BiCAA introduces bidirectional credit assignment for search-augmented agents by combining Forward Solvability Gain (FSG) and Subsequent Reasoning Gain (SRG). FSG measures how each search step increases the likelihood of the correct answer, while SRG evaluates how a step supports future reasoning, discouraging redundant retrieval. Experiments on seven open‑domain QA benchmarks and two Qwen backbones show that BiCAA improves average F1, especially on multi‑hop tasks.
By Yibin Huang, Bin Xu, Hailong Cao, Conghui Zhu
The paper investigates test-time scaling for small-scale open-ended reasoning models by introducing semantic cluster voting, a judge-free method that groups model outputs by meaning and selects the largest cluster’s representative. Experiments on a 250-prompt benchmark across three 3‑B models show that voting improves accuracy on math tasks (10–29 point gains) but only marginally on open-ended text (+0.30 points). A larger 200‑problem MATH‑500 run confirms a +8.5 point math improvement, while most open-ended samples fall into a single cluster, indicating limited diversity and that higher decoding temperatures reduce quality.
By Adam Y. Ahmed, Abdullah G. Elafifi
The paper introduces ConfabQA, a 784‑item benchmark designed to distinguish between confabulation and refusal in small language models when faced with unanswerable questions. Using linear probes, a sparse autoencoder, and direct optimization of intervention directions, the study maps the hidden‑state geometry on models such as Qwen3‑1.7B, Gemma 2‑2B, and Llama 3.2‑3B, showing that certain hidden‑state directions can flip confabulations into refusals. The work also demonstrates that the refusal signal is model‑dependent and that the hidden state’s contribution can be quantified beyond prompt features.
By Vladimir Kazei
DuSE-Bench is a new benchmark for evaluating how language models combine evolving dialogue and document facts to meet user requirements. It contains 1,000 tasks across five scenarios, with 700 tasks requiring premises from both dialogue and documents, and uses an evolving Support Graph to track changes and evidence links. Experiments on 11 baseline models show that joint reasoning remains difficult, with accuracy on jointly grounded tasks at most 56.7% and frequent attribution errors.
By Guanjie Cheng, Jianfeng Gao, Yufei Tong, Xinkui Zhao, Shuiguang Deng
The paper presents a method that uses relation knowledge—either manually crafted or learned from training data—to select a compact set of knowledge‑graph facts for large language models (LLMs). It introduces derivability, consistency, and validation checks to determine whether an answer can be derived from the selected facts, whether the facts conflict, and whether the derived answer should replace the LLM’s output. Experiments show that seven open LLMs perform better with the compact context than with relevance‑ranked facts, achieving higher exact‑match accuracy on ComplexWebQuestions and outperforming SubgraphRAG across 53 LLM, benchmark, and prompt settings.
By Seunghui Jwa, Minsu Oh, Sangwoo Hong, Yeo Chan Yoon
The paper examines how calibration rules for open‑ended question answering can be misled by the choice of correctness labels, noting that exact‑match labels can produce misleading confidence estimates. It introduces FactForm, a fact‑grounded form grader that combines reference matching with evidence retrieval, achieving lower error and higher AUROC than traditional graders on a large benchmark. FactForm’s design allows a single LLM call per answer and demonstrates stable performance across runs.
By Zhexi Feng, Bingrui Zhang, Wuxi Chen
The paper introduces FAID, a fine‑grained adaptive framework for detecting implicit hate speech. It first classifies samples into Shallow, Targeted, or Context‑Dependent categories and then applies tailored strategies—Prompt‑Tuning for Shallow, knowledge augmentation for Targeted, and an agentic prompt‑generation system for Context‑Dependent cases. Experiments on four benchmark datasets show that FAID outperforms state‑of‑the‑art baselines.
