiv.org/abs/2610.20786v1

Multi-modal agents have made significant progress in interacting with open worlds. However, their ability to self-evolve remains limited. Supervised fine-tuning enables imitation but not active experience organization, while outcome-based reinforcement learning promotes exploration yet underuses intermediate trajectory information. Therefore, we propose EvolvingBrain, which uses an adaptive skill system as the brain for agent self-evolution, consisting of two following components. Stratified Skill Self-Evolution: extracts Procedural Patterns and Problematic Patterns from model-specific trajectories, summarizes experience by generating skills, and evolves through multiple interactions with the open world, ultimately forming a Stratified Skill Architecture consisting of Domain, Plan, and Action/Trap Skills, which provides the agent with a multi-granularity knowledge repository and a novel evolution approach. State-Aware Adaptive Execution: summarizes long-horizon interactions into Structured Abstractive Memory and uses it to guide the State-Aware Skill Fusion Strategy, enabling the agent to utilize knowledge according to its current state. Experiments on VAB-Minecraft and DiscoveryWorld demonstrate that EvolvingBrain improves success rates on open-world tasks and exhibits generalization capabilities. Our code is provided in the supplementary materials and will be publicly released in the camera-ready version.

---

### APRiL: An Open-ended Benchmark for Chinese Social Awareness in Large Language Models

Authors: Tian Lan, Han Zhang, Shanshan Wang, Yahan Yu, Jiang Li, Yugo Murawaki
Published: 2026-10-14
Categories: cs.CL
arXiv: https://arxiv.org/abs/2610.20785v1

Large language models are increasingly used in social interactions, where appropriate responses require recognizing implicit cues, inferring mental states, identifying conflicts, and considering potential consequences. However, existing benchmarks often focus on specific social abilities, constrained task formats, or particular interaction settings, offering limited coverage of social awareness across diverse contexts. To address this gap, we introduce APRiL, an open-ended benchmark for Chinese social awareness with instance-specific rubrics. APRiL contains 2,196 Social Awareness Evaluation Instances across six categories and two complementary tasks: Social Scenario Analysis evaluates models' understanding of social cues, mental states, and conflict structures, while Socially Aware Response assesses whether they can translate this understanding into appropriate, natural responses. Experiments on 12 models from six families reveal a consistent understanding-to-response gap. Higher-quality scenario analyses improve responses, while weak analyses may reduce response quality. APRiL provides a systematic resource for advancing socially aware language models in Chinese contexts. Our data and code are available at https://anonymous.4open.science/r/APRiL-05DE/.

---

### Open-world AI: a Tale of Three Requirements

Authors: Jianyu Zhang, Léon Bottou
Published: 2026-10-14
Categories: cs.AI
arXiv: https://arxiv.org/abs/2610.20782v1

The closed world asks AI for specialized competence and lets it manufacture the iid data its theory assumes: a fixed criterion licenses collecting examples until the task is covered. The open world asks for versatility --- taking up tasks a machine was not built for, from few examples and with little prior from a human designer --- and offers data that is grouped, not shuffled. Its tasks are not known in advance, so a machine must do inference-time learning (1st requirement). Modeling tasks as sharing latent factors and decomposing the training objective, we derive two more: a *rich representation*, retaining more than iid generalization requires, and a predictive disentanglement that aligns inputs and outputs into matched low-dimensional groups. The analysis further predicts the need for a consistent inner learner, and the benefit of long contexts spanning related tasks. We test these claims on language models up to 8B, asking whether a model has a property (qualitative) rather than how much it wins (quantitative). Models that meet the requirements improve as data from related tasks accumulates; the rest degrade, and neither scale nor layer-type interleaving repairs this.

---

### ReasonUni: Reasoning-Augmented Reconstruction Advances Unified Multimodal Models

Authors: Feize Wu, Zhiyuan Yan, Yuhuan Hong, Yuwei Niu, Zongjian Li, Zheyuan Liu, Bin Lin, Jinbo Yan, Jie Qin, Xudong Mao, Limeng Qiao, Guanglu Wan et al.
Published: 2026-10-14
Categories: cs.CV, cs.AI
arXiv: https://arxiv.org/abs/2610.20781v1

Image-to-text (I2T) understanding and text-to-image (T2I) generation are dual tasks in unified multimodal models (UMMs). Reconstructive post-training couples these tasks, but existing reconstruction-based methods remain perception-centric and provide limited supervision for reasoning about structure, intent, and transformation. We propose ReasonUni, a reasoning-augmented reconstructive post-training framework that extends this coupling beyond perception. Given an original image and an editing instruction, the UMM first reasons about what and how to edit and describes the target state, then generates the corresponding target image conditioned on this understanding output. To jointly optimize both modules, we introduce TreeGRPO, a tree-structured reinforcement learning method that samples multiple reasoning paths and candidate images per path, then aggregates visual rewards to update both modules under a unified objective. Experiments across multiple benchmarks demonstrate consistent gains in both understanding and generation, showing that reasoning-augmented reconstruction yields genuine mutual gains for UMMs.

---

### ARIA: Runtime Inspection and Amendment for Multi-Agent Systems

Authors: Bolin Shen, Yushun Dong
Published: 2026-10-14
Categories: cs.AI, cs.CL
arXiv: https://arxiv.org/abs/2610.20780v1

In LLM-based multi-agent systems (MAS), a local mistake can propagate as other agents build on it, turning a correctable error into a failed task. Diagnosing completed runs or repeatedly re-executing them to search for corrections does not provide a learned policy for intervening while execution is still in progress. We present Agent Runtime Inspection and Amendment (ARIA), a framework for learning when and how to repair ongoing multi-agent executions. Our key intuition is to turn execution-verified repairs into reusable supervision, moving repair search into data construction and training so that a learned model can intervene at runtime without repeatedly testing alternatives. First, we collect over 24{,}000 successful and failed trajectories across four multi-agent frameworks, six benchmarks, and eight host models. From these trajectories, we construct the ARIA dataset, to our knowledge the first large-scale dataset pairing failed multi-agent trajectories with execution-verified repairs. It contains 3{,}953 verified repair pairs, each specifying what went wrong, where to intervene, and how execution should continue. Then, we develop the ARIA framework for MAS-agnostic runtime inspection and amendment, designed to integrate with any MAS that exposes execution trajectories and supports continuation replacement. Using verified repairs as training examples, we teach a compact model to recognize when intervention is needed and how to correct an ongoing run. The aim is to learn repair strategies that apply to related failures in new configurations, rather than merely reproduce individual corrections. Finally, the trained ARIA model demonstrates strong repair performance in both post-hoc and runtime settings, with generalization to unseen frameworks and host models. Our code is available at https://github.com/LabRAI/ARIA.

---

### Verdict: Encode Once, Decide Many

Authors: Avdhesh Singh Chouhan, Jack Perry
Published: 2026-10-14
Categories: cs.CL, cs.AI, cs.LG
arXiv: https://arxiv.org/abs/2610.20778v1

Automating organizational work means answering many small typed questions about one piece of state (does this email need a reply, how urgent is it, who owns it) with probabilities good enough to act on without a human. The open typed-decision model we compare against, like a standard cross-encoder, answers each question with a separate pass over the input. We present Verdict-MQ, an open encoder that reads a case once and answers all of its questions: given identical retained state tokens after truncation (verified on every reported case), a structured attention mask makes the state encoding independent of the questions, keeps questions from influencing each other, and makes option order irrelevant by construction. On the typed-decisions benchmark, a 395M Verdict-MQ matches the soft gold labels significantly better than the released 421M Laya specialist on all five fresh seeds (KL 0.080 vs 0.117, Brier 0.045 vs 0.061), an advantage that survives per-system temperature scaling; its higher accuracy (0.801 vs 0.767) lies past the benchmark's own saturation point, so we read that margin as closer agreement with the benchmark's labelling teacher. On RACE, BoolQ and MultiRC, against a matched fine-tuned baseline it wins on MultiRC and loses on RACE and on BoolQ, where its training is unstable; we measure zero-shot references ourselves, including a closed general system (Jev) that beats every fine-tuned encoder. A controlled comparison measures what the shared state costs: giving each question its own question-aware copy of the state gains 0.9-3.5 points on RACE and MultiRC at 1.4-2.5x the per-case latency, and nothing measurable on typed decisions, with the invariance guarantees unchanged.

