# values Published articles for values. This is one page of public article previews, not the complete archive. Follow Next page to continue. Summaries are not the original full articles. ## Beyond the Lower-Star Filtration: Alternative Filtration Methods for Persistent Homology DevFeed: [Beyond the Lower-Star Filtration: Alternative Filtration Methods for Persistent Homology]() Original publisher: [Read original article]() Author: Innocent Abaa, Umar Islambekov Published: 2026-11-09T05:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [Algorithms]() Tags: [aggregation](), [boundary](), [cs-lg](), [datasets](), [delay](), [diagram](), [dimension](), [loops](), [math-at](), [persistence](), [persistent-homology](), [stat-ap](), [stat-ml](), [values](), [vertex]() ### AI overview In the lower-star filtration (LSF), a simplex enters when its highest-valued vertex appears, so some higher-dimensional simplices arrive with their boundary faces and can immediately fill features that then do not appear in the persistence diagram. The proposed alternatives address this in different ways: ELSF adds a dimension-dependent delay to the maximum-vertex rule, while LpF and SMF aggregate values from all vertices. These assignments can let loops and voids persist longer and reveal features absent under LSF. Experiments on simulated and real-world datasets reportedly found the methods computationally efficient and more sensitive than LSF to structural differences; the abstract does not provide specific datasets or performance measures. ### Source excerpt arXiv:2611.08833v1 Announce Type: cross Abstract: In the lower-star filtration (LSF), simplices are assigned filtration values according to the maximum-vertex rule, with each simplex entering the filtration when its highest-valued vertex appears. A limitation of this construction is that some higher-dimensional simplices may enter simultaneously with their boundary faces, causing certain topological features to be filled immediately upon formation and thus not appear in the resulting persistence diagram. To address this limitation, we propose three alternative filtration methods: the \(\varepsilon\)-lower-star filtration (ELSF), the \(L_p\) filtration (LpF), and the softmax filtration (SMF). {ELSF introduces a dimension-dependent delay to the maximum-vertex rule,} allowing higher-dimensional simplices to enter the filtration after their boundary faces, whereas {LpF and SMF use aggregation functions that depend on all vertex values rather} than solely on their maximum. By modifying how filtration values are assigned, these methods allow simplices to enter the filtration more progressively, so that topological features such as loops and voids may persist over longer filtration intervals rather than being filled immediately. Consequently, the resulting persistence diagrams may include persistence features that are not observed under LSF, potentially providing useful information for distinguishing data populations. Experimental results on simulated and real-world datasets demonstrate that the proposed filtration methods are computationally efficient and can provide greater sensitivity to structural differences in the data than LSF. ## LowRAM: Low-Rank Keys and Amortised Values for Composable KV Cache Compression at Test-Time DevFeed: [LowRAM: Low-Rank Keys and Amortised Values for Composable KV Cache Compression at Test-Time]() Original publisher: [Read original article]() Author: Teresa Delgado, Hugo Hazard, Martin Benfeghoul, Jun Wang, Zafeirios Fountas Published: 2026-11-06T05:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [Compression](), [Cache](), [llm-reasoning](), [quantization]() Tags: [attention](), [autoregressive](), [cache](), [compression](), [context](), [cross](), [cs-cl](), [cs-lg](), [decode](), [decoding](), [dimensions](), [exploit](), [inference](), [keys](), [kv-cache](), [layers](), [learning](), [lightweight](), [rank](), [test](), [time](), [values]() ### AI overview LowRAM compresses attention keys and values using different structure: it applies cross-layer low-rank compression to keys and fits a lightweight per-head predictor to reconstruct values from keys, storing residuals for tokens reconstructed worst. The predictor can be meta-learned to reduce prefill overhead or fitted from scratch when a one-off cost is acceptable, such as in multi-query inference. The abstract reports 4x near-lossless value compression and, at 4x overall compression, up to 5x higher peak decode throughput by reconstructing only selected KV pairs per step. Combined with quantisation and eviction, LowRAM reaches 26.9x compression at 94.5% of baseline performance on long contexts. These are abstract-level reported results across three model families; the excerpt provides no detailed evaluation setup. ### Source excerpt arXiv:2611.07025v1 Announce Type: new Abstract: Autoregressive