The paper introduces a method for generating multilingual reasoning traces for medical question answering using large language models. It creates 500,000 reasoning traces in English, Italian, and Spanish by retrieving medical information from Wikipedia and applies them to MedQA and MedMCQA datasets extended into Italian and Spanish. The approach improves performance in both few‑shot in‑context learning and supervised fine‑tuning, achieving state‑of‑the‑art results for 8B‑parameter LLMs and releasing all resources for further research.
By Pietro Ferrazzi, Aitor Soroa, Rodrigo Agerri
The paper introduces DroneCATS-Agent, a modular framework that places a multimodal large language model (MLLM) at the core of a drone’s control loop, allowing the model to decide actions solely from prompts. It presents the DroneCATS benchmark, evaluating MLLMs on four tasks—approaching, tracking, searching, and multi‑drone commanding—without fine‑tuning or function‑calling. Results show that while small open models can navigate reliably, they often fail by mismanaging protocol termination, highlighting a gap between perception and action planning in current MLLMs.
By Jaewoo Park, Minyoung Lee, Sukmin Seo, Moonbin Yim, Hyunwook Yoon, Dohoon Ryu, Daehee Kim, Myungseo Song, Jihyuk Byun, Seunggyu Chang, Taeho Kil, Jiseob Kim, Bado Lee, Geewook Kim
Neurosymbolics for Data Engineering introduces a neurosymbolic layer that can be added to existing LLM backbones to improve logical reasoning and reduce long‑context token usage. The layer boosts accuracy by an average of 85% on benchmarks such as BIRD‑CRITIC and LiveSQLBench without any task‑specific finetuning or RLHF. It also cuts effective token usage by over 50% and lowers time complexity from O(n²) to roughly O(n) for long‑context tasks.
By Vishvesh Bhat
arXiv:2609.01560v1 Announce Type: cross
Abstract: We present H3-World, an efficient framework that turns the 33B MiniMax-H3 video generator into an interactive world model. Our key finding is that, a...
By Danze Chen, Zeqing Wang, Ziyue Lin, Xingyi Yang, Yeying Jin
The paper introduces WHALE, a method that alternates between updating a language model’s weights and searching for a better harness (the code that manages context and control flow). By iteratively fine‑tuning the model under the current harness and then optimizing the harness under the updated model, WHALE improves performance across search QA, math reasoning, and chess puzzles, outperforming weight‑only, harness‑only, and Fast‑Slow Training by 4.15–24.38 percentage points in mean@8 accuracy. The approach uses either fixed phase lengths or an adaptive patience rule to decide when to switch phases, and the authors provide code on GitHub.
By Haechan Kim, Yoonho Lee, Gisang Lee, Chelsea Finn, Kangwook Lee
The paper introduces CoLT-Drive, a 3,536-sample counterfactual long‑tail benchmark for evaluating decision‑level driving affordance prediction, which tests whether models can infer how rare objects affect an ego vehicle’s high‑level actions. It also proposes KPA, a knowledge‑preserving adaptation framework that combines structured prompting, expert merging, and a regime‑aware LoRA mixture‑of‑experts module to improve small VLMs on driving tasks. Experiments show KPA achieves 60.8% pair accuracy on CoLT‑Drive, outperforming the Qwen3‑VL‑2B baseline and LoRA SFT while keeping competitive in‑domain performance.
By Zhengxu Tang, Guofeng Cui, Ziyu Gong, Xiaozhou Zhang, Ruifeng Deng, Chengzhi Qi, Ke Chen, Sachin Patil, Tianjun Xiao, Langechuan Liu, Pichao Wang
The paper introduces the Prospective Intention Store (PIS), a method that places lifecycle logic in code and confines language tasks to a typed action space, enabling small models to perform prospective memory tasks more effectively. Using PIS, a small model (DeepSeek-Chat) achieves 82.9% Set‑F1 on PM‑Bench, surpassing the previous best of 65.1%. On Gemma‑E2B, PIS boosts Set‑F1 from 4.2% (without a store) to 66.2%, and reaches 70.1% Set‑F1, outperforming retrospective memory approaches that max out at 54.4%.
By Jinqing Zhao, Chengcan Wu
CoBRA is a counterfactual boundary‑learning framework designed to improve when a tool‑augmented language model should call an external tool. It builds internal and external experts from the same base model, collects paired trajectories, and estimates the reward margin between answering with and without tools. Using these margins, CoBRA partitions data into internal‑favored, external‑favored, and ambiguous cases, then applies Boundary‑Aware Cold‑Start SFT and MARS‑RL to optimize boundary decisions, leading to more efficient tool use and better accuracy on tool‑dependent out‑of‑distribution questions.
By Wenhao Zou, Xianglong Liu, Wendong Bi, Hanjie Wang, Simin Zhao, Gong Zhi
GazeRefine is a training‑free framework that uses eye‑gaze data as an inference‑time prompt for zero‑shot medical image segmentation. It converts sparse, duration‑weighted fixations into foreground and background priors that initialize semantic prototypes in a frozen DINOv3 feature space, then iteratively refines these prototypes through discrimination, affinity propagation, and anchoring to the gaze guidance. The method achieves strong results on colonoscopy polyp segmentation and competitive performance on prostate MRI, demonstrating that gaze‑guided prototype refinement can enable segmentation without dense expert annotations or model fine‑tuning.
By Mohammed Oussama Benyahia, Marouane Tliba, Mohamed Amine Kerkouri, Taifour Yousra, Bin Wang, Max Bengtsson, Gorkem Durak, Elif Keles, Zuheng Ming, Marek Penhaker, Azeddine Beghdadi, Ulas Bagci, Aladine Chetouani
arXiv:2609.01604v1 Announce Type: cross
Abstract: LLM-based evaluators of natural language generation (NLG) quality are widely deployed as scoring tools and as automated training signals, yet the int...
