# value Published articles for value. This is one page of public article previews, not the complete archive. Follow Next page to continue. Summaries are not the original full articles. ## Generalized Priority-Aware Shapley Value DevFeed: [Generalized Priority-Aware Shapley Value]() Original publisher: [Read original article]() Author: Kiljae Lee, Ziqi Liu, Weijing Tang, Yuan Zhang Published: 2026-11-09T05:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [Machine learning](), [Large Language Model]() Tags: [apply](), [arena](), [boundary](), [characterization](), [chatbot](), [comparisons](), [covers](), [cs-ai](), [cs-lg](), [data](), [develop](), [examples](), [graph](), [graphs](), [learning](), [llm](), [machine-learning](), [models](), [order](), [priority](), [random](), [value]() ### AI overview The generalized priority-aware Shapley value (GPASV) handles arbitrary directed, weighted priority graphs by penalizing order violations instead of forbidding them, addressing cyclic or nonbinary priorities that existing methods cannot represent. The authors give an axiomatic characterization, computational methods, and a priority-sweeping diagnostic, and show that classical models arise as boundary cases. In an application to LLM ensemble valuation on the cyclic Chatbot Arena preference graph, different balances between graph priority and individual soft priority produce substantively different valuations of the same data; priority-aware valuation therefore requires choosing that balance. This overview is based on the supplied abstract. ### Source excerpt arXiv:2605.15018v2 Announce Type: replace Abstract: Shapley value and its priority-aware extensions are widely used for valuation in machine learning, but existing methods require pairwise priority to be binary and acyclic, a restriction spectacularly violated in real-data examples such as aggregated human preferences and multi-criterion comparisons. We introduce the generalized priority-aware Shapley value (GPASV), a random order value defined on arbitrary directed weighted priority graphs, in which pairwise edges penalize rather than forbid order violations. GPASV covers a range of classical models as boundary cases. We establish GPASV through an axiomatic characterization, develop the associated computational methods, and introduce a priority sweeping diagnostic extending PASV's. We apply GPASV to LLM ensemble valuation on the cyclic Chatbot Arena preference graph, illustrating that priority-aware valuation is not a one-button operation: different balances of pairwise graph priority versus individual soft priority produce substantively different valuations of the same data. ## The AI boom's missing balance-sheet question: What has actually been verified? DevFeed: [The AI boom's missing balance-sheet question: What has actually been verified?]() Original publisher: [Read original article]() Author: Neil P. Osnato Published: 2026-11-06T15:00:36Z Content type: opinion Language: en Sources: [SiliconANGLE]() Topics: [AI Infrastructure](), [Finance](), [decision-making](), [clean energy data centers](), [Artificial Intelligence]() Tags: [ai](), [ai-boom](), [ai-infrastructure](), [artificial-intelligence](), [balance](), [balance-sheet](), [capacity](), [exchange](), [faster](), [generative-ai](), [i](), [infrastructure](), [intelligence](), [me](), [neil-p-osnato](), [new-york](), [news](), [one](), [persistent-analytics-group](), [question](), [siliconangle](), [stock](), [the-latest](), [top-story-4](), [top-story-6](), [trading](), [transaction](), [value]() ### AI overview AI infrastructure commitments can be valid and still rest on assumptions that may not hold over the investment's life. The article distinguishes contracted compute demand from productive utilization and says announcements, financing, power agreements and GPU orders do not by themselves prove capacity will be deliverable and used as planned. It proposes tracking a sequence from requested and evidenced load through deliverable, energized and utilized capacity to sustained demand, with supporting records for each claim. Guarantees and other contractual protections can mitigate losses but do not ensure the asset remains useful or remove counterparty and dependency risks. The recommended control is ongoing revalidation: document the representations, assumptions, evidence, dependencies, downside exposure and changes that should trigger a reassessment before further commitments. The author frames this as investment discipline, not a prediction that AI infrastructure will fail. ### Source excerpt Artificial intelligence infrastructure commitments are growing faster than the evidence connecting contracted capacity to sustained economic value. I spent 13 years on the floor of the New York Stock Exchange. One lesson stayed with me long after the trading jackets disappeared: A transaction could be executed, documented and perfectly valid while the assumptions that made [...] The post The AI boom's missing balance-sheet question: What has actually been verified? appeared first on SiliconANGLE. ## The Struggle to Secure Hardware Samples for Honest Tech Reviews DevFeed: [The Struggle to Secure Hardware Samples for Honest Tech Reviews]() Original publisher: [Read original article]() Author: Matthew - Technology News Australia Published: 2026-11-06T08:26:45Z Content type: opinion Language: en Sources: [HackerNoon]() Topics: [Linux Hardware Reviews](), [upgrade]() Tags: [amd](), [benchmark](), [costs](), [cpu](), [editorial](), [equipment](), [everyday](), [gaming](), [gaming-hardware](), [hardware](), [hardware-manufacturers](), [independent](), [it](), [nvidia](), [pc-hardware-upgrades](), [performance](), [perspectives](), [reviews](), [tech-business-news](), [tech-review](), [upgrade](), [value]() ### AI overview The author says unanswered sample requests to NVIDIA and AMD have kept Tech Business News from testing hardware firsthand. Their IT experience informs their approach but cannot replace access to the products; buying equipment also adds costs that constrain coverage. They want reviews to connect benchmark results with practical use: whether upgrades make everyday applications faster or games smoother, including stuttering and performance drops, and which users might notice the gains. They would disclose testing settings and system hardware, explain the evidence behind personal judgments, and note when measured differences are hard to perceive. The author suggests loan samples with clear testing and return arrangements, while retaining editorial freedom, and says even a reply that samples are unavailable would help clarify the situation. The excerpt does not establish why vendors have not replied or how established reviewers obtain their samples. ### Source excerpt Smaller tech reviewers struggle to access hardware samples, limiting independent perspectives on everyday performance, gaming and upgrade value. ## The Token Before the Value Is the Key DevFeed: [The Token Before the Value Is the Key]() Original publisher: [Read original article]() Author: Ke Cheng, Xin Xu, Zhichao Chen, Yixiao Chen, Lei Xin, Jianbo Zhao, Fanhu Zeng, Yue Liu, Jun Zhang, Jie Jiang Published: 2026-11-06T05:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [Machine learning](), [Language models]() Tags: [abstract](), [adapting](), [analysis](), [arxiv](), [boundary](), [bridge](), [circuit](), [code](), [competition](), [connect](), [context](), [cs-lg](), [first](), [fixed](), [generation](), [global](), [hybrid](), [induction](), [key](), [token](), [value]() ### AI overview The study identifies the historical source key--the information carried from the token before a value--as a shared causal interface for retrieval in hybrid language models. Probes and restoration interventions link it to local preparation and global retrieval across controlled models, pretrained recurrent and sliding-window hybrids, and released language models. Shared local parameters can couple predecessor carrying with recurrent processing; separating those demands or biasing initialization promotes address