By Han Wang, Hanqing Tao, Xuesong Wang, Yi Zhu
VeriCypher is an LLM-based framework that generates Cypher queries for question answering over Labeled Property Graphs (LPGs). It uses schema summaries to condition query generation, executes queries, evaluates results with an LLM, and iteratively corrects errors through named-entity verification. Experiments on CypherBench show that VeriCypher consistently improves answer correctness and query robustness compared to a baseline, and ablation studies confirm that named-entity verification corrects hallucinated schema elements and entity values.
By Anton Gusarov, Anastasia Volkova, Valentin Khrulkov, Andrey Kuznetsov, Evgenii Maslov, Ivan Oseledets
The paper introduces MedForget, a hierarchy‑aware multimodal unlearning benchmark that models medical data as nested institutional structures, allowing fine‑grained evaluation of retain and forget splits. Experiments reveal that existing unlearning methods struggle to achieve effective hierarchy‑aware forgetting without harming retained utility on clinically relevant tasks. To address this, the authors propose CHIP, a training‑free method that deletes target‑specific weight subspaces while preserving sibling‑shared information, achieving superior forget‑retain performance across all hierarchy levels while maintaining competitive downstream utility.
By Fengli Wu, Vaidehi Patil, Jaehong Yoon, Yue Zhang, Mohit Bansal
The paper introduces Latent Harness, a method that compiles external agent behaviors into a small rank‑8 activation shift applied to every decoder block of a frozen language‑model backbone, affecting only 0.03‑0.05% of parameters. This compiled shift eliminates the need for on‑call harness text, offers runtime controls for injection strength, phase, and mixing, and recovers most of the performance gains of the original harness across multiple settings. The approach also enables zero‑shot transfer across tasks, composability in shift space, and improves budget‑matched harness search, outperforming LoRA with far fewer parameters.
By Ran Li, Shimin Di, Zeyu Ren, Ling Yue, Jiachuan Wang, Lei Chen
The paper introduces Legible, a fine‑tuning technique that aligns a language model’s internal representation of a newly added fact with the model’s own reading of that fact, thereby restoring the model’s ability to perform two‑hop reasoning. By adding a loss term that forces the bridge phrase to match the model’s internal entity representation, Legible improves the model’s composition of learned facts and significantly boosts bridge accuracy across several large‑language‑model families. The method achieves these gains without requiring multi‑hop training data, outperforming prior approaches such as CaKE and RippleEdits on benchmark tests.
By Zhiyun Shi
FlyWheel introduces optimized TPU kernels for softmax and linear attention, reducing vector work and improving matrix operand reuse. By overlapping independent computations, FlyWheel achieves up to 4.14× speedups over the RPA kernel and 1.80× over Splash Attention, reaching 81% of peak BF16 throughput on TPU v6e. It also boosts Gated DeltaNet prefill by up to 3.61× and enhances vLLM serving of Qwen3.8-27B prefill throughput by up to 1.56× while preserving accuracy.
By Zishuo Bao, Zeshen Zhang, Linkun Chen, Yucheng Lu
The paper introduces Research Attention Prediction (RAP), a benchmark designed to evaluate large language model (LLM) agents’ ability to infer changes across entire research fields from limited evidence gathered during online scholarly searches. RAP is built on a local search environment comprising 356,357 dated arXiv papers and includes two evaluation sets: a Core set of 2,160 questions covering 54 AI/ML fields and an Extended set spanning 269 fields. Experiments show that while search improves agents’ recognition of recent changes, their forecasting of future shifts remains limited or even negative, indicating that LLM agents can detect some recent developments but struggle to reliably anticipate future research activity.
By Yingqian Wu, Jingcong Liang, Shengbin Yue, Siyuan Wang, Zhenfei Yin, Junchi Yu, Zhongyu Wei
The paper introduces a discrete-class framework for detecting human‑AI co‑authorship, distinguishing between superficial AI polishing and deeper AI‑driven discourse changes. It shows that adversarial paraphrasing can obscure detection signals, but augmenting training data with such paraphrases improves detector robustness across multiple datasets and generation models. The study demonstrates that principled data augmentation and a fine‑grained classification approach are essential for reliable co‑authorship detection.