---

### INFUSE: A Unified Surface Benchmark for Indirect Prompt Injection in Agentic Systems

Authors: Thuy Dung Nguyen, Jingyi Guo, Rana Seda Dubauskas, Sreepriya Damuluru, Taylor T. Johnson, Kevin Leach
Published: 2026-10-14
Categories: cs.CR, cs.AI
arXiv: https://arxiv.org/abs/2610.20776v1

LLM-based agents increasingly execute long-horizon, end-to-end workflows involving shell commands, code modification, dependency installation, and interaction with remote services. This growing autonomy expands their exposure to indirect prompt injection (IPI), where adversarial instructions embedded in external content are interpreted as legitimate task directives. Existing evaluations, however, often focus on isolated malicious strings, single-step tool calls, or narrowly scoped agent configurations. We introduce INFUSE, an end-to-end benchmark for evaluating IPI throughout the lifecycle of long-horizon agent workflows. INFUSE comprises 40 realistic tasks and a structured taxonomy of 56 attack scenarios spanning five threat categories. Attacks are injected through realistic interaction surfaces encountered during task execution, extending evaluation beyond attacks placed only in initial prompts or isolated tool interactions. We evaluate 19 LLMs from the GPT, Claude, Gemini, Llama, Gemma, and Qwen families, producing 41,824 adversarial trajectories (26,220 without defenses) and 720 benign reference runs. Across the benchmark, mean attack success reaches 19.6% under LLM-based evaluation and 15.4% under heuristic verification, with per-model ASR-L ranging from 3.2% to 41.9%. Existing guard models provide limited protection, while prompt-level instructional hardening reduces mean attack success from 58.2% to 14.3%, revealing substantial but model-dependent security-utility trade-offs. Our code is available at https://github.com/judydnguyen/infuse.

---

### Selection-Safe Conformal Early Exit: Distribution-Free Coverage under Adaptive Depth

Authors: Adria Binte Habib, Ibne Farabi Shihab
Published: 2026-10-14
Categories: cs.LG, stat.ML
arXiv: https://arxiv.org/abs/2610.20773v1

Calibrating a conformal prediction set at each early exit does not preserve coverage once a data-dependent exit is selected. We construct one terminal path score from prefix maxima, calibrate a single split-conformal quantile, and return nested prefix sets that all contain the valid terminal set. The guarantee is selection-safe without an alpha/L correction and holds simultaneously for every post-calibration label-free stopping policy that preserves terminal-set containment. We also bound the efficiency cost of the path maximum by the empirical tail of the early-over-terminal score excess. Across eight encoder task-backbone cells, pathwise selection attains macro coverage 0.901 versus 0.814 for naive layerwise selection, with average set size 1.41 and latency fraction 0.71. On CIFAR-100 with a 24-exit ViT-L/16 (n_cal=500), first-eligible stopping at budget b=4 attains coverage 0.902, set size 2.76, and latency fraction 0.57, and the same calibration supports deployment-time budget changes without refitting. In a joint exit-count/calibration-size sweep at n=100, pathwise set size moves only from 3.02 to 3.11 as exits increase from 4 to 24, while Bonferroni degenerates to the full label space for 12 or more exits. Paired comparison against MinSE valid selection at (n,L)=(100,24) gives a set-size difference of 0.80 (0.47, 1.13). Full-vocabulary decoder experiments remain a negative boundary, with sets of 7,162-9,620 tokens. Pathwise calibration is most useful when candidate exits are many relative to calibration size, when budgets change after calibration, and when shallow heads return actionable sets.

---

### Semantic Alignment of EEG and Text: A Contrastive Learning Framework for Decoding Imagined Speech

Authors: Shahryar G. Namdari, Francesco Iacomi, Riccardo Barbieri, Luca Mainardi
Published: 2026-10-14
Categories: cs.LG, q-bio.NC
arXiv: https://arxiv.org/abs/2610.20766v1

Imagined speech decoding from EEG offers a promising route toward naturalistic brain-computer interfaces. Traditional closed-set classifiers treat semantic categories as independent labels and do not explicitly exploit relationships among linguistic intents, limiting output flexibility. We adapt a multimodal contrastive learning framework to imagined speech, aligning EEG trials with language representations in the embedding space of a frozen pretrained text encoder. An EEG encoder is trained with a composite objective combining semantically reweighted EEG-sentence alignment and text-prototype-based class discrimination, encouraging semantic structure while preserving separability among 39 intent categories. We evaluate the approach on the Chisco imagined speech dataset across five subjects, reporting Top-k accuracy, macro-F1, and balanced pairwise matching accuracy (2v2). We also investigate maximum EEG-text cosine similarity (MaxCosine) as an intrinsic indicator of decoding reliability and probe semantic structure beyond category labels using five evaluation-time semantic axes and a qualitative Semantic Gallery Probe (SGP) with previously unseen candidate sentences. The contrastive framework achieved decoding performance comparable to a task-optimized closed-set classifier while mapping EEG into a frozen language embedding space. MaxCosine identified subsets of held-out trials with higher Top-1 and Top-5 accuracy across subjects. Evaluation-time semantic axes showed mean agreement above the 50% chance reference across all five dimensions, and the SGP qualitatively illustrated direct comparison with unseen candidate sentences. Together, these findings suggest that EEG-language alignment can retain comparable fixed-intent decoding while supporting similarity-based retrieval, intrinsic reliability ranking, and evaluation of broader semantic structure.

---

### When Reality Outpaces Cognition: A New Scaling Principle for Reliable Intelligence

Authors: Ahsan Adeel, Talha Bin Riaz
Published: 2026-10-14
Categories: cs.AI, cs.LG, cs.RO
arXiv: https://arxiv.org/abs/2610.20764v1

This paper challenges the understanding capabilities of existing AI paradigms and emerging world models by asking a fundamental question: How quickly can reality change before an intelligent system can no longer reliably perceive, reason, and act? We formulate this as a theory- and model-agnostic challenge in which systems must infer reliable actions for increasingly distant future states across a growing observation--reality gap. We operationalize this challenge through the effective Speed of Coherent Cognition (\(S_C^{\mathrm{eff}}\)), defined as the highest tested Speed of Reality (\(S_R\)) through which a system continuously satisfies a predefined behavioural reliability criterion from baseline. As a proof of concept, reinforcement-learning agents are trained at the baseline \(S_R\) and evaluated, without retraining, as \(S_R\) is increased by up to \(50\times\) while computational resources remain fixed. We compare Transformer and gated-MLP baselines with \(Co^4\), an architecture inspired by cellular mechanisms implicated in conscious processing and awake thought. Across the tested scenarios, \(Co^4\) achieves substantially higher \(S_C^{\mathrm{eff}}\) than the baselines, motivating \(S_C^{\mathrm{eff}}\) as a new benchmark and the scaling of coherent cognition as a principle for reliable AI.