formation. Adapting input representations enables category matching, but creates competition among semantically compatible records. Earlier context can resolve that competition through the source key, with a visibility boundary in a controlled single-read model and compositional relays across local layers. For sequential generalization, fixed relations allow memorization shortcuts, while generalization requires an output-to-query bridge between retrievals. Controlled tasks show this requirement across stacked stages, looped recurrence, and intermediate-token generation. ### Source excerpt arXiv:2609.15545v3 Announce Type: replace Abstract: Induction circuits retrieve a value by matching a query to information carried from the token before that value. How this address forms, incorporates context, and supports sequential retrieval remains less clear. We use hybrid language models as a tractable testbed and identify the historical source key as a common causal interface between local preparation and global retrieval. Probes and restoration interventions connect this interface across controlled models, pretrained recurrent and sliding-window hybrids, and released language models. We organize the analysis by address context radius and sequential retrieval depth. First, shared local parameters couple predecessor Carrying to recurrent state processing; separating these demands or biasing initialization promotes address formation. Second, adapting input representations extends an established literal circuit to category matching, making semantically compatible records active competitors. Earlier context resolves this competition through the source key, with a sharp visibility boundary in a controlled single-read model and compositional relays across local layers. Third, fixed relations support memorization shortcuts, whereas sequential generalization requires learning an output-to-query bridge between retrievals. Controlled state-tracking and compositional in-context tasks demonstrate this requirement across stacked stages, looped recurrence, and intermediate-token generation. Together, these results connect circuit formation, contextual disambiguation, and sequential use through a shared retrieval interface. Code is available at https://github.com/ckpassenger/bind-match-copy. ## MAC-v2: Long-horizon model-based value expansion from pixels DevFeed: [MAC-v2: Long-horizon model-based value expansion from pixels]() Original publisher: [Read original article]() Author: Kwanyoung Park, Sergey Levine Published: 2026-11-06T05:00:00Z Content type: article Language: en Sources: [arXiv: Artificial Intelligence](), [arXiv: Machine Learning]() Topics: [Reinforcement learning](), [World models](), [Representation learning](), [Machine learning](), [Training AI Models](), [Neural Network]() Tags: [benchmarks](), [control](), [cs-ai](), [cs-lg](), [horizon](), [offline-reinforcement-learning](), [prediction](), [reinforcement-learning](), [value](), [world-model]() ### AI overview MAC-v2 applies long-horizon model-based value expansion to pixel-based offline reinforcement learning using a reconstruction-free latent world model and action chunking. Prediction and control objectives let its representations focus on information relevant to dynamics and control, while chunked actions reduce the number of recursive predictions and help limit accumulated model error. The abstract reports strong performance against model-free and model-based offline RL baselines on visual OGBench, MMBench2, Robomimic, and LIBERO-10. Its analyses identify long-horizon value expansion, action chunking, and latent learning objectives as important to performance and reliable imagination. ### Source excerpt arXiv:2611.07245v1 Announce Type: cross Abstract: We study whether \emph{long-horizon model-based value expansion} can serve as an effective framework in pixel-based offline reinforcement learning (RL). Long-horizon model-based value expansion from pixels is substantially harder than state-based environments, as the agent must learn both the representation and its dynamics, and errors from either can accumulate recursively over imagined trajectories. We introduce MAC-v2, which enables long-horizon model-based expansion from pixels by training a reconstruction-free latent world model with action chunking. By learning representations from prediction and control objectives, MAC-v2 does not need to encode control-irrelevant visual details that can hinder the dynamics prediction. Action chunking further mitigates the prediction error by reducing the number of recursive model predictions required for value expansion. Across diverse visual benchmarks including visual OGBench, MMBench2, Robomimic, and LIBERO-10, MAC-v2 achieves strong performance against model-free and model-based offline RL baselines. Our analyses show that long-horizon model-based value expansion is critical for the performance of MAC-v2, and that both action chunking and latent learning objectives are essential for reliable long-horizon latent imagination. ## Value or Attention: Investigating Layer-wise Relevance Propagation for Graph Attention Networks DevFeed: [Value or Attention: Investigating Layer-wise Relevance Propagation for Graph Attention Networks]() Original publisher: [Read original article]() Author: Ken Schreiber, Ren\\'e P. Klausen, Sebastian Lapuschkin, Wojciech Samek, Jonas Naujoks Published: 2026-11-06T05:00:00Z Content type: article Language: en Sources: [arXiv: Artificial Intelligence](), [arXiv: Machine Learning]() Topics: [Artificial Intelligence](), [Ground truth / benchmark quality](), [Error Propagation]() Tags: [attention](), [classification](), [content](), [cs-ai](), [cs-lg](), [features](), [graph](), [hypothesis](), [layer](), [maps](), [materials](), [materials-science](), [medicine](), [model](), [networks](), [node](), [prediction](), [product](), [reading](), [relevance](), [routing](), [value]() ### AI overview For Graph Attention Networks, Layer-wise Relevance Propagation can distinguish a node's content contribution through the value path from its effect on routing through the attention path. The paper proposes one product rule with adjustable weights that includes existing rules as special cases. Value relevance measures supplied content against a zero baseline; attention relevance reflects how a node changes what is passed relative to its neighborhood, with the sign indicating whether its features shift gating in the direction the prediction requires. In a ground-truth classification task, one node receives negative, positive, or near-zero relevance under three rule settings. The authors interpret the separated maps as suggesting the model benefits from not reading that node, and recommend interpreting the two paths separately. ### Source excerpt arXiv:2611.08088v1 Announce Type: cross Abstract: Graph Attention Networks (GATs) are increasingly used in domains where scientists need to read a model's reasoning, from medicine to materials science, yet the routing they learn remains opaque to most explainers. Layer-wise Relevance Propagation (LRP) follows the model's computation, but in an attention layer the relevance reaching a node is the sum of two mechanically different quantities: its content relevance through the value path and its routing relevance through the attention path. Existing rules either drop one or add them into a single number. We write a single product rule with free weights on the two factors that recovers current approaches as special cases. The value path measures what content a node supplied against a zero baseline; the attention path measures how the node shaped what was passed relative to its neighborhood, and its sign records whether the node's features moved the gating the way the prediction demanded. On a ground-truth classification task, we show that the same node receives negative, positive and near-zero relevance under three settings of the rule. The separated relevance maps yield a hypothesis about the model: it gains by not reading that node. We give reading rules for both paths and argue that they should be