decoding relies on the KV cache, whose memory footprint grows linearly with context length and caps batch size and throughput. Existing post-training methods exploit redundancy across tokens, layers, and feature dimensions, or lower numerical precision. We identify a complementary axis: within each attention head, values are largely predictable from their corresponding keys. We exploit this with value cache amortisation, which fits a lightweight per-head predictor at test time and stores residuals only for the tokens it reconstructs worst, giving $4\times$ near-lossless value compression. The predictor can be initialised by meta-learning for substantially reduced prefill overhead, or fitted from scratch in settings that allow a one-off cost, such as multi-query inference. We pair value amortisation with cross-layer low-rank key compression to produce LowRAM, an asymmetric compression pipeline that stays close to lossless across three model families. By reconstructing only a selected subset of KV pairs at each decoding step, LowRAM can keep decoding overhead minimal and achieve up to $5\times$ higher peak decode throughput at $4\times$ compression. LowRAM further composes with quantisation and eviction, reaching $26.9\times$ compression at $94.5\%$ of baseline performance on long contexts. These results show that treating keys and values as structurally distinct enables efficient, aggressive, and composable KV cache compression. ## RecGroupMemory: Feedback-Aware Cohort Memory for E-Commerce Recommendation DevFeed: [RecGroupMemory: Feedback-Aware Cohort Memory for E-Commerce Recommendation]() Original publisher: [Read original article]() Author: Wendong Zhang, Weipeng Huang, Zhixin Ma, Bin Zhang, Dimin Wang Published: 2026-11-03T05:00:00Z Content type: article Language: en Sources: [arXiv: Artificial Intelligence]() Topics: [personalization](), [AI Chat]() Tags: [advice](), [commerce](), [comparisons](), [cs-ai](), [e-commerce](), [feedback](), [guidance](), [llm](), [memory](), [online](), [orchestration](), [planning](), [recommendation](), [revision](), [transaction](), [user](), [users](), [validation](), [values](), [workflow]() ### AI overview RecGroupMemory gives an orchestration LLM group-level user profiles and recommendation guidance, separating memory construction from later revision. It uses the previous day's outcome feedback to update decision guidance, with validation checks intended to keep selected source values unchanged. In an online comparison among Taobao users with low purchasing power selected by the business team, memory without feedback increased guided transaction users by 1.58% and memory with feedback by 1.78%, each relative to a no-memory baseline. The excerpt does not specify sample size or further experimental details. ### Source excerpt arXiv:2611.02453v1 Announce Type: new Abstract: Cohort memory can support LLM-based recommendation planning through group-level user profiles and advice. As outcome feedback becomes available, this guidance may need to be updated while preserving the underlying observations. We present RecGroupMemory, a workflow that provides this memory to an orchestration LLM and separates memory construction from feedback-based revision. It uses feedback from the previous day to revise decision guidance, while validation checks keep selected source values unchanged. Online comparisons were conducted among Taobao users with low purchasing power selected by the business team. Relative to a no-memory baseline, memory without feedback and memory with feedback showed increases of 1.58% and 1.78%, respectively, in the number of guided transaction users. ## Invertible Graph Aggregation Can Increase ReLU Readout Width DevFeed: [Invertible Graph Aggregation Can Increase ReLU Readout Width]() Original publisher: [Read original article]() Author: Hiroyasu Usami, Masanori Tsukamoto, Kohei Hashiguchi Published: 2026-11-03T05:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [Plover](), [Graphs]() Tags: [aggregate](), [aggregation](), [condition](), [cs-lg](), [error](), [features](), [fixed](), [free](), [graph](), [has](), [hidden](), [lazy](), [matrix](), [prediction](), [raw](), [relu](), [roman-empire](), [scalar](), [stat-ml](), [training](), [units](), [values]() ### AI overview An invertible graph aggregation can preserve all node features yet increase the ReLU readout width needed for classification. For each even N >= 4, the authors construct a maximum-degree-three tree where a shared one-hidden-layer classifier uses one unit on raw scalar features but exactly N after a lazy neighborhood mean; aggregate values remain distinct and the mixing matrix has condition number at most 8. The abstract also reports a complete error curve with a free affine term and persistence under scalar perturbations. Under a specified bounded uniform-noise model with one nuisance coordinate, every