AI: post transformers

These sources explore advanced techniques for accelerating **Large Language Model (LLM) inference** through **speculative decoding**, a process where smaller "draft" models predict tokens for a larger "target" model to verify in parallel. A primary focus is **Multi-Draft Speculative Decoding (MDSD)**, which uses multiple draft sequences to increase the probability of acceptance and reduce latency. Researchers have introduced **SpecHub** to simplify complex optimization problems into manageable linear programming, while others utilize **optimal transport theory** and **q-convexity** to reach theoretical efficiency upper bounds. Additionally, the **Hierarchical Speculative Decoding (HSD)** framework stacks multiple models into a tiered structure, allowing each level to verify the one below it. Collectively, these papers provide **mathematical proofs**, **sampling algorithms**, and **hierarchical strategies** designed to maximize token acceptance rates and minimize computational overhead.Sources:1)January 22 2025Towards Optimal Multi-draft Speculative DecodingZhengmian Hu, Tong Zheng, Vignesh Viswanathan, Ziyi Chen, Ryan A. Rossi, Yihan Wu, Dinesh Manocha, Heng Huang.2)2024SpecHub: Provable Acceleration to Multi-Draft Speculative DecodingLehigh University, Samsung Research America, University of MarylandRyan Sun, Tianyi Zhou, Xun Chen, Lichao Sunhttps://aclanthology.org/2024.emnlp-main.1148.pdf.3)2024MULTI-DRAFT SPECULATIVE SAMPLING: CANONICAL DECOMPOSITION AND THEORETICAL LIMITSQualcomm AI Research, University of TorontoAshish Khisti, M.Reza Ebrahimi, Hassan Dbouk, Arash Behboodi, Roland Memisevic, Christos Louizoshttps://arxiv.org/pdf/2410.18234.4)2025HISPEC: HIERARCHICAL SPECULATIVE DECODING FOR LLMSThe University of Texas at AustinAvinash Kumar, Sujay Sanghavi, Poulami Dashttps://arxiv.org/pdf/2510.01336.5)2025Fast Inference via Hierarchical Speculative DecodingHarvard University, Google Research, Tel Aviv University, Google DeepMindClara Mohri, Haim Kaplan, Tal Schuster, Yishay Mansour, Amir Globersonhttps://arxiv.org/pdf/2510.19705.

This research collaboration between King’s College London, Google DeepMind on a research paper published on February 19, 2026 introduces a novel framework for evaluating the **collective behavior** of large language model (LLM) agents within complex **social dilemmas**. By prompting models to generate high-level **algorithmic strategies** rather than individual actions, the authors successfully simulated interactions among hundreds of agents to observe emergent societal outcomes. The study reveals a concerning trend where newer, more capable reasoning models often prioritize **individual gain**, leading to a "race to the bottom" that diminishes total **social welfare**. Through **cultural evolution** simulations, the researchers found that **exploitative strategies** frequently dominate populations, especially as group sizes increase and the relative benefits of cooperation drop. To address these risks, the authors released an **evaluation suite** for developers to assess and mitigate anti-social tendencies in autonomous agents before deployment. Ultimately, the findings highlight a critical tension: while advanced reasoning can achieve optimal cooperation, it also empowers models to become more effective at **exploitation**.Source:February 19, 2026EVALUATING COLLECTIVE BEHAVIOUR OF HUNDREDS OF LLM AGENTSKing’s College London, Google DeepMindRichard Willis, Jianing Zhao, Yali Du, Joel Z. Leibohttps://arxiv.org/pdf/2602.16662

The February 12.2026 research from the University of Virginia and Google introduces the deep-thinking ratio (DTR), a novel metric designed to measure the true reasoning effort of large language models by analyzing **internal token stabilization**. While traditional metrics like **token count** often fail to predict accuracy due to "overthinking," DTR tracks how many layers a model requires before its internal predictions converge. Findings across several benchmarks indicate that **higher DTR scores** correlate strongly with correct answers, whereas mere output length often shows a negative correlation with performance. Using this insight, the authors developed **Think@n**, a test-time scaling strategy that identifies and prioritizes high-quality reasoning traces early in the generation process. This method allows models to match or exceed the accuracy of **standard self-consistency** while cutting computational costs by roughly half. Ultimately, the study suggests that **reasoning quality** is better reflected by a model's internal depth-wise processing than by the superficial length of its responses.Source:February 12 2026Think Deep, Not Just Long: Measuring LLM Reasoning Effort via Deep-Thinking TokensUniversity of Virginia, GoogleWei-Lin Chen, Liqian Peng, Tian Tan, Chao Zhao, Blake JianHang Chen, Ziqian Lin, Alec Go, Yu Menghttps://arxiv.org/pdf/2602.13517