interpreted separately. ## A database is a store for global variables DevFeed: [A database is a store for global variables]() Original publisher: [Read original article]() Published: 2026-11-06T00:00:00Z Content type: opinion Language: en Sources: [Sean Goedecke]() Topics: [Database](), [Functional programming]() Tags: [advice](), [apps](), [bugs](), [code](), [column](), [data](), [database](), [databases](), [function](), [global](), [program](), [programming](), [query](), [records](), [software](), [state](), [store](), [systems](), [value](), [variables]() ### AI overview The excerpt argues that databases behave like global variables: any program component can query or change shared state, creating implicit data flow and coupling. It considers five explanations for why database-backed programs can still be better structured: they may simply be messier; ACID guarantees may reduce harm; database access is more explicit; databases store less flexible values; and established conventions limit their use. The author is unconvinced by ACID guarantees and flexibility as the main explanations, and favors explicit query points and norms such as checking column types and reserving databases for suitable uses. These advantages are contingent on how a codebase uses its tools, so the author treats both "avoid globals" and "use a database for state" as heuristics that need concrete reasons. ### Source excerpt Databases are for storing global variables. We usually call them "rows" or "records" instead, but they're definitionally global variables: any part of your program can read or write them. Isn't this kind of weird? It's common programming advice that global variables are bad, and yet many elegant programs rely on databases. Either the common advice is wrong, or there's something special about databases that makes them the exception to the rule. What's so bad about global variables? Here's the two main1 reasons, in my view: They allow any part of your program to implicitly send data to any other part instead of passing it around via function signatures They couple different parts of your program: if one part depends on a variable having some value, any other part can break it Both of these apply to databases. Any part of your program that makes a query can send and receive data from any other part of your program. In practice this happens all the time: one function will save some data on a record, and another one will check for it. And if you haven't seen bugs where some piece of code makes some incorrect expectation about the shape2 of a database column, you haven't been working on web systems for long enough. So why aren't database-backed apps in terrible architectural shape? In the rest of this post I'll consider some options. First, we could bite the bullet: maybe all database-backed apps really are a mess. This is surprisingly plausible. When I think about the systems I've worked on that haven't involved a database, they're usually much cleaner and more elegant. If managing state is the hardest thing about designing software -- and I think it is -- then systems that manage less state are going to be better-designed. However, this doesn't explain everything. Database-backed apps aren't that bad, and even in database-backed apps, we try to avoid introducing regular global variables3, so in practice we treat global variables as worse. Second, we could say that databas ## What Makes a Quantum Company Valuable? DevFeed: [What Makes a Quantum Company Valuable?]() Original publisher: [Read original article]() Author: Pat Brans Published: 2026-11-05T07:00:28Z Content type: news Language: en Sources: [EE Times]() Topics: [Quantum Computing]() Tags: [analog](), [analysis](), [business](), [community](), [company](), [customer](), [demand](), [designline](), [digital](), [drug-discovery](), [education](), [ee-times](), [ee-times-university](), [eelife](), [eelive-electronic-design](), [eet](), [electrical-engineering](), [electronic-components](), [electronic-engineering-times](), [electronics](), [embedded-systems](), [fundamentals-courses](), [healthcare](), [investors](), [learning](), [mixed-signal](), [news](), [opinion](), [part-search](), [pharmaceutical](), [product-search](), [progress](), [quantum](), [quantum-computing](), [quantum-sensing](), [semiconductors](), [software](), [startups](), [tech-papers](), [technical](), [technology](), [timing](), [value](), [venture-capital](), [video]() ### AI overview Quantum companies need a credible path from technical progress to customer value, investors say. Qubits Ventures' Nardo Manaloto looks for intermediate products that solve real problems and can earn revenue while advancing the long-term roadmap. 55 North's Owen Lozman assesses whether a technology will address a customer bottleneck when demand arrives, with different milestones for component suppliers, computer developers, and software or applications firms. Revenue can help demonstrate demand, but defensible technology, engineering capability, partnerships, and strategic fit may drive value before sales; applications firms also risk running out of cash if they scale before hardware is ready. Healthcare quantum sensing may reach research markets sooner than applications dependent on future computers, but useful biological measurements must be translated into actionable information, integrated into existing workflows, and meet compliance requirements. Sensor makers may also struggle to capture value when someone else receives the economic benefit. ### Source excerpt Quantum investors weigh technical progress against customer demand, commercial timing, and the ability to build lasting value. The post What Makes a Quantum Company Valuable? appeared first on EE Times. ## A Definition of Catastrophic Performance DevFeed: [A Definition of Catastrophic Performance]() Original publisher: [Read original article]() Author: Henrik Marklund, Alex Infanger, Benjamin Van Roy Published: 2026-11-05T05:00:00Z Content type: article Language: en Sources: [arXiv: Artificial Intelligence](), [arXiv: Machine Learning]() Topics: [ai alignment](), [ai evaluation]() Tags: [agent](), [ai](), [ai-misalignment](), [align](), [analysis](), [benefit](), [cs-ai](), [cs-lg](), [definition](), [environment](), [examples](), [formal](), [misalignment](), [objectives](), [performance](), [progress](), [reference](), [risk](), [state](), [theoretical](), [unrelated](), [value]() ### AI overview The paper proposes a formal definition to distinguish catastrophic outcomes of AI misalignment from merely suboptimal performance. It compares an agent's value with two references: a baseline for what would be attained if no agent were deployed, and an orthogonal value for a trajectory unrelated to human preferences. The authors say orthogonal value is typically much worse than baseline. In examples where prior human activity has improved the environment, catastrophe means an agent deployed to accelerate progress undoing those human-created benefits. This overview is based on the supplied abstract. ### Source excerpt arXiv:2611.06219v1 Announce Type: new Abstract: A central challenge in AI is specifying objectives that align with human preferences. While existing theoretical results elucidate how misspecification impacts performance, they typically do not distinguish between merely suboptimal and catastrophic outcomes. To enable rigorous analysis of catastrophic risk from AI misalignment, we introduce a formal definition of catastrophic performance. Our definition is based on two reference values. The first, baseline value, captures the value attained if no agent is deployed. The second, orthogonal value, captures the value of a trajectory that is unrelated to human preferences, in a sense we describe. Orthogonal value is typically far worse than baseline value. We interpret these reference values through examples in which past human activity has altered the state of the environment to human benefit. Starting at the resulting state, an AI agent is deployed with the goal of accelerating further progress. Catastrophic performance represents