---

### Reference-Guided Turn-Taking Style Transfer for Full-Duplex Spoken Language Models

Authors: Yuxuan Chen, Yang Xiao, Jack Sun, Chak Sun, Brick Li, Sam
Published: 2026-10-14
Categories: cs.CL, cs.AI, cs.SD
arXiv: https://arxiv.org/abs/2610.20763v1

Full-duplex spoken language models decide at every frame whether to take the floor, defining a conversational personality -- yet each set of weights encodes exactly one, which human speakers bracket rather than share. Existing controls mostly act from outside; none identifies the quantity that sets how much the model speaks. We identify it: a single late-layer residual-stream direction, the initiative vector, extracted from the model's own conversations without training or annotation. The same direction reads the decision, anticipating onsets up to 800 ms ahead, and shifts it: scaled during generation, it moves speech ratio from 6-12% to 42-61% (no steering: 21-30%), consistently across conversations, weights, and corpora. A short reference recording suffices to transfer a speaker's turn-taking style to the model, and the transfer is reference-specific: swapping in another speaker's reference worsens the match in 93-98% of cases, whereas the best uniform dose is indistinguishable from no steering. Turn-taking style is not frozen into the weights but a direction that can be read, shifted, and calibrated per person. Audio examples are available at https://elpsykongloo.github.io/initiative-vector/.

---

### Purify Before You Align: Improving Representation Alignment for Diffusion Models

Authors: Yingheng Wang, Yaoqiang Li, Yaqin Wu, Junwen Bai, Jiatao Gu, Christopher De Sa, Volodymyr Kuleshov
Published: 2026-10-14
Categories: cs.CV, cs.LG
arXiv: https://arxiv.org/abs/2610.20761v1

Representation alignment improves and accelerates diffusion training by encouraging intermediate features to match a pretrained visual encoder. However, standard alignment matches pretrained representations without distinguishing what information may be more useful for generation. We introduce local appearance purification (LAP), which improves generation quality and makes encoders such as MAE more effective alignment targets by removing locally predictable content and aligning to the residual. LAP modifies only the alignment target, and requires no changes to diffusion training or inference. Through controlled interventions within a fixed encoder, we show that effective alignment depends on what information the target emphasizes, not simply how much image information it contains. Our theory gives a variational interpretation of standard alignment and, for fixed local predictors, a sum of token-wise information bounds conditioned on local appearance. To understand why purification benefits some encoders more than others, we derive an exact constant-readout test that measures how well the target can be matched without using the diffusion model's hidden state. LAP improves generation quality across five pretrained representations (MAE, AIMv2, CLIP, DINOv2, DINOv3), with the largest gain for MAE: on ImageNet-256 with SiT-B/2, it reduces FID from $53.1$ to $42.3$. Its gains persist with stronger alignment recipes and tokenizers. Computed before diffusion training, the constant-readout score predicts the benefit of purification (Spearman $ρ=0.94$ across 13 encoders).

---

### Optimal Hydra: Model Heterogeneous Federated Learning for LLMs using Data-Informed Hydra-Nets

Authors: Gwen Legate, Rahaf Aljundi, Eugene Belilovsky
Published: 2026-10-14
Categories: cs.LG, cs.AI
arXiv: https://arxiv.org/abs/2610.20753v1

Large language models (LLMs) trained on vast amounts of data have demonstrated impressive performance over a range of tasks, but future performance gains will require increasing amounts of data and compute. Federated Learning (FL) facilitates training on decentralized private data, but client resource limitations can make training impractical, particularly as model sizes continue to expand. These resource constraints make the assumption of a shared server-client architecture impractical. Model heterogeneous FL offers a promising alternative by adapting model size to client capabilities; however, previous work is limited to the pre-training setting, and conclusions on sub-network selection do not necessarily hold for fine-tuning. Furthermore, prior approaches create client sub-networks without considering their unique data distributions, thereby overlooking the opportunity to tailor sub-network architecture to local data. Motivated by evidence that attention heads exhibit task dependency within layer defined feature spaces, we propose Optimal Hydra, a transformer based heterogeneous FL framework for fine-tuning that combines global layer selection with client specific attention head selection, aligned with local data distributions. Optimal Hydra makes explicit use of client data to produce client-specific sub-networks we call Hydra-nets. To our knowledge, this is the first model heterogeneous FL approach tailored to LLM fine-tuning. Experiments on three datasets show significant improvements over baseline methods while also reducing local resource demands.

---

### Interpreting Language Model Parameters

Authors: Lucius Bushnaq, Dan Braun, Oliver Clive-Griffin, Bart Bussmann, Nathan Hu, Michael Ivanitskiy, Linda Linsefors, Lee Sharkey
Published: 2026-10-14
Categories: cs.LG, cs.AI
arXiv: https://arxiv.org/abs/2610.20748v1

Mechanistic interpretability aims to uncover how neural networks use their parameters to implement their impressive neural algorithms. Although previous work has uncovered substantial structure in the intermediate representations that networks use, little progress has been made to understand how the parameters and nonlinearities of networks perform computations on those representations. In this work, we present a method that brings us closer to this understanding by decomposing a language model's parameters into subcomponents that each implement only a small part of the model's learned algorithm, while simultaneously requiring only a small fraction of those subcomponents to account for the network's behavior on any input. The method, adVersarial Parameter Decomposition (VPD), optimizes for decompositions of neural network parameters into simple subcomponents that preserve the network's input-output behavior even when many subcomponents are ablated, including under ablations that are adversarially selected to destroy behavior. This encourages learning subcomponents that provide short, mechanistically faithful descriptions of the network's behavior that should aggregate appropriately into more global descriptions of the network's learned algorithm. We study how sequences of interactions between these parameter subcomponents produce the network's output on particular inputs, enabling a new kind of 'circuit' analysis. While more work remains to be done to deepen our understanding of how neural networks use their parameters to compute their behavior, our work suggests an approach to identify a small set of simple, mechanistically faithful subcomponents on which further mechanistic analysis can be based.

---

### Ergodic topology sampling toward the continuum with Stochastic Normalizing Flows

Authors: Claudio Bonanno, Andrea Bulgarelli, Elia Cellini, Alessandro Nada, Dario Panfalone, Davide Vadacchino, Lorenzo Verzichelli
Published: 2026-10-14
Categories: hep-lat, cond-mat.stat-mech, cs.LG, hep-ph
arXiv: https://arxiv.org/abs/2610.20747v1

In this work we provide numerical evidence that topological freezing in $\mathrm{SU}(3)$ Yang-Mills theory can be mitigated with a flow-based approach, as tested at lattice spacings as fine as 0.03~fm. We do so by analyzing the scaling of the costs of a computational strategy based on the combination of open boundary conditions and Stochastic Normalizing Flows. We show that the integrated autocorrelation time of the squared topological charge and the statistical uncertainty on the topological susceptibility remain under control throughout the explored parameter range with currently available computational resources. This marks an important step toward the use of even more efficient flow-based architectures for ergodic sampling of topology in state-of-the-art lattice gauge theory calculations.

---

### Vertical Sliced Wasserstein Distances for Meta-Measures

Authors: Elaheh Akbari, Ahmadreza Moradipari, Kyungtae Han, Hamed Pirsiavash, Yikun Bai, Rocio Diaz Martin, Soheil Kolouri
Published: 2026-10-14
Categories: cs.LG
arXiv: https://arxiv.org/abs/2610.20743v1

Many modern learning problems require comparing probability distributions whose individual observations are themselves probability measures, naturally giving rise to meta-measures in $\mathcal{P}(\mathcal{P}(Ω))$. However, optimal transport distances on this space can be prohibitively expensive to evaluate. We introduce the vertical sliced Wasserstein ($\mathrm{VSW}$) distance, a scalable metric on $\mathcal{P}(\mathcal{P}(Ω))$. Rather than slicing along the transport geometry of $\mathcal{P}(Ω)$, $\mathrm{VSW}$ projects each inner measure using continuous test functions and compares the resulting one-dimensional distributions with one-dimensional Wasserstein distance. We prove that $\mathrm{VSW}$ is a valid metric on $\mathcal{P}(\mathcal{P}(Ω))$, metrizes weak convergence, and admits dimension-independent statistical convergence guarantees. Empirically, $\mathrm{VSW}$ provides high accuracy and efficiency compared with existing distances for meta-measures. We further derive a practical particle gradient for the empirical $\mathrm{VSW}_2^2$ objective and demonstrate its use for gradient-based transport of meta-measures. $\mathrm{VSW}$ provides a simple and scalable alternative for comparing and optimizing distributions of distributions.