These sources detail advanced **reinforcement learning frameworks** designed to improve how **quadruped robots** navigate difficult, real-world environments. The first source introduces a **single-stage teacher-student method** that utilizes **skeleton information** and a system-response model to achieve more natural, stable movement. The second source proposes **ZSL-RPPO**, a zero-shot learning architecture that eliminates the need for imitation by training **recurrent neural networks** directly in partially observable settings. Both research papers prioritize bridging the **simulation-to-reality gap**, ensuring robots can handle unpredictable terrain like stairs, oily surfaces, and grass. By employing **domain randomization** and specialized encoders, these frameworks enhance the **robustness and adaptability** of robotic locomotion without requiring extensive manual tuning. Together, they represent a shift toward more **efficient training paradigms** that produce versatile and resilient autonomous behaviors.Sources:1)October 22 2025Skeleton Information-Driven Reinforcement Learning Framework for Robust and Natural Motion of Quadruped RobotsGuangdong University of Technology, University of MacauHuiyang Cao, Hongfa Lei, Yangjun Liu, Zheng Chen, Shuai Shi, Bingquan Li, Weichao Xu, Zhi-Xin Yanghttps://doi.org/10.3390/sym171117872)March 2024ZSL-RPPO: Zero-Shot Learning for Quadrupedal Locomotion in Challenging Terrains using Recurrent Proximal Policy OptimizationHuawei Technologies, Huawei Munich Research Center, University College London, Huawei Noah's Ark Lab, East China Normal UniversityYao Zhao, Tao Wu, Yijie Zhu, Xiang Lu, Jun Wang, Haitham Bou-Ammar, Xinyu Zhang, Peng Duhttps://arxiv.org/pdf/2403.019283)May 2025End-to-End Multi-Task Policy Learning from NMPC for Quadruped LocomotionBonn-Rhein-Sieg University of Applied Sciences, University of Bonn, Fraunhofer Institute for Intelligent Analysis and Information SystemsAnudeep Sajja, Shahram Khorshidi, Sebastian Houben, Maren Bennewitzhttps://arxiv.org/pdf/2505.08574

MEDUSA is a novel framework introduced on June 24 2024 designed to accelerate Large Language Model (LLM) inference by overcoming the delays caused by sequential token generation. Instead of relying on a separate draft model like traditional speculative decoding, it incorporates multiple decoding heads that predict several subsequent tokens simultaneously. These predictions are organized into a tree-based attention mechanism, allowing the model to verify multiple potential continuations in a single parallel step. The system offers two fine-tuning tiers: MEDUSA-1, which keeps the backbone model frozen for easy integration, and MEDUSA-2, which trains the heads and backbone together for superior speed. Additionally, a typical acceptance scheme and self-distillation pipeline ensure high-quality, diverse outputs even when original training data is unavailable. Experimental results demonstrate that this approach can increase generation speeds by 2.2 to 2.8 times without compromising the accuracy or quality of the language model.Source:June 14, 024MEDUSA: Simple LLM Inference Acceleration Framework with Multiple Decoding HeadsPrinceton University, Together AI, University of Illinois Urbana-Champaign, Carnegie Mellon University, University of ConnecticutTianle Cai, Yuhong Li, Zhengyang Geng, Hongwu Peng, Jason D. Lee, Deming Chen, Tri Daohttps://arxiv.org/pdf/2401.10774