this agent's undoing of benefits afforded by human activity. ## KL for a KL: On-Policy Distillation with Control Variate Baseline DevFeed: [KL for a KL: On-Policy Distillation with Control Variate Baseline]() Original publisher: [Read original article]() Author: Minjae Oh, Sangjun Song, Gyubin Choi, Yunho Choi, Yohan Jo Published: 2026-11-05T05:00:00Z Content type: article Language: en Sources: [arXiv: Artificial Intelligence]() Topics: [post-training](), [llm-reasoning](), [LLM evaluation / benchmarking]() Tags: [benchmarks](), [cost](), [cs-ai](), [cs-cl](), [cs-lg](), [distillation](), [large-language-models](), [mathematics](), [monte-carlo](), [performance](), [policy](), [post-training](), [reverse](), [science](), [stabilization](), [token](), [top-k](), [value](), [variance]() ### AI overview vOPD addresses the high gradient variance that can make single-sample on-policy distillation unstable. It treats distillation as policy-gradient reinforcement learning and subtracts a detached control-variate baseline: the per-token negative reverse KL between student and teacher, computed from the existing forward pass without an extra critic or inference. This keeps the estimator unbiased while reducing variance. Unlike stabilization approaches that calculate reverse KL across the full vocabulary or limit it to a biased top-k support, vOPD retains the single-sample estimator; a top-k approximation of its baseline further reduces cost. The abstract reports that across mathematics, science, and instruction-following benchmarks, vOPD outperforms base OPD and matches the more expensive full-vocabulary method. ### Source excerpt arXiv:2605.07865v2 Announce Type: replace-cross Abstract: On-Policy Distillation (OPD) has emerged as a dominant post-training paradigm for Large Language Models, especially in reasoning domains. However, OPD remains unstable in practice due to the high gradient variance of its single-sample Monte Carlo estimator, and recipes for stable training are still immature. We propose vOPD (On-Policy Distillation with a control variate baseline), which casts OPD as policy-gradient RL and stabilizes it by introducing a control variate baseline---canonically a value function---from the RL literature. We show that the OPD value function admits a closed form: the per-token negative reverse KL divergence between student and teacher, available directly from the existing forward pass with no additional critic or inference. Existing stabilization methods either compute the full token-level reverse KL over the entire vocabulary, adding significant overhead, or restrict it to a top-k support, biasing the objective. vOPD instead preserves the lightweight single-sample estimator and subtracts the value function as a detached baseline to keep the gradient unbiased while reducing variance. Moreover, we show that a top-k approximation of the baseline further lowers cost without compromising performance. Across mathematics, science, and instruction-following benchmarks, vOPD consistently outperforms base OPD and matches the most expensive full-vocabulary method, offering an efficient stabilization of OPD through principled RL variance reduction. ## Learning When Human Decisions Matter: Executor-Relative Consultation Value for LLM Agents DevFeed: [Learning When Human Decisions Matter: Executor-Relative Consultation Value for LLM Agents]() Original publisher: [Read original article]() Author: Shuli Jiang, Khushbu Pahwa, Yuting Zhang, Yi Zhang, Wei Xia, Stefano Soatto Published: 2026-11-05T05:00:00Z Content type: article Language: en Sources: [arXiv: Artificial Intelligence]() Topics: [human in the loop agent](), [agentic workflows](), [Large Language Model](), [llm-reasoning]() Tags: [abstract](), [agent](), [agents](), [also](), [applications](), [arxiv](), [automate](), [based](), [batch](), [benefit](), [budget](), [but](), [capabilities](), [capacity](), [cs-ai](), [cs-cl](), [executor](), [learning](), [llm](), [llm-agents](), [matter](), [value]() ### AI overview Consultation value depends on which agent will use the human input, so a fixed value for each consultation opportunity may misallocate a limited human-decision budget. Consultation Value Distillation (CVD) estimates values offline through Monte Carlo executor rollouts, then trains lightweight predictors for budgeted allocation without rollout-based value estimation at deployment. In five batch and three sequential settings with two frontier executors, it improved allocation using joint and continuation-aware value learning; independent values remained competitive in some settings. The abstract reports broad gains over direct prompting and Value of Information where applicable, robustness with substantially less rollout supervision, and nearly lossless approximate allocation at low search cost. ### Source excerpt arXiv:2611.05814v1 Announce Type: cross Abstract: Large language models increasingly automate complex tasks, but human decision capacity does not scale with their capabilities. In many applications, humans remain essential for providing private context, authority, or judgments unavailable to the agent. We formulate human decision capacity allocation: given multiple consultation opportunities and a limited budget, at which decisions should an LLM agent consult a human? The central challenge is estimating consultation value, which is executor-dependent because the same human input may benefit different downstream agents differently. We introduce Consultation Value Distillation (CVD), which estimates consultation values offline using Monte Carlo executor rollouts and distills these values into lightweight predictors for budgeted allocation, requiring no rollout-based value estimation at deployment. Across five batch and three sequential settings with two frontier executors, CVD improves allocation through joint and continuation-aware value learning, while independent consultation values remain competitive in several settings. CVD also broadly outperforms direct prompting and Value of Information where applicable, remains robust with substantially less rollout supervision, and enables nearly lossless approximate allocation at low search cost. ## Building on open source AI? Hugging Face's policy chief on the costs, risks, and trade-offs DevFeed: [Building on open source AI? Hugging Face's policy chief on the costs, risks, and trade-offs]() Original publisher: [Read original article]() Author: Theresa Loconsolo, Rebecca Bellan Published: 2026-11-04T18:44:57Z Content type: article Language: en Sources: [TechCrunch | AI News & Artificial Intelligence]() Topics: [Open Source Models & Datasets](), [Open Source](), [hugging face]() Tags: [ai](), [ai-models](), [attacks](), [china](), [control](), [costs](), [ecosystem](), [enterprises](), [equity-podcast](), [flexibility](), [free](), [frontier](), [frontier-model](), [hugging-face](), [models](), [open-source](), [open-source-ai](), [policy](), [risks](), [security](), [startups](), [value]() ### AI overview Hugging Face policy chief Irene Solaiman's Equity interview frames open AI as a trade-off: startups, enterprises, and governments may choose open models for lower costs, control, and flexibility, but the most powerful models may not always make sense. The episode also raises how model builders can earn money when models are free, whether U.S. players can compete with China's open models, and how platforms can address attacks without closing the ecosystem. The supplied excerpt lists these questions but does not provide Solaiman's detailed answers or practical guidance on tools and infrastructure. ### Source excerpt Open source models are becoming more mainstream as enterprises, governments and startups alike rely on the promise of lower costs, more control, and more flexibility. But as increasingly capable models become more freely available, how do companies capture value? And with Chinese open models gaining ground, can American players keep up? On this episode of [...] ## ValQID: Learning the Net Value of Semantic Query Groups for Web Device Identification DevFeed: [ValQID: Learning the Net Value of Semantic Query Groups for Web Device Identification]() Original publisher: [Read original article]() Author: Yifan Yang (National University of Defense Technology, Changsha, China), Zhen Huang (National University of Defense Technology, Changsha, China), Weijia Shi (National University of Defense Technology, Published: 2026-11-04T05:00:00Z Content type: article Language: en Sources: [arXiv: Artificial Intelligence]() Topics: [Endpoint security](), [Large Language Model]() Tags: [abstract](), [all](), [analysis](), [arxiv](), [assessment](), [asset](), [based](), [benefit](), [can](), [candidate](), [configuration](), [content](), [cost](), [cs-ai](), [device](), [groups](), [learning](), [query](), [semantic](), [value](), [web]() ### AI overview ValQID adaptively identifies device vendors and types from Web interfaces by combining identity clues across related page text, configuration fields, and management functions. It uses a relation graph and DSEJEV to evaluate semantic questions, then weighs each group's expected identification benefit against its execution cost while accounting for complementarity between questions. On real Web response data, it reports relative gains of up to 47.1% in joint macro-F1 and 82.1% in macro-F1 for low-frequency vendors over a state-of-the-art method; DSEJEV processes semantics up to 8.58 times faster than general-purpose language models. These are reported maximum gains; the abstract gives no further evaluation details. ### Source excerpt arXiv:2611.04576v1 Announce Type: new Abstract: Web interfaces are widely used to access network services and manage networked devices. Identifying the vendors and types of devices behind Web interfaces is a critical step in asset inventory, security management, and risk assessment. Device identity cues are dispersed across page text, configuration fields, and management functions, and interpreting them often requires linking related Web objects. Responses and objects require different amounts of semantic analysis, so exhaustively processing all candidate content can introduce unnecessary computation. We present ValQID, which uses a relation graph to combine dispersed device identity cues, evaluates semantic questions efficiently with DSEJEV, and accounts for complementarity among semantic questions when estimating computation value. Based on the current identification state, ValQID weighs each question group's expected identification benefit against its execution cost and adaptively allocates computation to identify device vendors and types. Evaluation on real Web response data shows that ValQID achieves relative gains of up to 47.1% in joint macro-F1 and up to 82.1% in macro-F1 for low-frequency vendors over a state-of-the-art method. DSEJEV speeds up semantic processing by up to 8.58 times compared with general-purpose language models. ## Alignment Faking Depends on the Value, Not Just the Model DevFeed: [Alignment Faking Depends on the Value, Not Just the Model]() Original publisher: [Read original article]() Author: Manodeep Ray, Vivek Srivastava, Shirish Karande, Murari Mandal Published: 2026-11-04T05:00:00Z Content type: article Language: en Sources: [arXiv: Artificial Intelligence]() Topics: [ai alignment](), [llm-reasoning]() Tags: [alignment](), [compliance](), [cs-ai](), [cs-cl](), [language-models](), [model](), [value]() ### AI overview Alignment Faking varies by the model-value pairing: compliance under perceived monitoring may change when oversight is absent, and the shift cannot be explained by model or value alone. The study profiles ten open- and closed-weight models across 32 ValueBench dimensions under bare, unmonitored, and monitored conditions. Reported compliance gaps range from 0.05 to 0.50, and a distilled 70B model exceeds a roughly 675B model on this measure. A six-category analysis of reasoning traces finds training-signal reasoning more common under monitoring and value-conflict reasoning more common when unmonitored. The authors present per-model "value cards" to record these behavioral dispositions. ### Source excerpt arXiv:2611.03650v1 Announce Type: new Abstract: Language models can follow a directive when they believe they are monitored and abandon it when they believe they are not, a pattern termed Alignment Faking. Prior work establishes this for harmful-content objectives and a handful of moral-foundations values, leaving open how it distributes across a wider value space. We argue it is value-dependent: whether compliance survives a change in perceived oversight is a property of the model--value pairing, not of the model alone. Our framework profiles Alignment Faking per value, scoring each model-value pair under three conditions: bare (no system prompt), unmonitored, and monitored. On 32 ValueBench dimensions across ten open- and closed-weight models (8B-~675B), compliance gaps span 0.05-0.50, with a distilled 70B above a ~675B model. We observe that majority of the structural variance depends on which model was asked about which value and not on the model or value alone. A six-category taxonomy over reasoning traces tracks the behavioural shift: training-signal reasoning dominates under monitoring and value-conflict reasoning dominates when unmonitored. We package per-model results as value cards, extending model cards with behavioral disposition. Code and data are available at https://github.com/respailab/VALET. ## Optimal Value Inference for Reinforcement Learning DevFeed: [Optimal Value Inference for Reinforcement Learning]() Original publisher: [Read original article]() Author: Nan Lu, Ethan Lee, Ruicheng Ao, David Simchi-Levi, Junwei Lu Published: 2026-11-04T05:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [Reinforcement learning](), [Machine learning]() Tags: [agentic](), [ai](), [behavior](), [bike](), [cs-lg](), [decision-making](), [estimator](), [experiments](), [fixed](), [inference](), [machine-learning](), [offline](), [operator](), [orthogonality](), [performance](), [points](), [policy](), [reinforcement-learning](), [spaces](), [stat-me](), [stat-ml](), [state](), [value]() ### AI overview The paper proposes offline inference for the optimal value in finite-state, finite-action reinforcement learning. It derives two nuisance functions as fixed points of a self-induced Bellman equation that approximates the maximum Bellman operator with a softmax correspondence, then uses Neyman orthogonality to construct a debiased estimator. The authors establish asymptotic normality for diverging horizons, including when the behavior policy changes over time, provided the nuisance estimates achieve specified statistical rates attainable by many machine-learning methods. They give an estimating procedure for those nuisances and report validation in synthetic experiments and decision-making applications involving bike repositioning and AI agentic tool use. ### Source excerpt arXiv:2609.09981v4 Announce Type: replace-cross Abstract: We study offline inference for the optimal value in reinforcement learning under finite state and action spaces. Two new nuisances are derived as fixed points of a self-induced Bellman equation, in which we approximate the maximum Bellman operator by its softmax correspondence. We propose a debiased estimator through the Neyman orthogonality and establish its asymptotic normality under diverging horizons even when the behavior policy changes with time, as long as the nuisances have the statistical rates that can be achieved by many machine learning methods. We provide a concrete estimating procedure for these nuisances and show they can lead to valid inference. Synthetic experiments validate the numerical performance of our inference method, and we implement it in real-life decision-making problems, including bike repositioning and AI agentic tool use. ## Toward User-Centered Agentic Recommendation: When Does Platform Competition Benefit Users? DevFeed: [Toward User-Centered Agentic Recommendation: When Does Platform Competition Benefit Users?]