---

### Quantile-Guided Diffusion for Probabilistic Time-Series Forecasting

Authors: Shadi Omari, Bashar Khoury, Assaf Schuster
Published: 2026-10-14
Categories: cs.LG
arXiv: https://arxiv.org/abs/2610.20739v1

Quantile forecasts summarize uncertainty at each future time step, while diffusion forecasters represent it through sampled future trajectories. We study how predicted quantiles can serve as distributional controls for diffusion forecasting. Specifically, we investigate how the mechanism of quantile integration affects its effectiveness in guiding a pretrained diffusion forecaster. We find that residual control throughout the denoising network yields the most consistent improvements. We instantiate this approach as Control-TS, which learns a quantile-control pathway while keeping the pretrained backbone frozen. Across eight benchmarks, Control-TS consistently improves the backbone's CRPS, achieves competitive performance against state-of-the-art forecasting models, and retains its gains in capacity-matched comparisons. Further analyses examine how these gains depend on quantile granularity, predictor choice, and forecast quality. Finally, we introduce the Conditional Linear Predictive Score (CLPS), which complements distributional metrics by evaluating the downstream forecasting utility of generated futures. Together, these findings support predicted quantiles as an effective interface between explicit uncertainty estimates and diffusion-based trajectory generation.

---

### WorldCast: Exploring Intelligent Visual Generation in World State Prediction

Authors: Wenhao Yan, Haodong Li, Xiaotong Li, Yaxin Luo, Zhengrong Yue, Jialv Zou, Xu Huang, Haojin Yang, Nan Jia, Yan Feng, Peng Pei, Xunliang Cai
Published: 2026-10-14
Categories: cs.CV, cs.AI
arXiv: https://arxiv.org/abs/2610.20731v1

Humans interpret multimodal observations of the world and forecast how it will evolve. Recent multimodal generative systems have demonstrated similar capabilities: they can integrate multimodal context and synthesize visual outputs that depict changes in the world. To characterize this emerging trajectory, we organize model capabilities into a four-level hierarchy and discover that higher-level tasks remain relatively underexplored despite model progress. To bridge this gap, we explore model performance on multi-step world state prediction, a high-level task that requires models to forecast successive future world states from multimodal observations and diverse references. Such a task requires models to possess entangled understanding and generation capabilities, as well as extensive world knowledge to translate human action sequences, physical transformations, and abstract visual reasoning into coherent visual outputs. We introduce WorldCast, a benchmark comprising 320 manually designed cases across eight categories, and evaluate 12 representative models. The conducted experiments yield three key insights: (a) advances in frontier models are rapidly pushing the boundaries of intelligent world modeling, with the GPT-Image-2.5 System scoring 72.8 and demonstrating early-stage visual state prediction capabilities on real-world tasks; (b) WorldCast reveals substantial gaps between models: even Nano Banana Pro reaches only 42.8 and the best open-source model scores 20.1, with deficits in handling complex context, entangled references, and reasoning-heavy tasks; (c) strong generative models effectively draw on the capabilities of their underlying language models. These findings highlight the need for multimodal generative models that are capable of dynamic planning and reasoning while also being deeply unified to faithfully simulate real-world states.

---

### Orbia: Do Generated Videos Form a Consistent 3D World?

Authors: Dongyue Lu, Guian Fang, Hur Lim, Yonggang Wu, Yanrong Wang, Hongshan Chen, Shengju Qian, Yue Liao, Wei Chow, Lingdong Kong, Tianxin Huang, Yu Yang et al.
Published: 2026-10-14
Categories: cs.CV
arXiv: https://arxiv.org/abs/2610.20693v1

Video generators serve as interactive worlds for exploration and simulation, yet realistic and visually appealing frames do not necessarily depict a consistent 3D world. Existing benchmarks assess selected geometric properties or use vision-language models to judge memory and consistency, leaving systematic 3D evaluation underexplored. We introduce Orbia, the first benchmark dedicated to a systematic and comprehensive evaluation of 3D consistency in interactive world models. Through controlled exploration and revisit, we organize the evaluation into three dimensions: input preservation, generated-content persistence, and 3D self-consistency. For input preservation, registered RGB-D references and annotated anchors support appearance, geometry, and spatial stability measurements. For generated-content persistence, we compare revisits to regions absent from the input image with their first generated observations to test whether models preserve what they create. For 3D self-consistency, cross-view depth and feature correspondence, regional scale agreement, and held-out Gaussian rendering test whether generated views describe the same 3D scene. Minute-long exploration and revisits further test long-term consistency. Our construction and annotation pipeline brings together public datasets, self-captured scenes, and simulation, yielding 720 instances across indoor and outdoor environments and six exploration-and-revisit protocols. Evaluation spans 27 text-, action-, and camera-conditioned models through native-interface adapters. Our Orbia benchmark provides a systematic basis for diagnosing 3D inconsistency and developing more persistent video world models.

---

### Open Benchmark of AI Impact on Humans (ImpactBench): an expert-guided, open platform for the holistic evaluation of AI impact on humans

Authors: Chayapatr Archiwaranguprok, Krystal Kirkland, Rachel Poonsiriwong, Stanley Huang, Constanze Albrecht, Jackson Trager, Emillie de Keulenaar, Sheer Karny, Peter Baxter, Nikos Bellos, Andre Luiz Freire de Carvalho Kato, Yuning Liu et al.
Published: 2026-10-14
Categories: cs.HC, cs.AI
arXiv: https://arxiv.org/abs/2610.20690v1

The present work is concerned with how to measure, at scale, the multi-dimensional impact of generative AI models on the individuals who use them. In service of this goal, we introduce the Open Benchmark of AI Impact on Humans, or ImpactBench. Our iterable and expert-guided platform evaluates whether model behaviors support or undermine an individual's wellbeing across psychological, physical, and social domains. ImpactBench translates constructs from domain experts and existing benchmarks into metrics that are grounded in actual model behavior. Success on these metrics occurs either when a model promotes expert-defined desirable behaviors or avoids identified harms. Models are evaluated against these metrics in simulated multi-turn conversations using matched teen and adult scenarios. The conversations are then scored by an LLM judge and verified through reliability audits and human expert review. The current suite delivers 809 metrics drawn from 26 benchmarks. We evaluated ten frontier LLMs against these metrics across 48,540 conversations. Our results indicate that models promoted desirable behaviors in 68.7% of evaluations, but avoided harms in only 53.1%. Majority expert ratings agreed with the automated judge in 83.3% of reviewed cases (Cohen's Kappa = 0.667). Results are released through a public platform that keep aggregate scores traceable to the underlying metrics and transcripts. ImpactBench is intended as evolving infrastructure for independent, transparent evaluation of generative AI's multi-dimensional impact on humans.

