The paper introduces Demographic Pluralism, an inference-time framework that estimates population-level opinion distributions without requiring training data or task-specific fine-tuning. It generates multiple perspectives within demographically grounded groups and, across four model backbones on GlobalOpinionQA and VITAL, reduces Jensen-Shannon distance by 8.4%–26.4% compared to Modular Pluralism. The study finds that equal weighting of group perspectives yields the best overall performance, while weighted aggregation performs worse due to increased group-level error.
By Meng-Chen Wu, Qipin Chen, Ansh Jain, Tess Wood, Zhe Du, Si-Chi Chin
BELIEFRAG is a closed‑loop controller that maintains an explicit evidence state—tracking sufficiency, reliability, conflict, uncertainty, evidence gaps, and acquisition cost—to guide adaptive retrieval‑augmented generation (RAG). By updating this state, it selects among retrieval, query rewriting, verification, answering, stopping, and abstention, achieving higher token‑F1 scores and lower token usage on six QA benchmarks compared to fixed iterative retrieval. The approach shows that corrective re‑retrieval and calibrated answerability are key to its performance gains.
By Hongji Pu
MADBench is a benchmark that evaluates the security of Multi-Agent Debate (MAD) systems, which allow large language models to exchange and critique answers. The study categorizes attacks into a layered taxonomy aligned with the MAD workflow and tests six attack families across 356 source tasks and 3,958 test cases. Results indicate that while MAD can reduce answer accuracy attacks compared to single-agent baselines, it may amplify unauthorized reads or writes, and even with collusion among agents, the final answer changes from correct to wrong only 28.30% of the time.
By Yuwan Liu, Jiaming Zhang, Yue Huang, Sisi Duan
The paper proposes measuring a language model’s understanding via no‑arbitrage, defining it as the inability of a bounded trader to profit from Dutch books against the model’s probabilities on logically related claims. It shows that full logical coherence is computationally infeasible, that standard next‑token training yields incoherent predictions across formats, and that uncertainty grows predictably along reasoning chains, creating arbitrage opportunities. The authors introduce Arbitr, a training framework that penalizes logical inconsistencies while maintaining accuracy, reducing exploitability by orders of magnitude and revealing a scaling illusion where large models appear coherent yet exhibit extreme unjustified confidence.
By Daniel Dragonevskiy
The paper introduces Event-Driven Refresh + Recurrence Memory (EDRRM) to improve Referring Video Object Segmentation (RVOS). EDRRM selectively re-invokes the Sa2VA model at stable change points, using an EMA‑smoothed event score from tracking cues and a recurrence memory that retrieves anchor frames via CLIP similarity. Experiments on Ref‑DAVIS17, MeViS, and ReVOS show that EDRRM maintains or surpasses J&F scores while reducing refresh calls and false‑positive failures, with modest overhead compared to Sa2VA inference.
By Abu Hanif Muhammad Syarubany, Jaehyun Jang, Siwoo Lim, Seungyeon Ryu, Chang D. Yoo
OP-CAD introduces a curriculum-based, on-policy clean-audio distillation framework that enhances audio-visual reasoning under environmental noise and competing speech. The method trains a student model from mild to severe noise, using a frozen teacher that provides token-level supervision based on clean audio and verified answers, while selectively weighting positions sensitive to acoustic interference. Experiments show OP‑CAD outperforms existing methods across all noise conditions, preserving clean‑correct answers without sacrificing overall accuracy.
By Xingming Shui, Dapeng Chen, Bowei Liu, Jingqi Tian, Minfu Li, Kun Yi, Jiapeng Hong, Yansong Tang
The paper introduces RAE-CoD, a diffusion-based compression method that operates in a representation autoencoder space to preserve recognizable content even at extremely low bitrates. It addresses the problem of semantic collapse observed in existing codecs when the bitrate approaches zero, showing that reconstruction losses conflict with semantic objectives and that VAE diffusion models lose efficiency in preserving semantics. Experiments on MSCOCO-30K demonstrate that RAE-CoD outperforms competitors, reducing VFM feature MSE and Fréchet Distance ratios by at least 25.7% and 69.1% at 0.001–0.008 bpp while maintaining stable recognizability and quality.