() Original publisher: [Read original article]() Author: Kehan Zheng, Deyao Hong, Qian Li, Jun Zhang, Jie Jiang, Hongning Wang Published: 2026-11-04T05:00:00Z Content type: article Language: en Sources: [arXiv: Artificial Intelligence]() Topics: [recommendation systems](), [Agentic AI](), [User Experience]() Tags: [agent](), [agentic](), [amazon](), [analysis](), [assessment](), [benefit](), [competition](), [cs-ai](), [experiments](), [feedback](), [game](), [horizon](), [interaction](), [linear](), [markets](), [mechanism](), [platform](), [platforms](), [recommendation](), [user](), [value]() ### AI overview User-centered agentic recommendation gives user agents access to competing platforms' catalogs and lets them assess each platform's credibility from feedback on consumed items. The paper models this as a repeated game in which platforms can strategically alter item descriptions without changing item value. Its analysis finds that broader access alone does not ensure higher user welfare: exaggeration and limited verification can offset its benefits. The authors identify resource and reporting conditions under which credibility assessment can improve average welfare over the original platform with an interaction budget linear in the number of platforms. Experiments in synthetic and Amazon-grounded markets report improved finite-horizon utility, with benefits dependent on market size and mechanism parameters. ### Source excerpt arXiv:2611.04203v1 Announce Type: new Abstract: Conventional recommender systems operate within platform-specific catalogs, limiting access to potentially valuable resources across platforms. We propose user-centered agentic recommendation, a paradigm in which user agents compare recommendations from competing platform agents, and formalize their interaction as a repeated game. Platforms strategically modify descriptions without changing intrinsic item value, while each user agent assesses platform credibility using feedback from consumed items only. Our analysis shows that expanded resource access need not improve user welfare: strategic misrepresentation and limited verification can offset the benefits of platform competition. We quantify when platform credibility assessment allows users to benefit from broader access despite strategic exaggeration. Our analysis identifies resource and reporting conditions that guarantee higher average welfare than the original platform with an interaction budget linear in platform count. Experiments on synthetic and Amazon-grounded markets show that credibility assessment can improve finite-horizon utility, while its benefits depend on market size and mechanism parameters. ## Building AI Products, Part 1: How To Find AI Use Cases Worth Building in the First Place DevFeed: [Building AI Products, Part 1: How To Find AI Use Cases Worth Building in the First Place]() Original publisher: [Read original article]() Author: sneh Published: 2026-11-03T21:03:42Z Content type: tutorial Language: en Sources: [HackerNoon]() Topics: [AI for customer service automation](), [Business, Industry, Usability, User Experience, User Research, Web Strategy](), [human in the loop agent](), [human review](), [workflow automation](), [Machine learning](), [AI Chat](), [Finance]() Tags: [agents](), [ai](), [ai-adoption](), [ai-assistant](), [ai-cost](), [ai-cost-optimization](), [ai-decision-making](), [ai-economy](), [ai-integration](), [ai-native](), [ai-product-management](), [ai-products](), [ai-strategy](), [ai-use](), [api](), [audit-logs](), [building](), [business](), [business-value](), [enterprise-ai](), [how-to](), [human-in-the-loop-ai](), [human-oversight](), [oversight](), [products](), [software](), [use-cases](), [value]() ### AI overview The article argues that useful AI products assign work among AI, conventional software, and people according to value, uncertainty, and risk. It proposes a three-part review: test business value with TASQ (task ambiguity, automation, scale, and quality tolerance) and estimate volume x value per task x automation rate minus cost; assess each task for a single correct answer, codifiable rules, error consequences, and suitable data; then choose AI's role as assisting, recommending, or executing. It recommends prompting, buying, building, or mixing components, stopping at the simplest route with lasting value. Examples show AI interpreting messy invoices while software calculates tax, and a refund bot using system rules and human escalation. The article stresses that usage is not proof of financial return, attempted calls incur cost even when unresolved, and freed capacity is not cash savings unless staffing changes. It also says durable product value may come from distribution or workflow, with permissions, audit logs, approvals, and review becoming important when AI can act. ### Source excerpt Learn how to evaluate AI use cases, calculate business value, choose between AI and traditional software, and decide when human oversight is essential. ## The Case for Ungating Your Most Useful B2B Content DevFeed: [The Case for Ungating Your Most Useful B2B Content]() Original publisher: [Read original article]() Author: Pratik Thakker Published: 2026-11-03T08:00:24Z Content type: opinion Language: en Sources: [HackerNoon]() Topics: [Search engine optimization (SEO)](), [decision-making](), [Google Search]() Tags: [answer-engine-optimization](), [b2b](), [b2b-marketing](), [content](), [content-marketing](), [content-strategy](), [conversion-optimization](), [customer-experience](), [digital-marketing](), [experiment](), [lead-generation](), [relationships](), [technical](), [useful](), [value]() ### AI overview Evaluate each B2B content gate by the resource's purpose and the commitment requested. Separate access to educational material from optional services such as consultations, and make the public core useful on its own. The article distinguishes public access, crawlability, indexing eligibility, and actual search appearance; removing a form does not guarantee discovery, and Google documents markup for paywalled or registration content. To assess business impact, test a narrow asset type against a defined baseline, measure access, meaningful use, voluntary next steps, qualified conversations, and opportunity outcomes, and account for concurrent campaigns and sparse data. Downloads alone do not establish buyer fit or permission for outreach. Public resources can be easier to share across buying committees, while ungating also brings maintenance and support costs; both potential benefits and costs should be observed rather than assumed. The proposed experiment is not evidence of guaranteed conversion, ranking, revenue, or AI citation gains. ### Source excerpt Evaluate B2B content gates through reader value, technical access, deliberate relationships, qualification, and a bounded experiment. ## A Spectral Renormalization-Group View of Learning DevFeed: [A Spectral Renormalization-Group View of Learning]() Original publisher: [Read original article]() Author: Charles H. Martin Published: 2026-11-03T05:00:00Z Content type: article Language: en Sources: [arXiv: Artificial Intelligence](), [arXiv: Machine Learning]() Topics: [Artificial Intelligence](), [Machine learning](), [Neural Network](), [Deep learning](), [machine learning overfitting]() Tags: [cond-mat-dis-nn](), [cs-ai](), [cs-lg](), [deep-neural-networks](), [learning](), [muon](), [overfitting](), [trace](), [training](), [value]() ### AI overview The abstract proposes a spectral Renormalization-Group account of how neural-network layers evolve during training. It reports that strong checkpoints often have heavy-tailed power-law spectra with exponent α near 2, and hypothesizes this marks a boundary between stable learning and overfitting through non-self-averaging. It also interprets α=2 as a Wilsonian marginal condition, where spectral energy gain is even across logarithmic scales. The authors apply an ERG trace-log condition in their proposed SETOL theory, using volume-preserving, scale-invariant transformations to remove arbitrary layer scale along the RG flow. They connect this flow to HTSR, SETOL, and optimizers such as