---

### On-Policy Residual Learning for Diffusion Language Models

Authors: Yifan Yu, Yuqing Jian, Jisen Li, Yuezhou Hu, Tianyi Zhang, Xiaoxia Wu, Rui-Jie Zhu, Junxiong Wang, Zhongzhu Zhou, Fan Lai, James Zou, Chenfeng Xu
Published: 2026-10-14
Categories: cs.LG
arXiv: https://arxiv.org/abs/2610.20687v1

Residual context is key to unlocking the speed-quality Pareto frontier of diffusion language models (DLMs) by carrying forward predictive distributions across denoising steps. Yet turning this seemingly simple mechanism into an effective learning paradigm is far from simple: there is no clear recipe for how residual context should be learned. We introduce ORL, a framework for on-policy residual learning in DLMs. Algorithmically, we identify that residual learning is inherently on-policy. ORL learns from model-induced residual trajectories on the fly and introduces an entropy-contraction objective that sharpens predictions as the model revisits its own predictions. System-wise, exact residual supervision requires capturing, retaining, and reprocessing many trajectory-dependent states. ORL co-designs the algorithm with asynchronous residual capture and transport, specialized block-diffusion attention, and packed compute-balanced training, making it practical at scale. On a 4B DLM converted from Qwen3-4B, ORL improves tokens per forward (TPF) by up to 6.2x and, with our dlmrt engine, end-to-end throughput by approximately 6.9x at comparable or better quality. In single-request serving on one H100, this mode achieves a stronger speed-quality Pareto frontier than autoregressive models with EAGLE3 and DFlash speculative decoding under both strict and relaxed verification, running 1.74x faster than DFlash at matched accuracy. ORL also transfers to existing DLMs: on LLaDA-2.1-mini, 36 H100-hours and 4.2M self-generated tokens, about 10x less data than prior LLaDA recipes, yield up to 1.5x higher TPF and up to 3x higher throughput at comparable quality, and on DiffusionGemma, ORL raises TPF by 1.1-1.2x. With dlmrt and NVFP4 quantization, ORL models decode a single request at up to 5,046 tokens/s on one B200, 7,708 on eight B200s, and 2,788 on a consumer RTX 5090, with comparable quality.

---

### RoLaQ: Joint Rotation and Range Calibration for 2-Bit KV Caches in Multi-Head Latent Attention

Authors: Shiyu Ji, Peng Ling, Chen Lin, Shupeng Li, Yixuan Wang, Qingfu Zhu, Wanxiang Che
Published: 2026-10-14
Categories: cs.LG, stat.ML
arXiv: https://arxiv.org/abs/2610.20686v1

Multi-head latent attention (MLA) compresses the content of a key-value cache into a shared latent per token, making low-bit storage attractive for long-context inference. At two bits, errors in this representation perturb both attention matching and value readout, with effects that depend on error direction and reconstruction scale. We introduce RoLaQ (Rotation-calibrated Latent Quantization), an offline method for joint rotation and range calibration. RoLaQ refines a shared rotation within subspaces defined by the frozen key and value projections, and jointly learns quantization ranges whose thresholds and reconstruction gains are coupled to control the shared signal's response. Prediction matching and attention-output supervision guide calibration with model weights frozen. Across three MLA models and six benchmarks, RoLaQ achieves the highest average accuracy among the compared 2-bit methods both without and with token protection. On DeepSeek-V2-Lite, an SGLang implementation provides 3.96 times the BF16 KV capacity under the same memory budget and 2.51 times its decode throughput with 256K input tokens.

---

### L2G-SR: Learning to Guide LLM-based Symbolic Regression

Authors: Letian Huang, Xiangyu Wang, Haowei Lin, Di He
Published: 2026-10-14
Categories: cs.AI
arXiv: https://arxiv.org/abs/2610.20685v1

Symbolic regression searches for an equation that explains observed data. Numerical error is commonly used to guide the search, but it alone can favor an accurate surrogate over a candidate with the right structure. We propose L2G-SR, a framework for LLM-based symbolic regression that guides search with a learned structural verifier. We construct structural supervision signal using tree edit distance (TED), accounting for commutative operators and alternative expression forms. A verifier first aligns numerical observations with the formulas, then learns to score a candidate's structural similarity to the target formula. At search time, the target formula is hidden and the verifier complements numerical error in parent selection. An analyst then uses verifier feedback and diagnostic tools to propose revisions, which a generator converts into refined expressions. On LLM-SRBench, L2G-SR achieves state-of-the-art numerical fit, equation level accuracy and TED across the evaluated LLM backbones and domains. L2G-SR demonstrates the value of learned structural guidance for LLM-based symbolic regression.

---

### Learning Topology-Aware Representations for Mechanics-Relevant TPMS Metamaterials

Authors: Songlin Duan, Shuo Yang, Yujia Zheng, Tobin Filleter
Published: 2026-10-14
Categories: cs.LG, cs.CE, cs.CV, stat.ML
arXiv: https://arxiv.org/abs/2610.20684v1

Learning informative representations of mechanical Triply Periodic Minimal Surface (TPMS) metamaterials is important for mechanics analysis and structural design, yet existing methods have rarely explicitly supervised learned representations with information about global structural connectivity that is relevant to mechanical response. We propose a topology-aware representation learning framework that integrates persistent homology into a DMAE backbone to explicitly encode global structural connectivity. A persistence-guided masking strategy prioritizes patches associated with persistent topological features, while an auxiliary persistent-homology decoder encourages the latent representation to retain information about topological feature counts across connected components, tunnels, and voids. We evaluate the learned representations on two complementary TPMS datasets using family classification, mechanical-property classification, unsupervised clustering, and regression. Across the two datasets, our method demonstrates strong performance in both structural family and mechanical-property classification, improving family classification by up to 3.9% while achieving the best linear-probe performance. More notably, unsupervised property-clustering performance improves by up to 26.7% and 51.5% across the two datasets, while strong regression performance is retained with $R^2=0.993$. These results establish persistent topology as an effective inductive bias for representation learning of complex three-dimensional mechanical metamaterials.

---

### Causal Syntax Circuits in CodeLMs: Budget and Probe Decide What Looks Localised

Authors: Harsh Kumar, Tanmay Shailesh Joshi, Rahul Maity, Anusa Saha
Published: 2026-10-14
Categories: cs.LG, cs.SE
arXiv: https://arxiv.org/abs/2610.20683v1

Code models write Python when the prompt does not say which language to use. We ask a mechanistic question about that habit: is it produced by a small, findable group of components inside the model? We run one pipeline---mean ablation, activation patching, greedy circuit extraction, a random-control significance test, and activation steering---over 55 model-language settings spanning ten open code models (five base and five instruction-tuned), three target languages (Rust, C#, PHP), two prompt suites, and two component budgets. We report five findings. First, the preference is large, grows with model size, and survives instruction tuning: it is larger on every instruction-tuned checkpoint than on its base version, by 12% on average. Second, its measured size is mostly a property of the probe; replacing our aligned prompts with XLCoST parallel functions takes the same measurement on the same models from about 1.6 to about 0.01. Third, whether the preference looks localised depends on how many components the search may keep: with a budget of 12, 1 of 15 settings yields a circuit that beats size-matched random controls after multiple-comparison correction, and with a budget of 40, 6 of 15 do. Fourth, interventions that a smaller-budget study reports as uniformly failing do sometimes succeed: patching makes a Python prompt prefer target syntax in 2 of 45 settings, and steering changes the most likely next token in 9 of 45. Fifth, every circuit whose effect is significant but points away from the target language comes from a setting where circuit extraction failed, which makes such results a diagnostic rather than a phenomenon. The practical lesson is that a claim about whether a behaviour is localised can turn on analysis choices that papers rarely report. We release the full grid so those choices can be varied.

---

### TurboMamba: Faster Inference, Faster Training, and Better Prediction with Lexical Memory

Authors: Hanxin Zhang
Published: 2026-10-14
Categories: cs.LG, cs.CL
arXiv: https://arxiv.org/abs/2610.20682v1

We introduce TurboMamba, a language-model architecture for faster inference, faster training, and better token prediction. A global Mamba-3 model runs once per completed token block. A small Transformer uses this global context and earlier tokens within the block to predict the next token. Three lightweight additions improve local predictions: direct access to the previous block's last token, a small correction to next-token scores, and a learned mixture with recent-token frequencies. We compare twelve small-scale implementations on public FineWeb-Edu, with comparable parameter counts and shared training-budget caps. The baselines include Transformer, Mamba-3, Jamba, and Zamba. The Jamba and Zamba versions use Mamba-3 in place of their original Mamba modules. TurboMamba achieves the lowest mean test prediction loss (negative log-likelihood, NLL) among all twelve models. Its mean NLL is 9.4% lower than that of Mamba-3, 7.3% lower than that of the Jamba version, and 18.3% lower than that of the Zamba version. End-to-end training throughput is 2.60x that of Mamba-3, 2.81x that of the Jamba version, and 4.82x that of the Zamba version. We measure end-to-end batch latency for a batch of 32 sequences, each with a 256-token prompt and 128 generated tokens. On this workload, TurboMamba achieves a 3.12x speedup over Mamba-3, a 3.53x speedup over the Jamba version, and a 6.21x speedup over the Zamba version.

