Hospital discharge summaries often blur when things happened. This framework tags every event in the note with an ID and ties it back to timestamped rows in the patient's records. On 40 critical-care summaries, it recovered 43% more events and came close to clinician annotations. The team also built GAVEL, a model that judges competing timelines, and human reviewers upheld most of its calls.

Authors: Sayantan Kumar, Nicolas Grimaldi, Jack Cummins, Jeremy C. Weiss

Paper: https://arxiv.org/abs/2609.13062v1

Building benchmarks for robots and other embodied AI is slow, and an early mistake can quietly spoil the result. Embodied-BenchForge has agents build the benchmark, check each intermediate piece, and redo only the parts that fail. It produced six question-answering benchmarks and one interactive benchmark with 220 tasks.

Authors: Baoyang Jiang, Fengchun Zhang, Leyuan Wang, Haotian Li, Yida Wang, Zhe Ji, Jinshan Lai, Xi Ren, Danyang Li, Zheng Yang, Jianwei Hu, Qiang Ma

Paper: https://arxiv.org/abs/2609.13082v1

One diffusion model treats language, camera views, goals and robot actions as the same kind of token. That lets it predict actions, the next view and the end state with a single model. Pretrained on about 1.33 million robot trajectories, it averaged 78.4% success on a real Franka arm across four conditions. A faster implementation cut action decoding time by up to 29 times.

Authors: Hoeun Lee, Jaeik Kim, Jusang Oh, Jinhyeok Kim, Geon Choi, Hyeonggeun Kim, Jaeyoung Do

Paper: https://arxiv.org/abs/2609.13053v1

Tabular foundation models are served differently from chatbots, so the usual speedups don't carry over. Here the bottleneck is the attention calculation. Converting queries, keys and values to FP8 gave up to 1.7 times the speed with no meaningful accuracy loss on TabPFN-v3 and TabICLv2. The catch: quantization error on test rows has to match the training rows, or accuracy drops sharply.

Authors: Jonas M. Kübler, Benjamin Jäger, Klemens Flöge, Noah Hollmann, Frank Hutter

Paper: https://arxiv.org/abs/2609.13031v1

Driving policies trained in simulators can inflate their safety scores with sudden, jerky moves no passenger would accept. This paper redraws the set of allowed controls at every step so they stay inside comfort limits, which tighten as speed rises. On Waymo driving data, comfort violations stayed under 1% and the policy still navigated better than the baselines.

Authors: Anna Rothenhäusler, Daniel Jost, Raghu Rajan, Faris Janjos, Oliver Scheel, Andreas Look, Joschka Boedecker

Paper: https://arxiv.org/abs/2609.13011v1

Research on steering language models mostly deals with yes-or-no concepts such as refusal. MAxBench tests concepts with many categories, such as animals or countries. Across 10 methods, 6 concepts and 4 models, affine subspaces worked best, largely because of better offsets. No method beat plain prompting.

Authors: Divya Appapogu, Freya Behrens, Yonatan Belinkov, Aaron Mueller

Paper: https://arxiv.org/abs/2609.13072v1

Voice agents are mostly tested in one-on-one conversation. MP-Bench tests them in groups. Across 12 agents, real-time systems scored 22% or less on comprehension and did no better than chance at knowing when to speak.

Authors: Yi-Jen Shih, Shih-Yun Shan Kuan, Guan-Ting Lin, Kai-Wei Chang, Siddhant Arora, Shu-wen Yang, Abdelrahman Mohamed, Shinji Watanabe, Hung-yi Lee, David Harwath

Paper: https://arxiv.org/abs/2609.13076v1

Can AI research agents handle a messy, open-ended industry problem? This study put them on telecom ticket retrieval. With little supervision they reached 90% of the best human result in 10 weeks instead of 10 months, at up to $200 a run. They were good at tuning but short on intuition, and the authors recommend pairing them with human researchers.

Authors: Junghyun Min, Huseyin Uzunalioglu, Mohamed Trabelsi

Paper: https://arxiv.org/abs/2609.13073v1

Many frontier models now use cheaper, subquadratic attention. This paper argues inference hardware should be split around that. SQD runs the parts of decoding that grow with context on different chips from the parts that don't. On an 8xB200 test setup it raised tokens per joule by 31% to 56% on GLM 5.2, Nemotron 3 Ultra and Gemma 4 31B. A modeled Rubin plus LPX system showed up to 3.6 times the throughput.

Authors: Arya Tschand, Yaosheng Fu, Vikram Sharma Mailthody, Nicolai Oswald, Po-An Tsai, Ritchie Zhao, Oreste Villa, Vijay Janapa Reddi, Karu Sankaralingam

Paper: https://arxiv.org/abs/2609.13134v1

Physics faculty and graduate students re-graded frontier model answers on six physics benchmarks. Most answers marked wrong turned out to reflect bad reference solutions, grading errors or unclear questions. After fixes, GPT-5.6-Sol rose from 47.3% to 78.7% on HLE-Physics and from 61.0% to 87.2% on CMT-Benchmark. The authors say these benchmarks understate what the models can do.

Authors: Ali Ansari, Haoran Sun, Andy Zeyi Liu, Mark Jabbour, Yongshan Ding, et al., Lucas Baker, Arman Cohan, John Sous (51 authors)

Paper: https://arxiv.org/abs/2609.13009v1