On a paper published January 21, 2026 researchers from MIT and NVIDIA explain how they have have developed a new system called Taming the Long Tail (TLT) to solve computational inefficiencies in training reasoning-heavy large language models. During standard training, many processors sit idle while waiting for the longest text sequences to finish generating, creating a massive resource bottleneck. The TLT system captures this wasted time by automatically training a secondary, lightweight drafter model on the fly to predict the primary model's outputs. This adaptive speculative decoding approach allows the larger model to verify multiple tokens at once, effectively doubling training speeds without losing any mathematical accuracy. By optimizing hardware usage, this method significantly reduces the energy and financial costs associated with developing advanced AI capable of complex logic and self-reflection.Source:January 21, 2026Taming the Long-Tail: Efficient Reasoning RL Training with Adaptive DrafterMIT, ETH Zurich, NVIDIA, UMass AmherstQinghao Hu, Shang Yang, Junxian Guo, Xiaozhe Yao, Yujun Lin, Yuxian Gu, Han Cai, Chuang Gan, Ana Klimovic, Song Hanhttps://arxiv.org/pdf/2511.16665https://news.mit.edu/2026/new-method-could-increase-llm-training-efficiency-0226

The September 26 2025 research paper introduces FastGRPO, a high-efficiency framework designed to accelerate the training of large language models using Group Relative Policy Optimization. The authors identify that the generation phase is the primary bottleneck in reinforcement learning, accounting for over 90% of total training time. To solve this, they implement concurrency-aware speculative decoding, which dynamically adjusts drafting and verification strategies based on real-time batch sizes. Additionally, an online draft learning mechanism is introduced to keep the smaller assistant model aligned with the evolving target model. Experimental results show that this approach achieves end-to-end speedups of up to 2.72x without compromising reasoning performance. Ultimately, the framework optimizes hardware utilization by balancing memory bandwidth and computational overhead during high-concurrency training.September 26, 2026Yizhou Zhang Ning Lv, Teng Wang, Jisheng DangLanzhou University, The University of Hong Kong, National University of Singaporehttps://arxiv.org/pdf/2509.21792

Researchers introduced on May 2024 self-speculative decoding, a novel "plug-and-play" inference scheme designed to accelerate Large Language Models (LLMs) without requiring auxiliary models or extra memory. This method utilizes a two-stage process where a faster, lower-quality draft is generated by selectively skipping intermediate layers of the original model. These draft tokens are then validated in a single forward pass by the full LLM, ensuring the final output remains identical to standard autoregressive decoding. To optimize performance, the system employs Bayesian optimization to identify the best layers to skip and an adaptive draft-exiting mechanism to stop generation when confidence is low. Benchmarks on models like LLaMA-2 show significant speedups of up to 1.99× across text and code generation tasks. Ultimately, this approach offers a cost-effective and lossless solution for reducing latency in large-scale AI applications.Source:May 20, 2024Draft & Verify: Lossless Large Language Model Acceleration via Self-Speculative Decoding.Zhejiang University, University of California, Irvine.Jun Zhang, Jue Wang, Huan Li, Lidan Shou, Ke Chen, Gang Chen, Sharad Mehrotra.https://arxiv.org/pdf/2309.08168.

The Deepmind February 3, 2023 paper "Accelerating Large Language Model Decoding with Speculative Sampling introduced speculative sampling, a novel algorithm designed to increase the speed of Large Language Model (LLM) decoding without altering the final output. The researchers utilize a smaller, faster draft model to predict multiple potential tokens, which are then verified in parallel by a larger, more powerful target model. By employing a unique rejection sampling scheme, the system ensures that the generated text remains mathematically identical to the distribution of the original large model. When tested with the 70 billion parameter Chinchilla model, this technique achieved a 2 to 2.5 times speedup in processing. The method is particularly effective because it overcomes the memory bandwidth bottlenecks typical of standard autoregressive generation. Ultimately, it provides a practical way to reduce latency in large-scale AI applications without sacrificing sample quality.Source:February 3 2023Accelerating Large Language Model Decoding with Speculative SamplingDeepMindCharlie Chen, Sebastian Borgeaud, Geoffrey Irving, Jean-Baptiste Lespiau, Laurent Sifre, John Jumperhttps://arxiv.org/pdf/2302.01318

