The paper introduces GAP-DPO, a method for personalizing large language models by selecting preference pairs based on gradient alignment with user utility. It formalizes personalized preference learning as a geometry‑aligned optimization problem, showing that off‑policy sampling can shift DPO updates from error correction to reinforcement when preference margins align with utility gradients. Experiments demonstrate that GAP‑DPO improves stylistic fidelity, preference alignment, and overall generation quality over standard DPO variants.
By Ruoming Jin, Xinyu Li, Hao Zhou, Jianfeng Zhu, Ruixin Guo, Feodor Dragan, Lei Xu, Haixun Wang, Yang Zhou
The paper investigates how long-horizon language agents use external memory, noting that memory tends to concentrate on a small core while rare states form a long tail where prediction errors accumulate. Through a tail audit, the authors find that random-walk policies produce log-normal retrieval artifacts, whereas semantic LLM policies generate truncated power-law core–tail traces. They introduce the Core–Tail World Model (CTWM), a rank-based memory controller that allocates prompt budget using a single exponent while summarizing the tail, achieving token savings and reduced tail prediction error across several benchmarks.
By Xinyuan Song, Zekun Cai
LensVLM is an inference framework and post‑training recipe that lets Vision‑Language Models (VLMs) process compressed images of text by selectively expanding only the relevant parts back to full resolution. Using Qwen3.5‑9B‑Base, LensVLM achieves accuracy comparable to full‑text models at 4.3× compression and outperforms other compression baselines up to 10.1× across seven text QA benchmarks, while also improving performance on multimodal document and code tasks as compression increases.
By Roy Xie, Dan Friedman, Donghan Yu, Bowen Pan, Christopher Fifty, Jang-Hyun Kim, Xianzhi Du, Zhe Gan, Vivek Rathod, Bhuwan Dhingra
The study investigates whether the organization of mathematical training data for large language models (LLMs) should be based on topic or on reasoning approach. Using two counterbalanced experimental designs, the authors compare fine‑tuned models trained on problems that share the same solution method (same‑approach, SA) versus those that share the same topic (same‑topic, ST). Across five base models and multiple training seeds, SA consistently outperforms ST, with mean gains of 10.8–14.3 percentage points, even though ST sources are more similar to targets in embedding and lexical metrics.
By Sajad Goudarzi, Samaneh Zamanifard, Seyed Amin Seyed Haeri, Moloud Nasiri, Hamed Rahimian
VeriHarness is a method that enhances verification for large language model agents tackling long‑horizon tasks without needing reference answers at test time. It transforms the base LLM into an agentic verifier by providing a workspace, evidence tools, and reusable verification skills, using disagreement resolution and consensus challenge to evaluate competing claims. Across five benchmarks and two frontier models, VeriHarness outperforms baselines, achieving significant performance gains and demonstrating self‑improvement of verification skills from failure feedback.
By Caiqi Zhang, Rujun Han, Zifeng Wang, Zoey CuiZhu, Nigel Collier, Tomas Pfister, Chen-Yu Lee
arXiv:2609.39082v1 Announce Type: new
Abstract: As new evidence arrives, a sequence model must update what it remembers and how memory influences predictions. While Transformers incur computation and...
By Wentao Wang, Hengyu Zhong, Yunhan Jiang, Jialiang An, Meng Lu
The paper introduces Decision Titan, a variant of the Decision Transformer that incorporates Test‑Time Training (TTT) layers to store episodic memories in network parameters. It evaluates this architecture on the X‑Maze environment, showing that Decision Titan can learn long‑term dependencies up to 20 times longer than its context window and generalise to sequences 1.7 times longer than the training data. The study also finds that temporal generalisation depends on the choice of time embeddings and that the ability to learn long‑term dependencies hinges on how relevant information is encoded.
By Jude Waide, Robert Lieck
VideoEvolve is a framework that automatically evolves agent harnesses for video temporal grounding, a task that localizes events in videos based on natural-language queries. It introduces a Cloze-Structured Harness Representation to keep stage interfaces stable while allowing agent workflows and instructions to change, and uses Branch-Guided Harness Evolution to preserve promising code branches, guide local edits with execution feedback, and validate improvements. Experiments show that this automated evolution improves grounding performance across multiple benchmarks, with instruction refinement consistently yielding gains.
By Bingjun Luo, Yuhuan Fan, Jialin Guo, Siqi Li
TopK-Guided is a training‑free method that improves activation sparsity for large language model inference by combining token‑level sparsity adaptation with block‑level budget allocation that accounts for block sensitivity. It addresses limitations of existing methods like TEAL, which adapts sparsity per token but lacks tight control, and WINA, which enforces a fixed sparsity across all tokens and blocks. Experiments on Llama‑2 and Llama‑3 show that TopK‑Guided consistently yields better perplexity and downstream accuracy while maintaining similar compute costs to WINA, especially at high sparsity levels.
By Mukund Agarwalla, Chih-Jen Lin
The paper evaluates legal text classification models for Korean sexual offense cases, comparing traditional machine learning, large language models, and fine‑tuned domain models. Fine‑tuned KLUE‑BERT achieved the highest accuracy of 99.3%, outperforming GPT‑3.5, GPT‑4.0, and other traditional approaches. Explainable AI techniques were used to analyze predictions, revealing linguistic features that influence decisions and highlighting limitations in capturing subtle textual cues, especially in real‑world KICS data.