By Himil Vasava, Ming Jiang
The paper investigates a privacy‑hallucination tradeoff in differentially private (DP) language models. It shows that DP pre‑training or fine‑tuning increases hallucinations compared to non‑DP models, especially as the privacy budget becomes stricter. The authors attribute this to DP mechanisms flattening output distributions, and demonstrate that controlling the frequency of facts in training data can mitigate hallucination risks.
By Krithika Ramesh, Krishna Pillutla, Danish Pruthi, Anjalie Field
The paper investigates why fine‑tuned reasoning models lose coverage, observing that pass@k accuracy degrades relative to the base model. The authors attribute this shrinkage to decision‑point or “forks in the road” scenarios in the fine‑tuning data, where the model faces multiple valid reasoning paths. Controlled experiments confirm a strong correlation between such decision‑point prevalence and coverage loss, and show that targeted data synthesis and diversity‑encouraging decoding can partially mitigate the effect.
By Ngoc-Hieu Nguyen, Parshin Shojaee, Phuc Minh Nguyen, Nan Zhang, Chandan K Reddy, Khoa D Doan, Rui Zhang
The study evaluates large language models (LLMs) as rerankers in conversational movie recommendation, comparing proprietary, open-weight, and fine-tuned LLMs against collaborative-filtering and sequential baselines on the ReDial benchmark. Results show that the best proprietary LLM achieves an NDCG@10 of 0.1497 with a shared semantic candidate pool, outperforming non-LLM baselines, while open-weight LLMs do not surpass a tuned shallow autoencoder under the same protocol. The analysis also highlights that reranker performance is highly sensitive to candidate generation, pool size, scoring policy, and decoding temperature, suggesting these factors should be reported as standard evaluation fields.
By Ante Kapetanovic, Tomislav Duricic, Andro Mercep, Emanuel Lacic
The paper introduces MiNER, a fine‑tuned biomedical NLP system that uses BioBERT to extract malaria‑related named entities from scientific literature. It builds a large, annotated corpus of malaria articles, preprocesses the text, and applies supervised learning to improve extraction performance. Experiments show that MiNER outperforms other encoding and machine‑learning methods in precision, recall, and accuracy, and the authors release the human‑labeled dataset for further research.
By V. S. Anoop, Devika N
The paper investigates how multimodal large language models (MLLMs) handle conflicting evidence presented in text, image, or both forms. Across 13 MLLMs and two datasets, the authors find that models are not robust to knowledge conflict: they tend to accept contradictory image evidence more readily than contradictory text, and when both modalities conflict the preference is arbitrary, depending on input order, model, and dataset. The instability degrades multimodal retrieval-augmented generation and can be exploited by adversarial attacks, while simple mitigation techniques such as prompting, steering, and direct preference optimization largely fail, with supervised fine‑tuning offering only moderate improvement.
By Jungyeon Lee, Yejin Yoon, Taeuk Kim
Online Self-Weighted Fine‑Tuning (OSW‑FT) augments standard supervised fine‑tuning by adding online, trajectory‑level weighting: for each query the model estimates its current success rate from a small number of inference‑only rollouts and rescales the SFT loss accordingly. The method keeps the optimization direction anchored to the expert trajectory while adapting the update magnitude online, and it is shown to be unbiased for any finite rollout count with a convergence analysis. Across Qwen3 models from 0.6B to 4B, OSW‑FT consistently outperforms plain SFT on challenging benchmarks such as AIME, achieving a favorable compute‑performance trade‑off with only two online rollouts.
By Haiquan Wen, Yiwei He, Bei Peng, Guangliang Cheng
arXiv:2609.00231v1 Announce Type: new
Abstract: Existing research on object hallucination in multimodal large language models (MLLMs) predominantly attributes the problem to language priors such as o...
By Peiyang Xu, Xiaopei Zhu, Jun Zhu, Xiaolin Hu
The study investigates whether post‑training methods—GRPO, SFT, and DPO—improve language models’ ability to follow prompt evidence that conflicts with memorized knowledge. By comparing nine training variants across different scales and families, the authors find that grounding gains are modest for GRPO, moderate for Conflict‑SFT, and near‑ceiling for DPO, but all largely rely on the same causal attention‑head set present in the starting checkpoint. Removing the starting‑model grounding direction suppresses these gains, while adding it back recovers a significant portion of DPO’s improvement, indicating that existing model machinery drives most of the observed gains.
By Prakhar Gupta, Vaibhav Gupta
TimeSteer introduces inference‑time speech scheduling for joint audio‑visual diffusion models, enabling users to place speech and visual articulation within specified time intervals without fine‑tuning the backbone. The method leverages two properties of the denoising process: a timing‑sensitive text‑to‑audio cross‑attention head that reveals each utterance’s source span, and a predicted clean latent that already organizes coupled speech and visual content. TimeSteer localizes each utterance’s source span and remaps the associated audio‑visual latent to the target interval, and the authors present SpeechShift as the first benchmark for interval‑level speech scheduling.
By Chao Zhou, Yiling Chen, Qi Chu, Tao Gong, Nenghai Yu, Tianyi We
arXiv:2609.01016v1 Announce Type: new
Abstract: Current speech synthesis struggles with code-switching, which mixes a foreign language phrase into a primary language utterance, causing the phrase to...
By Che Hyun Lee, Sangkwon Park, Donghun Kang, Dongwook Lee, Youngho Cho, Heeseung Kim, Sungroh Yoon