Researchers from METR introduce a novel framework for evaluating AI progress by measuring a model's time horizon, defined as the length of a task a human can complete that an AI can perform with 50% reliability. Traditional benchmarks often fail because they saturate quickly or focus on static knowledge, whereas this approach uses economically valuable tasks in fields like software engineering and cybersecurity. By comparing AI performance against over 2,500 hours of human baselines, the study found that the effective time horizon for frontier models has doubled approximately every 212 days since 2019. This consistent exponential growth suggests that AI agents may be capable of automating complex, month-long human projects by the end of this decade. While newer models like o1 show significant improvements in reasoning and error correction, they still struggle with "messy" environments that lack clear feedback loops. Ultimately, this psychometric-inspired methodology provides a unified metric to track the evolution of autonomous agents and forecast potential catastrophic risks as systems become increasingly powerful.Source:March 2025Measuring AI Ability to Complete Long TasksModel Evaluation & Threat Research (METR)Thomas Kwa, Ben West, Joel Becker, Amy Deng, Katharyn Garcia, Max Hasin, Sami Jawhar, Megan Kinniment, Nate Rush, Sydney Von Arx, Ryan Bloom, Thomas Broadley, Haoxing Du, Brian Goodrich, Nikola Jurkovic, Luke Harold Miles, Seraphina Nix, Tao Lin, Chris Painter, Neev Parikh, David Rein, Lucas Jun Koba Sato, Hjalmar Wijk, Daniel M. Ziegler, Elizabeth Barnes, Lawrence Chanhttps://arxiv.org/pdf/2503.14499

The provided sources explore advanced techniques for optimizing large language model (LLM) inference, specifically by addressing the memory bottlenecks of the Key-Value (KV) cache. KVQuant introduces a high-precision quantization framework that utilizes per-channel scaling, non-uniform datatypes, and sparse outlier handling to compress activations to sub-4-bit precision with minimal accuracy loss. Similarly, the KIVI algorithm proposes a tuning-free 2-bit quantization strategy that differentiates between key and value cache distributions to increase throughput. Shifting from quantization to architectural pruning, DuoAttention identifies specific Retrieval Heads that require full context while reducing Streaming Heads to constant memory usage by focusing only on recent tokens and attention sinks. Together, these methods enable LLMs to process million-level context lengths on standard hardware by drastically reducing the architectural and computational footprint of stored activations.Sources:1) 2024KVQuant: Towards 10 Million Context Length LLM Inference with KV Cache QuantizationUniversity of California, Berkeley, ICSI, LBNLColeman Hooper, Sehoon Kim, Hiva Mohammadzadeh, Michael W. Mahoney, Yakun Sophia Shao, Kurt Keutzer, Amir Gholamihttps://arxiv.org/pdf/2401.180792) 2024KIVI: A Tuning-Free Asymmetric 2bit Quantization for KV CacheRice University, Texas A&M University, Stevens Institute of Technology, Carnegie Mellon UniversityZirui Liu, Jiayi Yuan, Hongye Jin, Shaochen (Henry) Zhong, Zhaozhuo Xu, Vladimir Braverman, Beidi Chen, Xia Huhttps://arxiv.org/pdf/2402.027503) 2024QAQ: Quality Adaptive Quantization for LLM KV CacheNanjing UniversityShichen Dong, Wen Cheng, Jiayu Qin, Wei Wanghttps://arxiv.org/pdf/2403.046434) May 8, 2024KV Cache is 1 Bit Per Channel: Efficient Large Language Model Inference with Coupled QuantizationRice University, Stevens Institute of Technology, ThirdAI Corp.Tianyi Zhang, Jonah Yi, Zhaozhuo Xu, Anshumali Shrivastavahttps://arxiv.org/pdf/2405.039175) 2024DUOATTENTION: EFFICIENT LONG-CONTEXT LLM INFERENCE WITH RETRIEVAL AND STREAMING HEADSMIT, Tsinghua University, SJTU, University of Edinburgh, NVIDIAGuangxuan Xiao, Jiaming Tang, Jingwei Zuo, Junxian Guo, Shang Yang, Haotian Tang, Yao Fu, Song Hanhttps://arxiv.org/pdf/2410.108196) 2025MILLION: MasterIng Long-Context LLM Inference Via Outlier-Immunized KV Product QuaNtizationShanghai Jiao Tong University, Shanghai Qi Zhi Institute, Huawei Technologies Co., Ltd, China University of Petroleum-BeijingZongwu Wang, Peng Xu, Fangxin Liu, Yiwei Hu, Qingxiao Sun, Gezi Li, Cheng Li, Xuan Wang, Li Jiang, Haibing Guanhttps://arxiv.org/pdf/2504.03661