By Ziruo Zhao, Jie Fu, Wendy Hui Wang, Zining Zhu
Strategist is a meta‑agent that automatically evolves inference strategies for AI agents, turning model calls into responses. It stores reusable strategy blueprints in a library, retrieves relevant ones for new tasks, generates and tests candidate strategies, and refines them based on results. Across 30 benchmarks, Strategist outperforms the strongest baseline on 24 tasks, matches its average accuracy while cutting inference cost by 39%, and its strategies transfer to unseen tasks and models without additional search.
By Taiming Lu, Qi Xu, Zhigang Hua, Zhuang Liu
DiabAgent introduces a research suite for agentic physiological control in type 1 diabetes, comprising the APC framework, DiabetSim simulator, and SweetBench benchmark. The APC framework uses medical-device tools and ongoing observations to infer patient states, accumulate experience, and select treatment actions. In virtual testing with DiabetSim, APC outperforms the open-source baseline OpenAPS, achieving a COGI of 73.5 versus 71.2, and ablation studies show that individual adaptation and contextual information improve glucose control.
By Chengxi Li, Shaoxuan Li, Deyao Hong, Qingle Liu, Yi Huang, Puheng Bian, Minzhi Xu, Lanyue Zhang, Mingze Sun, Xin Li, Calvin Xiao, Shengyuan Wang, Huan-ang Gao, Jiazheng Xu, Kaisen Yang, Hongning Wang
The paper introduces D&A (Describe & Assess), a prompt-based method that separates progress estimation into a textual description of scene dynamics and a subsequent assessment using explicit rubrics. By requiring Vision‑Language Models to generate an explicit definition of absolute success before scoring, D&A anchors evaluations to an external criterion, reducing confirmation bias and providing fine‑grained, partial progress metrics. Empirical results show that D&A matches or surpasses post‑trained reward models across tasks, embodiments, and benchmarks, and its reward estimates support efficient downstream reinforcement learning.
By Marius Memmel, Shawn Yin, Urvi Rutia, Angad Wadhwa, Eldon Li, Gavin Angeloff, Ankit Goyal, Jesse Zhang, Dieter Fox, Fabio Ramos, Anqi Li, Abhishek Gupta
AStockArena is a real‑time decision‑auditing benchmark for continuous trading in China’s A‑share market, evaluating 13 open‑ and closed‑source models over 3,198 sequential decisions. The benchmark reconstructs actual account actions by aligning reports, trade requests, execution responses, and account states, then audits them with eight checks covering execution traceability, evidence sufficiency, and financial meta‑awareness. Across 1,070 trading points, 72.99% show at least one audit gap, with evidence gaps most common, and historical replay demonstrates that flagged trades significantly affect portfolio outcomes.
By Zeyu Liu, Sing Kwong Lai, Xuanzhi Feng, HaoYang Shang, Yang Li, Zihao Fang, Xuan Liu
The paper introduces GAVAGAI, a method that adds a referential alignment variable to vision‑language models by solving a small assignment problem via entropic optimal transport. By allowing words to be assigned to a special NOTHING column and enforcing a capacity constraint on regions, the approach improves word‑to‑picture accuracy and localization on the SAYCam infant video corpus. Experiments show that the alignment loss boosts performance over a standard next‑token captioning baseline, especially under high non‑reference rates.
By Rajdeep Pal, Jotiraditya Banerjee, Saptarshi Pani
SWE-Opt is a new benchmark that evaluates large language models (LLMs) on repository‑level software performance optimization. It is built from 102 expert‑written optimizations from open‑source projects and measures runtime, peak memory, and Time‑Weighted Memory Usage (TWMU) under varied input data, execution conditions, and noise‑aware measurements. The study finds that current LLMs achieve few runtime gains and almost no memory improvements, whereas expert implementations deliver a 15.5× speedup and 171.3× memory reduction.
By Ezgi Sar{\i}kayak, Wenchao Gu, Hesham Ghonim, Chunyang Chen