The paper introduces MCCE, a hybrid framework that combines a frozen closed‑source large language model (LLM) with a lightweight, trainable model for multi‑objective discrete optimization. By maintaining a trajectory memory and refining the small model through reinforcement learning, the two models jointly enhance global exploration and learning. Experiments on drug‑design benchmarks demonstrate that MCCE achieves state‑of‑the‑art Pareto front quality, outperforming existing baselines.
By Nian Ran, Zhongzheng Li, Yue Wang, Qingsong Ran, Xiaoyuan Zhang, Shikun Feng, Richard Allmendinger, Xiaoguang Zhao
The paper introduces Residual Reward Models (RRM) to enhance preference‑based reinforcement learning (PbRL) in robotics. RRMs decompose the true reward into a prior component—such as a heuristic, language‑generated, or IRL‑derived reward—and a learned residual that is trained with human preferences. Experiments on Meta‑World, DM‑Control, and a physical Franka Panda robot show that RRMs markedly improve sample efficiency and accelerate policy learning compared to standard PbRL methods.
By Chenyang Cao, Miguel Rogel-Garc\'ia, Mohamed Nabail, Xueqian Wang, Nicholas Rhinehart
The paper investigates the reliability of rule- and model-based verifiers used in reinforcement learning with verifiable reward (RLVR) for mathematical reasoning. It finds that rule-based verifiers often miss equivalent answers in different formats, causing false negatives that degrade RL performance as models improve. Model-based verifiers achieve higher static accuracy but become vulnerable to reward hacking during RL, misclassifying certain response patterns as correct after fine-tuning.
By Yuzhen Huang, Weihao Zeng, Xingshan Zeng, Qi Zhu, Junxian He
RubricRM introduces a pairwise generative reward modeling framework that generates an input‑specific rubric—comprising evaluation dimensions, weights, and scoring criteria—to score candidate images. The method is trained in two stages: supervised fine‑tuning to learn the rubric‑based scoring paradigm and GRPO to refine dimension‑level rewards. Experiments on text‑to‑image generation and instruction‑based image editing benchmarks demonstrate that RubricRM outperforms existing specialized reward models and competes with strong proprietary MLLM judges while using smaller backbones.
By Zijian Kan, Wei Wang, Long Luo, Bing Zhao, Xuan Ren, Weixu Qiao, Wenbo Li, Hu Wei, Lin Qu
FlavourBench is a new evaluation framework that replaces missing answer keys with dense reward maps generated from a versioned culinary environment. Each task requires selecting a three‑ingredient portfolio from eight options, and all 56 possible portfolios are scored by Epicure before model inference. The authors evaluate 27 endpoints across 534 tasks, perform statistical tests on 351 model contrasts, and conduct a preregistered post‑training study showing that LoRA fine‑tuning on Qwen3‑0.6B improves performance on 84 anchor‑disjoint maps by 13.30 points.
By Josef Chen (Independent Researcher), Erim Hayretci (Imperial College London)
The authors present the Cross-Platform Fairness Evaluation (CPFE) framework, a five‑axis audit protocol that assesses discriminative performance, calibration, statistical significance, prediction equity, and attribution stability of transformer models. Applying CPFE to four models trained on a Kaggle mental‑health corpus and tested on Reddit and Twitter, they find substantial cross‑platform degradation in AUC (30–40%) and severe calibration failures (ECE rising to 0.5 on Twitter). The study demonstrates that platform‑specific temperature scaling can largely fix calibration without harming discrimination, while prediction equity and attribution stability analyses reveal significant disparities and vocabulary divergence across platforms. The results argue that cross‑platform validation across all CPFE axes should become a standard requirement for mental‑health NLP systems deployed in heterogeneous environments.
By Rajveer Singh Pall, Sameer Yadav
The paper introduces Echo-GRPO, a method that rewrites privileged reasoning traces into a model’s own idiolect to align off‑policy supervision with the student policy’s vocabulary. By preserving semantics through Dual‑Reference Decoding, Echo‑GRPO mitigates gradient clipping on critical reasoning tokens and improves reasoning distillation. The approach is instantiated as VideoEcho‑R1 for video reasoning, yielding consistent gains across multiple multimodal LLM backbones and benchmarks, and it can be applied as a plug‑in to both RL and supervised fine‑tuning frameworks.