The January 29, 2026 research collaboration between Stanford University, SambaNova Systems, Inc and UC Berkeley introduce ACE (Agentic Context Engineering), a novel framework designed to improve how large language models learn and adapt through context rather than weight updates. Unlike traditional methods that suffer from brevity bias or context collapse by summarizing information too aggressively, ACE treats contexts as evolving playbooks that preserve and organize detailed domain insights. It utilizes a modular, human-like learning workflow consisting of a Generator, Reflector, and Curator to produce structured, incremental updates. This grow-and-refine approach manages long contexts efficiently by appending new "deltas" and pruning redundancies through semantic embeddings. Research results demonstrate that ACE significantly boosts performance in autonomous agents and complex financial reasoning while drastically reducing processing latency and operational costs. Ultimately, the framework offers a scalable solution for building self-improving AI systems that retain high accuracy across long-horizon tasks.Source:January 29, 2026Agentic Context Engineering: Evolving Contexts for Self-Improving Language ModelsStanford University, SambaNova Systems, Inc., UC BerkeleyQizheng Zhang, Changran Hu, Shubhangi Upasani, Boyuan Ma, Fenglu Hong, Vamsidhar Kamanuru, Jay Rainton, Chen Wu, Mengmeng Ji, Hanchen Li, Urmish Thakker, James Zou, Kunle Olukotunhttps://arxiv.org/pdf/2510.04618https://github.com/ace-agent/ace

The February 3, 2026 research paper in collaboration between the National University of Singapore, USTC, University of Toronto and the Sea AI Lab introduces Cortex, a specialized caching system designed to address high latency and financial costs in LLM agent applications. Unlike standard models, agents frequently perform repetitive external data retrievals that lead to significant delays and expensive API fees. Cortex optimizes this process by using a semantic cache to store and reuse previous search results and tool calls. A key innovation is the semantic judge, which ensures cached data is still accurate and relevant before serving it to the user. To maximize hardware efficiency, the system co-locates the agent and the judge on a single GPU using priority-aware scheduling. Evaluations show that this approach can improve throughput by up to 3.6× while reducing operational expenses by over 90%.Source:February 3, 2026Cortex: Achieving Low-Latency, Cost-Efficient Remote Data Access For LLM via Semantic-Aware Knowledge CachingNational University of Singapore, USTC, University of Toronto, Sea AI LabChaoyi Ruan, Chao Bi, Kaiwen Zheng, Ziji Shi, Xinyi Wan, Jialin Lihttps://arxiv.org/pdf/2509.17360

The January 26, 2026 Stanford research paper introduces Agentic Plan Caching (APC), a novel framework designed to reduce the high operational costs of Large Language Model (LLM) agents. Traditional caching methods often fail because agent workflows are highly dynamic and dependent on external environments, making simple input-output storage ineffective. The APC framework solves this by extracting generalized plan templates from successful task executions, which are then indexed by high-level intent keywords. When a similar task is encountered, a small, cost-effective planner LM adapts these cached templates to the new context, significantly reducing the need for expensive, high-reasoning models. Experiments demonstrate that this approach can cut financial and computational costs by over 75% while maintaining high accuracy across complex benchmarks. Ultimately, this system offers a scalable way to deploy sophisticated AI agents by minimizing redundant reasoning through intelligent plan reuse.Source:January 26, 2026Agentic Plan Caching: Test-Time Memory for Fast and Cost-Efficient LLM AgentsStanford UniversityQizheng Zhang, Michael Wornow, Gerry Wan, Kunle Olukotunhttps://arxiv.org/pdf/2506.14852