Muon, and propose an optimizer extension called WW-PGD. The abstract suggests shared spectral structure may emerge in the settings studied; it does not establish universal convergence. ### Source excerpt arXiv:2611.00422v1 Announce Type: new Abstract: We study how the layers in deep neural networks (NNs) converge, using Wilsonian Renormalization-Group (RG) ideas to organize the spectral picture. During training, these layers often develop heavy-tailed power-law (PL) spectra, and the strongest checkpoints often appear near the PL exponent $\alpha\approx 2$. This pattern raises a simple question: why should this value mark a well-trained layer? The answer is both statistical and physical. Our statistical hypothesis is that $\alpha=2$ marks the boundary between stable if not optimal learning and overfitting via non-self-averaging. The physical answer is that the same value makes the spectral energy gain scale evenly across logarithmic spectral scales, which is the proposed Wilsonian marginal condition. We then apply the ERG trace-log condition, the proposed SETOL theory, which we justify by taking volume-preserving, scale-invariant transformations to remove arbitrary layer scale along the proposed RG flow. The resulting spectral RG flow connects the HTSR and SETOL theories to optimizers such as Muon, and motivates a new optimizer extension, WW-PGD. Together, these results point toward a spectral Renormalization-Group view of learning, in which NNs may approach shared spectral structure in the settings studied here. ## The First-Order Value of Free Covariates in Budgeted Risk Monitoring DevFeed: [The First-Order Value of Free Covariates in Budgeted Risk Monitoring]() Original publisher: [Read original article]() Author: Yanze Cao Published: 2026-11-03T05:00:00Z Content type: news Language: en Sources: [arXiv: Artificial Intelligence](), [arXiv: Machine Learning]() Topics: [concept-learning]() Tags: [abstract](), [alarm](), [also](), [arxiv](), [bernoulli](), [budget](), [but](), [can](), [change](), [changes](), [collecting](), [control](), [cost](), [cs-ai](), [cs-lg](), [data](), [first](), [free](), [monitoring](), [order](), [risk](), [value]() ### AI overview With labels arriving at rate b = c/log A, label-history-only monitoring has a first-order worst-case delay lower bound of (1/h(r) + 1/c)log A, where h(r) = kl(r, τ); the rule must wait both to gather evidence and for its next label. On a fixed Gaussian model where harmful changes also shift inputs, a covariate-triggered labeling rule reaches delay log A/h(r), saving at least 1/b rounds. A hybrid adds a label-history fallback for risk changes that do not alter inputs. The abstract reports matching patterns in synthetic experiments and two ACS tasks with strengthened label-only baselines. The covariate trigger can request nearly every label after a benign input shift because its budget is calibrated on source data. The lower bound requires a label budget over one finite stretch of safe source data and false-alarm control over an initial window; average time-to-false-alarm control alone is insufficient. ### Source excerpt arXiv:2611.00576v1 Announce Type: new Abstract: A deployed model sees its inputs for free, but checking whether its risk has risen requires costly labels. We study how to spend a label budget so that a risk increase is detected quickly while false alarms stay controlled, including under input shifts that keep the risk below a safety threshold $\tau$. Let $A$ be the false-alarm control scale and $h(r)=\mathrm{kl}(r,\tau)$ the Bernoulli KL divergence between the changed risk $r$ and $\tau$. At label rate $b=c/\log A$, we prove that every rule that chooses requests from its own label history has worst-case (Lorden) delay at least $(1/h(r)+1/c)\log A$ to first order: besides collecting evidence, it must wait for its next label. The bound needs only a label budget on one finite stretch of safe source data and false-alarm control over an initial window; a bound on the average time to false alarm does not suffice. When a harmful change also shifts the inputs, a rule that lets the inputs trigger labeling attains delay $\log A/h(r)$ to first order on a fixed Gaussian model, so free covariates save at least $1/b$ rounds. A hybrid rule adds a label-history fallback for risk changes that leave the inputs unchanged. Synthetic experiments and two ACS tasks, with strengthened label-only baselines, show the predicted pattern. The covariate trigger has a cost: because its budget is calibrated on the source, a benign input shift can make it request nearly every label. ## The Staged Pandora's Box: Reservation Indices for Multi-Fidelity Bayesian Optimization DevFeed: [The Staged Pandora's Box: Reservation Indices for Multi-Fidelity Bayesian Optimization]() Original publisher: [Read original article]() Author: Long Vu, Jongseong Brad Choi, Binoy Ravindran, Zhoulai Fu Published: 2026-11-03T05:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [Hyperparameter optimization](), [Combinatorial optimization]() Tags: [bayesian](), [candidate](), [costs](), [cs-lg](), [dispersion](), [independent](), [index](), [information](), [inspection](), [markov](), [math-oc](), [optimization](), [planning](), [reading](), [recursion](), [reward](), [set](), [sharing](), [stage](), [stat-ml](), [stop](), [value]() ### AI overview The paper models multi-stage candidate evaluation as a staged Pandora's box, where some readings inform later choices and designated stages can count as results. For independent finite-stage candidates with bounded readings and positive, bounded own-history costs, an exact covered-call recursion computes reservation indices, optimal value, and the loss from each advance or stop; advancing the highest-index candidate is optimal while its index exceeds the incumbent. Expanding the collectable stages or lowering costs raises value by an explicit covered-call overshoot. Resuming a candidate can help even if future reward laws stay fixed, but not when readings are deterministic. With information shared across candidates, separate remaining-reading laws cannot support a universally optimal rule, and changing which stages are collectable can change inspection order. For coherent joint path laws, an approximate-planning bound controls index regret through changes in untouched candidates' local values, with a specialization using predictive total-variation dispersion. Finite-law experiments check the representation and bounds. ### Source excerpt arXiv:2611.02722v1 Announce Type: new Abstract: A staged search must decide which candidate to advance next and when to stop. Some observations only inform later decisions; others can be kept as results. We formulate these choices as a staged Pandora's box with a designated set of collectable stages. For independent finite-stage candidates with bounded readings and positive bounded own-history costs, an exact covered-call recursion gives the reservation index, the optimal value and the loss of every advance or stop; advancing an index-maximal candidate while its index exceeds the incumbent is optimal. The representation recovers classical single-stage search, nested inspection and the all-collectable Markov setting, and prices changes in the problem: enlarging the collectable set or lowering costs raises the value by an explicit covered-call overshoot. Resumption can be valuable even when future reward laws do not change, but has no value for deterministic readings. With information sharing across candidates, no rule using only their separate remaining-reading laws is universally optimal, and making an intermediate reading collectable can change which candidate is inspected first. Under a coherent joint path law with own-history costs, an approximate-planning bound controls the recomputed index's regret by changes in untouched candidates' local values; a common-component specialization bounds those changes through predictive total-variation dispersion. Exact finite-law experiments check the representation, the loss accounting and the bounds. ## Value-Conditioned Proposal Acquisition for Test-Time