By Jeongmin Lee
The paper investigates token‑level certainty as a proxy for correctness in large language models. It finds that certainty better predicts whether a model will answer a question correctly than it does whether a specific response is correct, and that certainty varies by token type and position. The authors show that using certainty early in generation to allocate responses and later to weight votes improves accuracy while dramatically cutting token cost.
By Yunfan Zhou, Ye Zhu, Zhihai Wang, Jianguo Yao, Haibing Guan, Xijun Li
Tokenized Key-Gated Adapter Routing (Locket) is a framework that embeds fine‑grained, policy‑driven access control into large language models by training lightweight LoRA adapters for different privacy policies. A gating module associates a learned keyed entry token with a specific adapter, allowing authorized tokens to unlock private knowledge while invalid or missing tokens trigger privacy‑preserving adapters that redact or sanitize sensitive content. Experiments on datasets such as Enron, ECHR, and Yelp with models like Qwen3, Llama‑3.2, and Gemma‑2‑2B show that Locket maintains perplexity comparable to fine‑tuning when the correct token is provided, and significantly reduces PII leakage when the token is absent or invalid, without sacrificing utility.
By Mohamed Shaaban, Mohamed Elmahallawy
arXiv:2605. 27901v2 Announce Type: replace-cross Abstract: Chain-of-thought (CoT) monitoring has been proposed as a promising safety mechanism for detecting misaligned behavior in large language models.
By Eric Onyame, Runtao Zhou, Kowshik Thopalli, Bhavya Kailkhura, Chirag Agarwal
DuplexSpeechBench-Document Grounding (DSB‑DG) is a benchmark that evaluates how well voice agents can ground their responses in external documents across five professional domains. It focuses on three failure modes—Context Saturation, Grounding Decay, and Proactive Grounding—and includes 1,636 adversarially verified QA pairs from 50 documents. The benchmark measures grounding accuracy, hallucination, and response latency, revealing that cascaded pipelines achieve the highest accuracy while open‑weight systems suffer from abrupt capacity collapse and multi‑turn decay.
By Puneet Mathur, Nedim Lipka, Zeyu Jin, Dinesh Manocha
LEGO-OPD introduces a factorized teacher composition for multimodal on‑policy distillation, combining a Language Expert and a Grounding Expert into a single teacher distribution. By treating the language expert as a prior over tokens and the grounding expert as a visual likelihood that updates this prior, the method decouples language reasoning from visual grounding. Adaptive calibration further adjusts the influence of visual evidence at each decoding prefix, preventing over‑ or under‑supervision. Experiments with Qwen3 models demonstrate that LEGO‑OPD outperforms both single‑ and multi‑teacher baselines on multimodal and text‑only reasoning tasks, improving visual perception while preserving language reasoning.
By Jaeyun Shin, Hangeol Chang, Jong Chul Ye
The paper investigates how the set of prior iterates influences success in large language model (LLM)-driven discovery tasks. It introduces 12 new harnesses called 'Modular' and evaluates them on five diverse discovery problems, revealing that success is fragile and highly dependent on harness design. The study identifies mode collapse—a sharp loss of iterate diversity—as a common failure, and shows that early discoveries predict final outcomes, leading to a new initialization strategy that consistently improves performance across harnesses and applications.
By Mansi Sakarvadia, Marco Ciccone, Colin Raffel
The paper investigates the relationship between a large language model’s internal probability distribution and its verbalized confidence statements. By systematically manipulating training and in‑context data, the authors show that both internal and verbalized probabilities are influenced by distributional and asserted uncertainty in the data. They find that verbalized probabilities align with internal ones beyond what would be expected if they tracked the same sources independently, indicating that verbalized confidence can serve as a probe of the model’s internal distribution.
By Sinead Williamson, Jiaxuan Li, Nick Foti, Russ Webb, Masha Fedzechkina
The paper introduces token‑wise incremental trajectories, a method that tracks how contextual word embeddings evolve as each word is added to a sentence. Using garden‑path sentences, the authors show that these trajectories capture known processing disruptions and can distinguish ambiguous sentences from disambiguated controls. They propose metrics for measuring representational displacement and find that ambiguity signals are present not only in sentence‑level CLS vectors but also in individual token embeddings, suggesting distributed meaning construction across scales.
By Grayson Wycliffe Storer, Julia Witte Zimmerman
The paper investigates challenges in open‑source large‑language‑model (LLM) based multi‑agent systems (MAS). By analyzing 944 issues extracted from 21 projects, it finds that orchestration and execution problems are most common, with workflow, tool integration, and memory issues as primary causes. The predominant remedy identified is optimizing workflow, and the study offers empirically grounded implications for improving orchestration, tool integration, and memory mechanisms in LLM‑based MAS.
By Asad Ur Rehman, Syed Mohammad Kashif, Ruiyin Li, Peng Liang, Zengyang Li, Arif Ali Khan
The paper introduces Diffusion Layer Integrated Gradients (DLIG), a token attribution technique for diffusion language models that extends Integrated Gradients to any layer and denoising step. DLIG tracks how a model progressively commits to a self-generated or fixed completion for a prompt, and it satisfies the IG axioms of completeness, implementation invariance, linearity, and symmetry preservation. The method is lightweight and complements interventional analysis, and the authors apply it to tasks such as word-sense disambiguation, multi-hop graph reasoning, and sentence infilling to show how diffusion language models utilize inputs across positions, layers, and denoising steps.
By Darpan Aswal, C\'eline Hudelot