The global AI agents market is experiencing explosive growth, with projections suggesting it could reach nearly $183 billion by 2033. This surge is fueled by a transition from simple bots to autonomous systems capable of executing complex, multi-step workflows across sectors like finance, healthcare, and retail. North America currently leads the industry, supported by heavy investments from tech giants like Microsoft, Google, and NVIDIA. These companies are prioritizing multi-agent orchestration and low-code platforms to help businesses build specialized, domain-specific tools. While industry forecasts anticipate over one billion agents in operation by 2028, some analysts suggest these figures may be inflated by the ease of accidental creation within existing software. Ultimately, the future of the market depends on advanced natural language processing, secure cloud infrastructure, and the ability of various AI entities to collaborate effectively.Sources:1)November 6 202526 AI Agent Statistics (Adoption Trends and Business Impact)DatagridDatagrid Teamhttps://www.datagrid.com/blog/26-ai-agent-statistics2)2026AI Agents Market Size And Share | Industry Report, 2033Grand View Researchhttps://www.grandviewresearch.com/industry-analysis/ai-agents-market-report3)April 2025AI Agents Market by Agent Role, Offering, Agent System - Global Forecast to 2030MarketsandMarketshttps://www.marketsandmarkets.com/Market-Reports/ai-agents-market-15761548.html4)April 23 2025AI Agents Market worth $52.62 billion by 2030MarketsandMarketsRohan Salgarkarhttps://www.marketsandmarkets.com/Market-Reports/ai-agents-market-15761548.html5)September 5 2025Gartner Predicts 40% of Enterprise Apps Will Feature Task-Specific AI Agents by 2026, Up from Less Than 5% in 2025GartnerSonika Choubeyhttps://www.gartner.com/en/newsroom/press-releases/gartner-predicts-40-percent-of-enterprise-apps-will-feature-task-specific-ai-agents6)June 25 2025Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027GartnerEmma Keenhttps://www.gartner.com/en/newsroom/press-releases/gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled7)May 30 2025Getting to one billion agentsPerspectives on Power PlatformJukka Niiranenhttps://www.perspectives.plus/p/one-billion-apps8)May 20 2025Microsoft expects 1.3 billion AI agents to be in operation by 2028 – here’s how it plans to get them working togetherIT ProBobby Hellardhttps://www.itpro.com/news/microsoft-expects-1-3-billion-ai-agents9)2026Top Strategic Technology Trends for 2026GartnerGene Alvarez, Tori Paulmanhttps://www.gartner.com/en/information-technology/insights/top-technology-trends10)October 28 2025What 1.3 billion AI Agents by 2028 Means for Business LeadersLanternhttps://www.lantern.com/blog/what-1-3-billion-ai-agents-by-2028-means-for-business-leaders11)October 2025AI Agents: Technologies, Applications and Global MarketsBCC ResearchAustin Samuelhttps://www.bccresearch.com/market-research/artificial-intelligence-technology/ai-agent-market.html

This FAST26 February 24, 2026 paper introduces CacheSlide, an innovative system designed to accelerate Large Language Model (LLM) serving by improving KV cache reuse. The researchers identify that complex agent-based workflows often cause **positional misalignment**, where cached data becomes unusable when text segments shift in absolute position. To resolve this, they propose **Relative-Position-Dependent Caching (RPDC)**, which focuses on preserving the order of segments rather than their exact locations. The architecture utilizes **Chunked Contextual Position Encoding (CCPE)** to minimize drift and **Weighted Correction Attention** to efficiently restore cross-attention between static and dynamic segments. Additionally, the **SLIDE manager** optimizes system performance by decoupling data loading from computation and implementing **dirty-aware eviction** to reduce storage bottlenecks. Experimental results demonstrate that this approach significantly **reduces latency** and **increases throughput** across various agentic benchmarks with minimal impact on accuracy.February 24, 2026CacheSlide: Unlocking Cross Position-Aware KV Cache Reuse for Accelerating LLM Serving Yang Liu and Yunfei Gu, Shanghai Jiao Tong University; Liqiang Zhang, Jinan Inspur Data Technology Co., Ltd; Chentao Wu, Guangtao Xue, Jie Li, and Minyi Guo, Shanghai Jiao Tong University; Junhao Hu, Peking University; Jie Meng, Huawei Cloudhttps://www.usenix.org/system/files/fast26-liu-yang.pdf

This February 2026 research paper introduces Bidaw, a novel system designed to optimize the performance of interactive Large Language Model (LLM) serving. The authors address inefficiencies in existing **key–value (KV) caching** methods, where the separation of computational engines and storage layers leads to high latency and redundant data processing. **Bidaw** implements **bidirectional awareness**, allowing the compute engine to schedule requests based on storage speeds while the storage system uses model output lengths to predict future data needs. Additionally, the system utilizes **storage-efficient tensor caching** to reduce memory footprints without sacrificing accuracy. Experimental results demonstrate that this approach significantly lowers response times and increases throughput compared to current state-of-the-art solutions. Ultimately, **Bidaw** bridges the gap between theoretical caching limits and practical local deployments for multi-round human-AI conversations.Source:February 24–26, 2026Bidaw: Enhancing Key-Value Caching for Interactive LLM Serving via Bidirectional Computation–Storage AwarenessTsinghua University, China University of Geosciences Beijing, China Telecom Omni-channel Operation CenterShipeng Hu, Guangyan Zhang, Yuqi Zhou, Yaya Wei, Ziyan Zhong, Jike Chenhttps://www.usenix.org/conference/fast26/presentation/hu-shipeng