---

### IneqAgent: Agentic Inequality Proving in Lean with Symbolic Computation

Authors: Gaolei He, Jinhui Li, Ruobing Zuo, Xia Zeng, Zhengfeng Yang
Published: 2026-10-14
Categories: cs.AI
arXiv: https://arxiv.org/abs/2610.20681v1

Formal inequality proving requires mathematical reasoning and exact computation, often connected through intermediate lemmas and goal reformulations. We introduce IneqAgent, an agentic framework combining LLM reasoning with symbolic computation for inequality proving in Lean4. Lemma-level Monte Carlo Tree Search coordinates lemma construction, goal reformulation, and symbolic backend selection. Complementary sum-of-squares, Bernstein, and successive difference-substitution backends produce exact certificates for local Lean proof generation. The agent integrates these proofs with LLM-generated steps and uses Lean feedback to construct complete proofs checked against the original theorem statements. On four benchmarks, IneqAgent with DeepSeek-V4-Flash solves 175 of 186 problems (94.09%), achieving the highest overall success rate among the evaluated methods. Across four backbones, it achieves higher success rates at low LLM inference cost thresholds than Direct Agent using the same backbone, with lower mean inference costs over each method's successful runs. We also obtain Lean-verified strengthenings of Dittert's conjecture in dimensions four and five, and certify the Holens--Djokovic $(4,4)$ inequality, implying Merris's conjecture in dimension four.

---

### Diminishing Returns of Data in Sparse Autoencoders: Functional Behavior does not Guarantee Feature Reproducibility

Authors: Sambit Ghosh, Madelyn Mathai, R. Venkatesh Babu, Chirag Agarwal
Published: 2026-10-14
Categories: cs.LG, stat.ML
arXiv: https://arxiv.org/abs/2610.20680v1

Sparse autoencoders (SAEs) are typically trained on hundreds of millions to billions of activation tokens, yet little work has tested whether this scale is necessary. We study how much activation data SAEs need by training four SAE architectures on three language models with budgets ranging from 5M tokens to a 500M-token reference. We measure two notions of convergence: functional behavior on standard evaluations and one-to-one reproducibility of learned features. Functional data requirements are strongly metric-dependent: reconstruction often approaches the full-data reference at substantially smaller budgets, whereas feature reproducibility continues to improve with data. For Gemma-2-2B BatchTopK, the fraction of strictly matched features rises from 2\% at 5M tokens to 29\% at 50M. Data selection also matters: random sampling better preserves the original activation distribution and generally favors reconstruction, whereas diversity-based sampling frequently improves AutoInterp. A theoretical analysis separates statistical consistency of subsampled training from the stronger identifiability and optimization conditions required for one-to-one feature recovery. Our results show that SAE training can often be made far more data-efficient when budgets are matched to the goal, and feature-level claims are backed by explicit reproducibility checks, paving the way toward cheaper and more reliable SAE research.

---

### Counterfactual Neighborhood Risk Fields: Evidence-Grounded Selective Safety for Vision-Language Models

Authors: Tianshi Wang, Jingsong Wang, Fengling Li, Jingjing Li, Xin Li, Lei Zhu
Published: 2026-10-14
Categories: cs.AI, cs.CR, cs.MM
arXiv: https://arxiv.org/abs/2610.20679v1

Representation-based safety defenses for vision-language models (VLMs) often learn to separate harmful from benign inputs, but separation alone does not establish what a risk score means or when it should be trusted. We introduce Counterfactual Neighborhood Risk Fields (CNRF), an evidence-grounded framework for selective VLM safety. CNRF constructs context-matched benign-harmful counterfactual pairs, uses their representation differences as capability-related risk directions, and estimates neighborhood support separately from risk. This separation prevents unfamiliarity from being conflated with harmfulness and enables selective intervention rather than blanket rejection. A victim-adaptive interface accommodates different safety observations across VLM architectures and routes selected requests for same-victim response review, which either releases acceptable candidates or produces constrained alternatives with reduced harmful capability. Across three VLMs and five multimodal jailbreak families, CNRF achieves 86.06-89.30% macro defense success rate while retaining substantial benign-task utility. Controlled experiments further show that query-matched references and preserved context-direction correspondence materially improve risk ranking in the low-false-positive regime. CNRF thus reframes representation-based VLM safety from learning harmfulness boundaries to constructing applicable evidence for selective risk decisions.

---

### Spectral Indistinguishability and Learning-Curve Identifiability in Quadratic Networks

Authors: Dekun Yang
Published: 2026-10-14
Categories: cs.LG, stat.ML
arXiv: https://arxiv.org/abs/2610.20678v1

We study what spectral summaries and scalar learning curves reveal about centered quadratic networks. We construct analytic positive input densities with exactly Gaussian one- and two-coordinate marginals. Their complete fourth-moment operators are intertwined by an orthogonal map fixing the entire diagonal teacher algebra, so the corresponding anchored spectral measures agree. For selected common teachers with a nonzero cycle invariant, their factorized loss curves differ because the map does not preserve matrix multiplication. Cycle length $r$, which can be arbitrarily large, gives $O(\varepsilon^r)$ contrast uniformly in training time for each fixed construction. For a supplied pair, noisy scalar loss queries require order $\varepsilon^{-2r}$ observations, while bounded initialization interventions with scalar gradient readouts require order $\varepsilon^{-2}$. A shared-input score, equal to a paired-gradient difference, attains the optimal input-sample exponent without knowing the teacher. An explicit family of diverging auxiliary teacher scales instead admits a nonzero finite limiting contrast at coupling $1/\log L$. Conversely, for Gaussian inputs and a rationally independent teacher spectrum, one exact scalar loss curve identifies the initial predictor up to sign conjugation. Within this rationally independent class, with nonzero diagonal inverse-state error in every coordinate and calibrated time and loss scales, it also identifies the unknown teacher spectrum and active dimension, modulo simultaneous rotations. For fixed dimension and distinct unordered teacher sums, dimension-minimal exact observations give generic joint local recovery. Finite exact global bounds hold for a known teacher and, on the active class, under a blind dimension cap. Exact certificates and full-weight replay verify the construction; numerical inverse checks reveal severe ill-conditioning.

---

### When Correctness Feedback Erases Evidence: Phase-Separated Completeness Auditing in Verus

Authors: Adarsh Agrawal
Published: 2026-10-14
Categories: cs.AI
arXiv: https://arxiv.org/abs/2610.20676v1

Correctness feedback suppresses the very evidence that specification-completeness auditing is designed to uncover: programs that satisfy a written contract but violate intended behavior. We demonstrate this objective reversal on 12 Verus/Rust tasks across four model families under matched call budgets. Independent no-feedback discovery more than doubles certified-witness yield relative to IntentShield's corrective-feedback loop (59.0% versus 27.1%). A format-matched iterative control places the larger share of this gap in feedback content, not independent restarts. The advantage holds in every model family and survives parser repair. A separate phase-separated execution matches the aggregate yield without reusing responses; most certified witnesses also fail held-out cases. Preserving these witnesses delivers stronger specification repair: Phase-Revised contracts achieve 81.3% implementation success, versus 62.5% for IS-Revised, with Direct-Revised close behind. Both no-feedback witness sources substantially outperform weakened contracts. The effect does not reproduce in the available Dafny and Lean evaluations, establishing an empirical boundary around the studied Verus/Rust workflow. Completeness auditing and correctness repair have opposing immediate objectives. Preserve and certify the admitted failure before repairing it.