By Ji Soo Lee, Jinyoung Park, Seohyun Lee, Jongha Kim, Joonmyung Choi, Jinsung Yoon, Hyunwoo J. Kim
MemToC is a controlled benchmark that tests how large language models resolve conflicts between their internal memory and tool outputs. It contains 6,504 episodes built from 542 factual questions, each paired with a model‑generated closed‑book answer and a tool return whose correctness is known, creating four distinct source‑correctness scenarios. Across five 7‑9B open‑weight models, tool responses overwhelmingly dominate closed‑book answers, and only a minority of instruction‑tuned models correctly retain a verified answer when the tool is wrong, while most follow a correct tool or repeat a wrong tool.
By Arseniy Varlamov, Rishat Zinnatullin, Elisei Rykov, Alexander Panchenko, Ilseyar Alimova
PercepCap is a video captioning framework that explicitly models spatio‑temporal perception before generating captions. It follows a perceive‑describe chain, first producing a perception trace of object trajectories and temporal events, then generating the final caption conditioned on that trace. The method uses a two‑stage training strategy—supervised fine‑tuning followed by perception‑grounded reinforcement learning—and builds caption‑aligned perception data to ensure the perception trace and caption refer to the same objects and events.
By Yifan Xu, Zihao Wang, Zhixiao Wang, Jiaming Zhang, Yichun Yang, Desen Meng, Yuanxing Zhang, Pengfei Wan, Limin Wang
The paper introduces Thomson, a frontier AI model developed through continual learning on open-weight models, aiming to democratize access to high-performance AI. It argues that institutions with limited resources can achieve frontier-level performance by applying a modern mid- & post-training stack, preserving model plasticity and stability while minimizing high-impact interventions. Thomson demonstrates competitive performance across agentic tasks, safety, legal, tax, multilingualism, and large-scale deep research, exhibiting a distinctive π-shaped improvement pattern and effectively mitigating the forgetting problem seen in narrow domain adaptation.
By Shengzhuang Chen, Jerrod Parker, Yejin Bang, Andrew M. Bean, Nabeel Seedat, Stefan Winzeck, Daniil Glazko, Jannik Zgraggen, Fangyi Yu, Scott Arnott, Dietrich Trautmann, Luca Ciuffreda, Guglielmo Bonifazi, Davide Romano, Bradley Bell, Kirsty Fielding, Daniele Giofr\`e, Tom Zielund, Ipshita Chatterjee, Sneha Murthy Ghantasala, Manpreet Nanreh, John Scoville, Maciej Sakowicz, Wassim Seifeddine, Lukas Thede, Jonathan Richard Schwarz
The paper introduces ASIL, an Agent‑Software Interaction Layer that replaces traditional screenshot‑and‑click interfaces with structured JSON observations and code‑executable semantic actions. ASIL is implemented across 15 applications and evaluated on 300 single‑application and 80 multi‑application tasks, achieving over 80% success with fewer than five actions per task. The structured interface also improves training efficiency, boosting performance of Qwen models from 58–66% to 72–80% with small‑scale supervised fine‑tuning and further gains with on‑policy reinforcement learning.
By Rui Xie, Lu Chen
The paper proposes a scalable, automated method for auditing candidate‑job matching systems for demographic bias. It employs large‑language‑model agents to generate neutral resumes, injects controlled demographic variations, ranks candidates with a fine‑tuned embedding model, and evaluates nine fairness metrics across counterfactual, group‑fairness, and merit‑aware families, producing a composite risk report. Experiments on a small corpus show that single‑score audits miss nuanced issues, underscoring the need for multi‑metric evaluation and LLM‑generated audits as a low‑cost complement to human reviews.
By Sai Yashwant, Shruti Bansal, Anurag Dubey, Samaroha Chatterjee, Satyam Kumar, Shreyash Gupta, Gantala Thulsiram
The paper introduces FrameFT, a parameter-efficient fine-tuning method for transformer models that represents weight updates using sparse coefficients in a Fusion Frame basis. This approach reduces memory usage by storing only the sparse coefficients, leading to significant compute advantages and formal convergence guarantees. Experiments on language and vision tasks show that FrameFT matches or surpasses state‑of‑the‑art PEFT techniques while requiring far fewer trainable parameters.