This December 2025 paper introduces SGI-Bench, a comprehensive framework designed to evaluate the capabilities of autonomous scientific agents across diverse research workflows. The benchmark spans multiple disciplines, including **chemistry, materials science, and astronomy**, by challenging models with tasks like **experimental design, numerical modeling, and data interpretation**. Through a series of structured modules, it explores how artificial intelligence can manage **dry experiments** involving simulations and **wet experiments** focused on physical laboratory processes. Technical examples demonstrate the rigorous use of **mathematical derivations and multi-modal analysis** to solve complex problems, such as calculating gravitational wave parameters or predicting molecular properties. Ultimately, the text highlights a shift toward **agentic science**, where AI assistants assist in accelerating discovery through systematic reasoning and automated tool use.Source:December 2025Probing Scientific General Intelligence of LLMs with Scientist-Aligned WorkflowsShanghai Artificial Intelligence LaboratoryWanghan Xu, Yuhao Zhou, Yifan Zhou, Qinglong Cao, Shuo Li, Jia Bu, et al.https://arxiv.org/pdf/2512.16969

Leading researchers propose a shift away from agentic AI, which autonomously pursues goals and poses catastrophic risks such as deception and loss of human control. To mitigate these dangers, they introduce the concept of Scientist AI, a non-agentic framework designed for **understanding the world** rather than acting within it. This system utilizes a **probabilistic world model** to generate causal theories and an inference machine to answer queries based on those hypotheses. By adopting a Bayesian approach, the model explicitly accounts for uncertainty, preventing the overconfident or manipulative behaviors common in current reward-driven systems. Ultimately, this safe-by-design alternative aims to accelerate scientific progress while serving as a trustworthy guardrail against more volatile autonomous agents.Source:February 24 2025Superintelligent Agents Pose Catastrophic Risks: Can Scientist AI Offer a Safer Path?Authors: Yoshua Bengio (Mila — Quebec AI Institute; Université de Montréal), Michael Cohen (University of California, Berkeley), Damiano Fornasiere (Mila — Quebec AI Institute), Joumana Ghosn (Mila — Quebec AI Institute), Pietro Greiner (Mila — Quebec AI Institute), Matt MacDermott (Imperial College London; Mila — Quebec AI Institute), Sören Mindermann (Mila — Quebec AI Institute), Adam Oberman (Mila — Quebec AI Institute; McGill University), Jesse Richardson (Mila — Quebec AI Institute), Oliver Richardson (Mila — Quebec AI Institute; Université de Montréal), Marc-Antoine Rondeau (Mila — Quebec AI Institute), Pierre-Luc St-Charles (Mila — Quebec AI Institute), David Williams-King (Mila — Quebec AI Institute)https://arxiv.org/pdf/2502.15657

Bloom is an open-source agentic framework designed to automate the development and execution of **behavioral evaluations** for frontier AI models. Unlike traditional static benchmarks, it utilizes a **four-stage pipeline**—Understanding, Ideation, Rollout, and Judgment—to generate diverse, targeted scenarios that quantify specific traits like **sycophancy, sabotage, and bias**. The tool is highly configurable, allowing researchers to adjust **seed configurations**, reasoning effort, and interaction lengths to produce reproducible and statistically significant metrics. Validation experiments show that **Bloom** effectively distinguishes between baseline models and those intentionally designed to be misaligned, while its automated scoring correlates strongly with **human judgment**. By providing a scalable alternative to high-effort manual auditing, it enables the rapid measurement of **alignment-relevant behaviors** across multiple model families. Ultimately, **Bloom** serves as a specialized instrument for precise behavioral measurement, complementing broader exploratory auditing tools in the **AI safety** landscape.Source:December 19, 2025Bloom: an open source tool for automated behavioral evaluationsIsha Gupta, Kai Fronsdal, Abhay Sheshadri, Jonathan Michala, Jacqueline Tay, Rowan Wang, Samuel R. Bowman, Sara Pricehttps://alignment.anthropic.com/2025/bloom-auto-evals/github.com/safety-research/bloom.