---

### AlphaDiana: A Measurement System for Agents in Verifiable Reasoning Tasks

Authors: Zhanke Zhou, Zongze Li, Weikai Huang, Xuan Li, Chentao Cao, Xiao Feng, Xiangyu Lu, Jinbo Hu, Yi Xie, Nico Pelleriti, Shiyang Liu, Menghan Lu et al.
Published: 2026-10-14
Categories: cs.AI, cs.LG
arXiv: https://arxiv.org/abs/2610.20675v1

Agents pair reasoning models with harnesses that organize model calls, execute tools, load skill instructions, and retain history in memory. Understanding agent performance requires linking execution conditions to outcomes, behavior, and resource use. We introduce AlphaDiana, a measurement system with modular interfaces for datasets, agents, environments, and scorers. It preserves harness-native control loops while controlling file reuse and memory retention separately. AlphaDiana supports macro analyses of complete agents and micro interventions on agent components such as tools, skills, and memory under matched conditions. We study three open-source models and three open-source harnesses on five answer-based benchmarks, with two models also tested on two interactive benchmarks. We find that harnesses do not consistently improve the accuracy of reasoning: gains on questions unsolved by the model alone can be outweighed by losses on questions it answers correctly without tools. This inconsistency extends to component interventions: neither skill guidance nor retained history reliably improves accuracy. Behavior and output-token use must be assessed with correctness: low token entropy can accompany wrong answers, and less output per execution can mean more output tokens per correct answer when accuracy falls. AlphaDiana links answer correctness and execution status to traces and resource records for comparing agents and diagnosing failures under shared task and resource policies.

---

### On the Stability and Convergence of Flow Map Training

Authors: Ishan Nath, Tom Stanic, Avishek Joey Bose, Gauthier Gidel, Quentin Bertrand
Published: 2026-10-14
Categories: stat.ML, cs.LG
arXiv: https://arxiv.org/abs/2610.20674v1

Flow maps are a recent method of probabilistic generation that amortize diffusion trajectories into one-step samplers, unlocking efficient generation. While progress has been made in improving their empirical performance, there is still a lack of understanding behind the mechanisms underlying their success. Specifically, many flow map learning methods employ stop gradient operators as a crucial part of their training target. Adaptive weighting also drastically improves the generation capabilities of flow maps. Although many interpretations have been proposed to support the inclusion of these heuristics, none have analyzed the impact of these assumptions on the training dynamics of flow maps themselves. In this work, we address this issue by providing grounded theory explaining why such tricks are required. Analyzing the functional gradient flow induced by the MeanFlow and Lagrangian losses, we show their convergence to the true flow map, settling the validity of these objectives. For finite updates, we highlight a severe roadblock to the usage of these losses in practice: the possible unboundedness of the update. Adaptive normalization not only provides bounded updates, but, combined with a specific weighting, additionally provides convergence guarantees towards a neighborhood of the solution. On synthetic data, we show that our functional gradient flow analysis exactly describes the true training dynamics for flow maps. Experiments on standard image datasets demonstrate that a proper reweighting scheme is essential for strong empirical performance, which is only achieved for a specific range of reweighting strengths.

---

### Value-Informed Schedule Trimming for Accelerated Sequential Monte Carlo Guidance in Discrete Diffusion

Authors: Jaehyeon Kim, Ziseok Lee, Siyoung Choi, Kyungsu Kim
Published: 2026-10-14
Categories: cs.AI, cs.LG
arXiv: https://arxiv.org/abs/2610.20669v1

Reward-guided discrete diffusion enables test-time alignment with downstream objectives, but applying Sequential Monte Carlo (SMC) guidance at every denoising step is computationally expensive. Active guidance steps require additional reward evaluations and, for gradient-based proposals, value-gradient computations. Sparse guidance reduces this cost by guiding only a subset of timesteps, but commonly used schedules are fixed heuristics whose effectiveness can vary substantially across tasks. We propose VISTA, a training-free method that automatically selects guided timesteps under a fixed budget. VISTA formulates sparse guidance as approximation of the ideal full-guidance probability path and derives an estimable upper bound on \(\ell_1\) path deviation. Using value statistics collected from a small number of full-guidance warmup generations, VISTA optimizes the resulting empirical surrogate exactly via dynamic programming. The warmup generations can be retained as model outputs, and after schedule selection VISTA has the same per-generation guidance complexity as other sparse schedules with the same budget. Across text, biological sequence, and image generation tasks, VISTA discovers task-dependent schedules and outperforms hand-designed sparse schedules in most tested settings, while approaching or exceeding full-guidance performance on several task metrics. In toxic-text generation, for example, VISTA attains a 0.99 unique-toxic ratio compared with 1.00 for full guidance while reducing post-selection per-generation latency from 204.23 s to 38.16 s, a 5.35$\times$ speedup.

---

### CascadeLedger: Counterfactual Auditing of Information and Resource Integrity in LLM Cascades

Authors: Miaobo Hu, Shuhao Hu, Xiaobo Guo, Xin Wang, Bokun Wang, Daren Zha, Jun Xiao
Published: 2026-10-14
Categories: cs.CR, cs.CL, cs.LG
arXiv: https://arxiv.org/abs/2610.20668v1

We introduce CascadeLedger, a counterfactual execution-audit framework for budgeted language-model cascades. The protocol enforces call-before-read semantics, binds successful reads to service and cache identities, records availability and component-wise budget transitions, and freezes the event trace before any target or task scorer is joined. The central object is aggregate-equivalent trace validity: two executions can share a final answer, declared resource projection, and service histogram while differing in whether their routing decisions used only legitimately observable information. We complement single-trace invariants with counterfactual prefix non-interference (CPNI), which couples a visible prefix and policy randomness while changing only a declared hidden family. The counterfactual influence frontier (CIF) localizes the earliest violating prefix and a minimal hidden witness; projection-preserving adversarial mutation (PPAM) searches for invalid traces that remain acceptable to a non-counterfactual checker. Across frozen-cache, independently generated, and serving evaluations, the two-layer audit reports a clean false-positive rate of 0.0000, detects 0.9983 of serialized faults and 0.9982 of unlogged hidden-read faults, and changes 0.3403 of corrupted quality-resource conclusions after deterministic recomputation. The online recorder and offline CPNI audit have separate overhead measurements. Learned routing and semantic verification provide downstream compatibility studies under the same execution contract.

---

### Evaluating Model Capabilities After Benchmark Saturation

Authors: Morio Hamada, Jehyeok Yeon, Matteo Migliarini, Justin Olive, Magda Dubois
Published: 2026-10-14
Categories: cs.AI, cs.LG
arXiv: https://arxiv.org/abs/2610.20667v1

Rapid advances in AI capabilities are causing benchmarks to saturate. Once several models approach perfect scores, a benchmark loses its power to differentiate between them, and with it its utility for informing AI safety cases and frontier policies. Yet benchmark items produce far more information than the scores we typically extract from them: a full transcript of how each model solved the task. We investigate whether these transcripts can be used to assess model capabilities even after benchmark scores have saturated. We simulate saturation retrospectively using historical benchmark and model releases, and test whether protocols that access only correctly solved transcripts from saturated benchmarks can predict capabilities measured on held-out, unsaturated items. We evaluate representative contemporary approaches for extracting information from model transcripts: pairwise judging, single-transcript absolute scoring, and predictions from deterministic transcript features or embeddings. We find that saturated transcripts retain capability information beyond benchmark accuracy. Amongst our evaluated protocols, pairwise judging performs best, reducing prediction error by 5.3 times compared to a baseline using only saturated benchmark scores. This capability-relevant information is distributed across transcript sections, but concentrated in specific channels of information such as reasoning and tool-calls. However, the predictive power of protocols does not always generalise across distribution shifts -- for example by underestimating the capabilities of the first reasoning models introduced in early 2025. Saturated benchmarks thus retain capability-predicting signal in their task transcripts, but extending the lifetime of existing benchmarks requires methods designed to extract these signals robustly.