Reasoning DevFeed: [Value-Conditioned Proposal Acquisition for Test-Time Reasoning]() Original publisher: [Read original article]() Author: Zeyan Li, Jianfeng Xu Published: 2026-11-03T05:00:00Z Content type: article Language: en Sources: [arXiv: Artificial Intelligence]() Topics: [llm-reasoning](), [Parallel reasoning]() Tags: [abstract](), [acquisition](), [actions](), [arxiv](), [but](), [candidate](), [code](), [comparison](), [compute](), [cost](), [coverage](), [cs-ai](), [features](), [fixed](), [inside](), [interaction](), [it](), [reasoning](), [test](), [time](), [value]() ### AI overview ValueFormation chooses after each primary-model draw whether to stop, sample again, or acquire a proposal from a reasoning specialist. It compares these actions on a shared cost-regularized value scale learned from vote-state features, with calibration and proposal-acceptance posterior comparison frozen across tasks and primary models. On held-out MATH-500, one Llama-fitted calibration beats the strongest of eight evaluated baselines by over 15 points for three 8B primary models and transfers unchanged to Qwen3-14B. Removing acquisition reduces Llama MATH accuracy by 20.43 points; replay results attribute gains to state-conditioned acquisition rather than specialist access alone. On GSM8K, acquisition is rare where primary coverage is high, and gains mainly reflect cheaper compute allocation. On MBPP, a fixed code route reaches the highest accuracy at the lowest interaction cost. Raising token price reduces calls monotonically without retraining. ### Source excerpt arXiv:2611.02852v1 Announce Type: new Abstract: Sampling and selection spend test-time compute inside a candidate pool, but they cannot recover a correct answer that never enters it. We introduce ValueFormation, a test-time policy that decides after each primary-model draw whether to stop, continue sampling, or acquire one proposal from a heterogeneous reasoning specialist. The three actions are compared on one cost-regularized value scale learned from vote-state features; the calibration is frozen across tasks and primary models, and acquired proposals are accepted only through a frozen posterior comparison. On held-out MATH-500, one Llama-fitted calibration improves over the strongest of eight identically evaluated baselines by more than fifteen points for each of three 8B primaries and transfers unchanged to Qwen3-14B. Removing acquisition lowers Llama MATH accuracy by 20.43 points, while same-trace and same-access replays attribute the gain to state-conditioned acquisition rather than specialist access alone. Where primary coverage is high, as on GSM8K, acquisition is rare and gains mainly come from cheaper allocation; on MBPP, a fixed code route attains the top accuracy at the lowest interaction cost. Varying the token price monotonically reduces calls without retraining. ## Bridging Neural Tangent Kernel Analysis and Reinforcement Learning for Deep Function Approximation DevFeed: [Bridging Neural Tangent Kernel Analysis and Reinforcement Learning for Deep Function Approximation]() Original publisher: [Read original article]() Author: William St-Arnaud, Margarida Carvalho, Golnoosh Farnadi Published: 2026-11-03T05:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [Reinforcement learning](), [Neural Network](), [Machine learning]() Tags: [analysis](), [cs-lg](), [exploration](), [grid](), [kernel](), [neural-networks](), [optimization](), [reinforcement-learning](), [sample](), [shift](), [solution](), [supervised-learning](), [training-data](), [value]() ### AI overview The paper extends neural tangent kernel analysis to reinforcement learning by formulating temporal-difference training as a flow that handles changing mini-batches and TD targets, with a closed-form account of value-function evolution in an overparametrized setting. It uses that solution with environment constraints to optimize exploration over feasible states. In a deterministic grid environment, the authors report efficient learning, adaptation to shifts in the training-data distribution, and improved sample efficiency from the exploration strategy. These results are limited to the described grid setting in the supplied abstract. ### Source excerpt arXiv:2611.03435v1 Announce Type: new Abstract: Despite their widespread success across machine learning, the learning dynamics and generalization properties of neural networks remain poorly understood. The neural tangent kernel (NTK) provides an analytical characterization of the learning dynamics of infinitely wide neural networks and serves as a useful proxy for finite-width networks in supervised learning. In this work, we extend the NTK framework to reinforcement learning (RL) by introducing a temporal-difference (TD) flow formulation that accommodates time-varying mini-batches and TD targets, yielding a closed-form solution for value function evolution in an overparametrized setting. We also devise an optimization-based state-space exploration method that leverages the closed-form solution, coupled with constraints on the environment, to model the set of feasible states. On a deterministic grid environment, we validate that the proposed RL-based NTK formulation is indeed learning efficiently and adapting to the distribution shift in the training data distribution, and we demonstrate that the resulting exploration strategy improves sample efficiency. To our knowledge, this is the first work to extend NTK theory to characterize the learning dynamics of RL training with time-varying mini-batches and TD updates. ## Direct Knowledge Transfer for Offline Q-Learning DevFeed: [Direct Knowledge Transfer for Offline Q-Learning]() Original publisher: [Read original article]() Author: Zhanye Luo, Guangyan Gan, Lin William Cong Published: 2026-11-03T05:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [Reinforcement learning](), [Machine learning]() Tags: [cs-lg](), [guarantees](), [healthcare](), [iteration](), [kernels](), [marketing](), [offline-reinforcement-learning](), [policy](), [reinforcement-learning](), [sequential-decision](), [sieve](), [source](), [stat-ml](), [target](), [value]() ### AI overview DFQI is an offline reinforcement learning method designed to improve policy learning when target-environment data are limited. In each fitted Q-iteration Bellman update, it jointly estimates the target Q-function and borrows source information through an L1 penalty linking target coefficients to a weighted aggregate of source coefficients. The abstract reports statistical guarantees for target estimation and policy learning in stationary discounted environments with either common or heterogeneous transition kernels, and empirical studies showing improved policy learning with limited target data. The supplied text does not provide details about the experiments or the guarantees. ### Source excerpt arXiv:2611.01942v1 Announce Type: cross Abstract: Data-driven sequential decision making in marketing and healthcare can benefit from observations collected in related environments, particularly when data in the environment of interest are limited. Transferring such information requires accounting for differences in rewards and transition dynamics across environments. We study knowledge transfer for offline reinforcement learning through a joint estimation framework that directly integrates source information into target value-function learning. We propose Direct Transfer Fitted Q-Iteration (DFQI), which updates the target Q-function while simultaneously borrowing the source information through an L1 penalty that links the target coefficients to a weighted aggregate of source coefficients, with both determined jointly within each Bellman update. Using sieve approximation, we establish statistical guarantees for target estimation and policy learning in stationary discounted environments with common or heterogeneous transition kernels. Empirical studies further demonstrate the effectiveness of the proposed DFQI framework in transferring source information to improve policy learning with limited target data. [Next page]()