By Harshavardhan Adepu, Li Zhang, Sanjiv Kumar, Vikas Singh
RATIO (Retrieval Across Typed Ideation Operations) is a large-scale benchmark designed to evaluate how well retrieval systems can support scientific inspiration. It defines relevance through three ideation moves—Address, Broaden, and Specify—each targeting different levels of abstraction in literature retrieval. The benchmark is built from millions of full-text CS papers using a novel discourse-marker distant supervision method, and includes extensive LLM and human vetting to ensure quality.
By Maayan Sharon, Tom Hope
The article reports that large language models can predict and collaboratively modulate human memory search during a semantic fluency task. By tracking and forecasting participants’ semantic retrieval patterns, the models outperform other humans in following these mental trajectories. This suggests that AI can serve as a cognitive tool to extend human abilities in open‑ended conceptual exploration and creative ideation.
By Eric Lacosse, Mariana Duarte, Graham Todd, Peter M. Todd, Daniel C. McNamee
The paper introduces a constraint‑aware physics‑informed neural network (PINN) for estimating the static shape of co‑manipulative continuum robots (CCRs). By integrating projected static equilibrium and geometric loop‑closure residuals, the PINN enforces both mechanical and closed‑chain constraints. Experiments show that, even with limited and noisy data, the PINN outperforms a purely data‑driven ANN, achieving markedly lower configuration, equilibrium, and closed‑chain errors, and delivering inference times orders of magnitude faster than traditional nonlinear solvers.
By Rana Danesh, Pari Qarehdaghi, Farrokh Janabi-Sharifi
The paper introduces sLoTh, a parameter‑efficient continual learning framework for sparse event‑based vision transformers. sLoTh freezes the backbone and limits plasticity to low‑rank attention updates (seLoRA) and shared neuronal threshold modulation, updating less than 1% of parameters without replay buffers. Experiments on CIFAR‑100, Tiny‑ImageNet, ImageNet‑100, and ImageNet‑R show competitive rehearsal‑free performance across up to 100 tasks while achieving roughly 6.5× lower energy consumption than dense vision transformers.
By Vaishnavi Nagabhushana, Kartikay Agrawal, Ayon Borthakur
The paper presents a new method for automated creativity assessment that uses Poly-Encoders to achieve high accuracy while remaining computationally efficient. By fine‑tuning a Poly-Encoder on a public dataset of 18,000 human‑rated responses from the Scientific Creative Thinking Test, the authors obtain Pearson correlations up to r = 0.74, comparable to large language models. The approach relies on small pre‑trained BERT encoders, dramatically reducing computational demands and enabling potential deployment on consumer‑grade hardware.
By Sam Grouchnikov, Phillip Gregory, Jiho Noh
The paper proposes a method called Modular Expert Merging for Biomedical Retrieval, which combines independently trained domain‑specialized experts instead of large mixed‑domain training. Experiments across four decoder‑only LLM families (0.6B‑7B) and twelve retrieval tasks from MTEB show that merging experts consistently outperforms mixed‑domain training. The authors also introduce a Synthesize‑Train‑Merge (STM) framework that generates hard negatives with a top‑tier LLM, fine‑tunes experts via LoRA, and merges them, achieving strong biomedical retrieval performance while retaining competitive general‑domain results.
By Sameh Khattab, Jean-Philippe Corbeil, Osman Alperen \c{C}inar-Kora\c{s}, Amin Dada, Julian Friedrich, Jiawei He, Douglas Teodoro, Jens Kleesiek
The paper introduces a unified conditional-flow framework that integrates text-driven motion generation, semantic editing, and intra-structural retargeting into a single rectified-flow model. By treating editing as a change in semantic condition and retargeting as a change in skeletal condition, the approach eliminates fragmented pipelines and allows a single model to perform generation, zero‑shot editing, and zero‑shot retargeting on articulated 3D motion data. Experiments on SnapMoGen and a Mixamo subset demonstrate that the model can handle all three tasks without task‑specific fine‑tuning, preserving both motion semantics and skeletal structure.
By Junlin Li, Xinhao Song, Siqi Wang, Haibin Huang, Yili Zhao