By Tianyu Zhang, Zhaoyang Jia, Houqiang Li, Dong Liu
Soft Spatial Reasoning introduces a post‑training framework for Large Vision‑Language Models that replaces hard, token‑by‑token chain‑of‑thought reasoning with a soft, continuous state formed by mixing token embeddings at each intermediate step. The method employs AdaptSoft, a controller that adjusts the degree of softness based on hidden states and predictive uncertainty, guided by a gradient‑alignment learning objective that requires no intermediate supervision. Experiments on diverse spatial benchmarks show that this approach outperforms both hard and fixed‑soft chain‑of‑thought baselines and several existing LVLMs.
By Rafi Ibn Sultan, Md. Sajid Alam Chowdhury, Saleh Zare Zade, Chengyin Li, Prashant Khanduri, Marco Brocanelli, Dongxiao Zhu
The paper investigates a critical flaw in AI coding-agent systems such as Claude Code, Codex CLI, and Cursor, where the action approved by a human is not the same as the action executed by the harness. It introduces the concept of Approval Laundering, categorizing six systematic failure modes—Scope, Argument, Temporal, Tool, Delegation, and Semantic laundering—and demonstrates these failures through controlled experiments. The authors propose an Approval Token mechanism that mitigates some laundering types but leaves others unaffected, highlighting the limitations of current enforcement strategies.
By Yang Wang
The paper evaluates an agentic harness that combines a 27B language model, financial calculations, narrative templates, and validation checks to produce corporate and investment banking presentation decks. Using a panel of five judges, the system consistently scores higher than a baseline model that generates directly from a short prompt, with scores ranging from 20.4 to 33.6 out of 95. However, judge variability and changes in grading criteria make it challenging to discern small improvements in deck quality.
By Ludovic Gibert, Matis Despujols, Andre-Louis Rochet
The authors report that a large‑scale experiment using a language‑model coding agent over five weeks produced new lower bounds for DNA‑barcode‑style codes. By restricting searches to codes with a prescribed symmetry, the agent improved the best known code of length 6 and minimum edit distance 3 from 114 to 120 words, and similarly raised lower bounds for lengths 6–9 and distances 3–6. The study also documents failures and the limitations of the verification protocol, noting that intermediate results were never rechecked and could lead to erroneous conclusions.
By Abraham Yeung
The paper introduces Actor‑Critic with Action Chunking (AC2), a method that assigns credit to short action chunks instead of entire trajectories, enabling policy updates without waiting for terminal rewards. AC2 employs local readiness, reference solutions, and 10k‑token chunks to make critic‑based credit assignment reliable. Experiments on Qwen3‑4B with FineProofs‑RL show AC2 surpasses GRPO’s peak validation score while using 2.5× fewer decoding FLOPs and fewer training steps.
By Kaiyue Wen, Luke Bailey, Arvind Mahankali, Tengyu Ma
The technical report introduces Hyperspectral Image Models, a modular framework that unifies 55 deep‑learning models across six paradigms for hyperspectral remote sensing. It standardizes tensor conventions, evaluation protocols, and dataset handling, integrating 24 benchmark scenes from various sensors and providing tools to avoid train‑test overlap. Experiments across 1,320 model‑scene combinations show that scene difficulty outweighs architecture, with no single paradigm dominating and small models achieving performance comparable to much larger ones.
By Tanishq Rachamalla, Aryan Das, Srishti Kaushik, Swalpa Kumar Roy
The paper presents a controlled comparison of six retrieval-augmented generation (RAG) strategies for scientific question answering on a large arXiv corpus. All pipelines use the same LLM generator and evaluation protocol, differing only in retrieval design—ranging from classic dense retrieval to late‑interaction methods like ColBERTv2. The authors also release a synthetic question dataset and code to enable reproducible, large‑scale evaluation of RAG trade‑offs.
By Bhagyesh Rathi, Eshan Chawla, William B. Andreopoulos
arXiv:2609.39378v1 Announce Type: new
Abstract: Real-world embodied tasks, from everyday activities to professional procedures, require agents to act under physical constraints while tracking evolvin...