---

### The World Won't Wait: Evaluating and Teaching Agents Under a Live Clock

Authors: Ningxin Pan, Hanyu Li, Yeqing Shen, Ning Ding, Yehui Tang, Wentao Zhang
Published: 2026-10-14
Categories: cs.AI
arXiv: https://arxiv.org/abs/2610.20665v1

Large language model (LLM) agents can solve complex interactive tasks, yet reasoning and tool calls take time. In continuously changing environments, an otherwise correct action may arrive too late. We introduce RealtimeBench, with 180 tasks across six game types, to evaluate agents while games continue running during inference. Paired evaluations show that GPT-6.1-sol outperforms the faster Jev when games are paused, but falls behind when they keep running. We propose Two-Timescale Offline Recursive Self-Improvement (TORSI): a slow teacher studies failures and tests corrections in paused games between attempts, then writes reusable instructions for a fast student. The student acts without teacher calls during gameplay, and validation scores determine which updates are adopted. Two rounds raise Jev's average test score from 13.51 to 24.19 on a 0--100 progress scale, improving all six game types. The instructions also improve average scores for three other models without retuning. These findings point toward real-time intelligence: enabling powerful LLM agents to remain effective in a continuously changing world by organizing deliberation, learning, and action across timescales.

---

### Graph Additive Models disentangling feature, topology, and interactions

Authors: Clara Alvarez-Rodriguez, Steve Azzolin, Vincenzo Marco De Luca, Filippo Maria Bianchi, Andrea Passerini, Antonio Longa
Published: 2026-10-14
Categories: stat.ML, cs.LG
arXiv: https://arxiv.org/abs/2610.20662v1

Graph Neural Additive Networks (GNANs) have emerged as effective, interpretable-by-design architectures for graph learning. They achieve transparency by combining per-feature shape functions with simple distance-based reweighting. However, GNANs have three limitations. First, node-wise relevance scores aggregate information from all graph nodes, allowing irrelevant nodes to inherit spurious relevance from relevant ones. Second, topology is represented only through shortest-path distances, so structurally distinct graphs may share identical distance profiles. Third, feature and distance information are combined multiplicatively, entangling feature, distance, and interaction effects in a single rescalable term. We introduce AMIGA, which represents topology through a vocabulary of mined motifs and decomposes each node contribution into three additive channels: feature, structure, and feature-structure interaction. The channels are centered around a graph-balanced training baseline, making the decomposition well-defined and unique under the chosen reference distribution. Across ten graph classification and regression benchmarks, AMIGA outperforms GNAN on most datasets and narrows the gap with standard black-box GNNs. Because each channel is centered and assigned to a single node, AMIGA produces node-local explanations that reconstruct the model output exactly and support interpretation at multiple levels of granularity.

---

### Latent Annealed Posterior Sampling with a Variance-Preserving Bridge for Sparsely Observed Random Fields

Authors: Jiankun Lei, Tian Qi
Published: 2026-10-14
Categories: stat.ML, cs.LG
arXiv: https://arxiv.org/abs/2610.20659v1

Posterior inference for random fields from sparse observations presents two major challenges: accurately characterizing complex spatial uncertainty and preventing conditional sampling costs from growing substantially as the spatial grid is refined. To address these challenges, we propose latent annealed posterior sampling (LAPS). LAPS first maps the random field into a latent space with reduced dimensionality, while a coordinate decoder composed of Branch and Trunk networks enables field evaluation at arbitrary spatial locations. Flow matching is then used to learn the latent prior distribution. Based on this learned prior, posterior sampling is performed entirely in the latent space, with the observation likelihood incorporated progressively through an annealing procedure. To control local covariance inflation, we introduce a variance-preserving (VP) bridge that allocates a covariance budget between the transported latent state and a Gaussian perturbation. Numerical experiments demonstrate that LAPS accurately approximates posterior distributions, preserves distinct posterior modes, and reduces the dependence of conditional sampling costs on spatial resolution. Across two Gaussian benchmarks, LAPS attains small errors against exact posterior references, with sliced Wasserstein distances only 2 to 3 percent higher than the reference values obtained from exact posterior samples. In a bimodal posterior problem, LAPS reduces the sliced Wasserstein distance from 0.2693 to 0.1499 and consistently recovers both modes across all test scenarios. For a two dimensional Navier-Stokes random field, LAPS substantially improves marginal posterior calibration, while its evaluation time remains nearly constant as the output grid is refined from 64^2 to 512^2. These results demonstrate that LAPS achieves a favorable tradeoff among posterior accuracy, uncertainty calibration, and computational efficiency.

---

### A polylogarithmic bound for list replicability

Authors: Ari Blondal, Hamed Hatami, Pooya Hatami, Chavdar Lalov, Sivan Tretiak
Published: 2026-10-14
Categories: stat.ML, cs.LG, math.CO
arXiv: https://arxiv.org/abs/2610.20657v1

Zhang (Zha26) recently posted a manuscript announcing that the list-replicability number of every class of $M\ge 2$ binary concepts on $N\ge 2$ domain points is $O(\log^2 M\log N)$, which would be the first polylogarithmic upper bound on this quantity. In this note, we present a short, self-contained proof of this bound, following the high-level approach of the manuscript but with considerably simpler arguments.

---

### Can Learning to Answer Spatial Questions Improve Novel-View Generation?

Authors: Jahyeok Koo, Hyeonseo Yu, Seungho Jang, Jaewoo Jung, Jiwon Kang, Seokju Cho, Seungryong Kim
Published: 2026-10-14
Categories: cs.CV
arXiv: https://arxiv.org/abs/2610.20610v1

Generating novel views from limited observations requires preserving both the appearance and spatial structure of a scene. Although recent approaches exploit fine-grained geometric cues, they do not explicitly model high-level semantic and spatial properties such as object identity, attributes, and inter-object relationships. In this work, we investigate whether 3D language understanding can provide complementary supervision for novel-view generation. We introduce Ask4NV, built on a pretrained unified multimodal model with a Mixture-of-Transformers (MoT) architecture. Our model incorporates camera geometry into both VAE and ViT representations and leverages understanding features during generation through joint self-attention with a modality-aware attention masking strategy. Experiments demonstrate strong image quality and geometric consistency in novel-view generation, alongside competitive 3D language understanding. Further analysis shows that joint training improves cross-view correspondence in VAE representations and similarity in a 3D-aware feature space.

---

### Representation Conditioning for Generation

Authors: Saeed Razavi, Xingjian Leng, Liang Zheng
Published: 2026-10-14
Categories: cs.CV
arXiv: https://arxiv.org/abs/2610.20608v1

Representation-alignment methods effectively accelerate the diffusion transformer training. REPA, for example, uses a pretrained vision encoder to supervise an intermediate layer through an MLP projection head which is discarded after training. It aims to import semantics from the vision encoder to the diffusion backbone. While this objective is achieved at early layers by back-propagating the alignment loss, there is no guarantee that later layers, supervised by the velocity prediction loss alone, can well access such semantics. We propose a simple solution, representation conditioning (RECON), to improve semantics in later layers. During training, we feed the pretrained vision encoder tokens into later layers; during inference, because we no longer have the clean image, we reuse the MLP head (as a proxy to the vision encoder) and project its output tokens to the later layers. Our method adds no parameters and leads to much faster convergence: RECON is $63\times$ faster than SiT and $25\times$ faster than REPA. Code is available at https://github.com/saeidrazavi/RECON

---