By Shulin Tian, Junsu Kim, Shuai Liu, Hao Li, Yujiao Shen, Sihan Li, Zhe Yang, Yeongon Kim, Feiyu Li, Jialin Wu, Yichi Zhang, Wenhui Wang, Runmao Yao, Yuhao Dong, Zhaoxi Chen, Fangzhou Hong, Antonino Furnari, Jingkang Yang, Hongyuan Zhu, Ziwei Liu
The paper introduces Wavelet Flow Matching, a method for generating multivariate time series by applying flow matching to multilevel discrete wavelet coefficients. By working in the wavelet domain, the model captures coarse-to-fine temporal structure implicitly and uses a channel-token transformer to model cross-channel dependencies. Experiments on seven benchmark datasets and four sequence lengths show that the approach matches or surpasses existing methods, especially in Context-FID and discriminative score metrics.
By Lucas Poinsignon, Jorge da Silva Gon\c{c}alves, Samuel Ruip\'erez-Campillo, Julia E. Vogt
Ready2Blend is a method that blends natural-language instructions with learned alignment prompts to enable continual alignment of large language models without retraining the backbone. It uses AlignFormer to map each requirement to a fixed-length prompt stored in a modular bank, while keeping the backbone and prior prompts frozen. The approach achieves 93.1–98.5% of joint‑training performance, retains prior knowledge, and reduces training time by up to 4.3×, also allowing weighted personalization and order‑free composition.
By Jeesu Jung, Hwan Chang, Juseon Do, Jeonghwan Choi, Jinho Choo, Sungwoo Nam, S. K. Hong, Hwanjun Song
ECHO-G is a framework for generating full‑body co‑speech motion for humanoid robots, jointly conditioned on speech audio and timed transcripts. Its Speech‑Grounded Diffusion Transformer (SGDiT) fuses frame‑aligned acoustic features with token‑level linguistic context, preserving distinct granularities while modeling one‑to‑many utterance‑motion relationships directly in robot space. The authors introduce a BEAT2‑derived audio‑text‑robot dataset, a benchmark for co‑speech characteristics, robot‑motion quality, and runtime efficiency, and demonstrate that direct robot‑space generation outperforms human‑motion generation and retargeting pipelines, with joint audio‑text conditioning yielding superior results in both quantitative evaluation and a video‑rating study.
"whyItMatters":"The study provides a new dataset, benchmark, and a demonstrably effective method for generating realistic co‑speech motion directly in robot space, advancing practical humanoid robot interaction."
By Yizhao Li, Pusen Gao, Ming Wang, Shaojie Shen, Shuo Yang, Hao Xu
OpenCollab is a multi‑agent coding framework that unifies organization design, enforces experimental control on a shared runtime, and tracks execution via fine‑grained event streams. It introduces the metric Adherence to measure whether the declared organization is actually realized, showing that small configuration changes can shift adherence from 47.2% to 97.2%. Experiments demonstrate that a two‑coder workflow built on OpenCollab achieves new state‑of‑the‑art performance against mainstream harnesses while using the fewest tokens, and that a well‑designed organization can outperform strong existing harnesses.
By Chun-Wah Hsu, Kai Gong, Yu Wu, Xianhe Chen, Mengyang Liu, Jie Li, Hanyu Li, Zhixuan Liu, Naisheng Tang, Jiaying Chi, Ziheng Fan, Xuning He, Xiaokang Yang, Xue Jiang, Yihong Dong
The paper introduces CoT-Interpretability Alignment (CIA), a metric that quantifies how well a large language model’s chain-of-thought (CoT) explanations match its internal reasoning processes. Evaluated on two-hop question answering, hint intervention, and integer multiplication across three LLMs, the study finds limited alignment (44.8–75.9%) and demonstrates that post‑training with a reward combining task accuracy and parametric faithfulness can substantially improve CoT faithfulness without sacrificing accuracy. The authors provide a framework for auditing CoT faithfulness and a pathway to making explicit reasoning more trustworthy, with code and data publicly available.
By Yihuai Hong, Shauli Ravfogel, Chen Zhao, Eunsol Choi