one-unit aggregate head has expected error at least 1/12, while a fixed N-unit head achieves zero. In controlled fits on Roman-empire and Tolokers, aggregation improves fitted prediction, while retaining raw features helps at the largest tested parameter budget. Training-only diagnostics give conditional output-layer optimality bounds for eight fixed runs. ### Source excerpt arXiv:2611.02274v1 Announce Type: new Abstract: An invertible graph aggregation preserves the complete feature table, but can increase the size of a classifier that reads one node at a time. For every even $N\ge4$, we construct a labeled tree with $2N$ vertices and maximum degree three. A shared one-hidden-layer rectified linear unit (ReLU) classifier needs one hidden unit on the raw scalar features and exactly $N$ units after one lazy neighborhood mean. All aggregate values are distinct, and the mixing matrix has condition number at most 8. We determine the complete error curve for heads with a free affine term. The width separation persists under scalar perturbations. Under a specified bounded uniform-noise model with one nuisance coordinate, every one-unit aggregate head has expected error at least $1/12$, while a fixed $N$-unit head has zero error. Controlled fits compare trained heads with these explicit feasible scores. On Roman-empire and Tolokers, aggregation improves fitted prediction, and retaining raw features helps at the largest tested parameter budget. Training-only diagnostics include conditional output-layer optimality bounds for eight fixed runs. ## Towards Multi-Value Alignment via Value Localization in LLMs DevFeed: [Towards Multi-Value Alignment via Value Localization in LLMs]() Original publisher: [Read original article]() Author: Xueqi Ma, Yanbei Jiang, Xingjun Ma, James Bailey, Sarah M. Erfani Published: 2026-11-03T05:00:00Z Content type: article Language: en Sources: [arXiv: Artificial Intelligence]() Topics: [ai alignment](), [Large Language Model](), [LLM evaluation / benchmarking]() Tags: [activation](), [activation-steering](), [alignment](), [alignment-benchmarks](), [attention](), [control](), [cs-ai](), [fine-tuning](), [functional](), [inference](), [interpretability](), [large-language-models-llms](), [llms](), [localization](), [mechanistic-interpretability](), [theoretical](), [time](), [value](), [values]() ### AI overview The paper identifies sparse sets of attention heads linked to specific values and style preferences in LLMs, using a probing-based localization framework. Across multiple model families, it reports specialized heads, structured interactions, and conflicts; targeted ablations produce value-specific behavioral changes, supporting a causal role for these heads. The authors then use inference-time activation steering to control individual values or combinations without changing model parameters. On multiple alignment benchmarks, they report effective single- and multi-value control while largely preserving performance on non-target behaviors. Evidence here is limited to the supplied abstract. ### Source excerpt arXiv:2611.02503v1 Announce Type: new Abstract: Aligning large language models (LLMs) with multiple, potentially conflicting human values remains challenging, as existing multi-objective alignment methods often require costly fine-tuning and provide limited insight into the internal mechanisms underlying value-related behaviors. In this work, we investigate multi-value alignment from a mechanistic interpretability perspective by identifying value-sensitive attention heads (value heads) in LLMs. Using a probing-based value localization framework, we discover sparse sets of attention heads associated with diverse dimensions, spanning both human values (e.g., harmlessness, honesty) and stylistic preferences (e.g., coherence, complexity). Analyses across multiple LLM families reveal functional specialization and structured interactions among value heads, including conflicts between certain values. Targeted head ablations further induce substantial, value-specific behavioral changes, providing causal evidence of their functional importance. Building on these findings, we propose an inference-time activation steering strategy that enables controllable alignment with individual values or combinations of values without updating the underlying model parameters, supported by theoretical analysis. Experiments across multiple value alignment benchmarks demonstrate effective single- and multi-value control while largely preserving non-target model performance. Our findings connect mechanistic interpretability with flexible multi-value alignment, offering a foundation for more interpretable and pluralistic LLM behavior. The code is available at https://github.com/sihuo-design/ValueHeads. ## Training Autoregressive Language Models on Rounded, Bounded, and Censored Numerical Labels via Exact Likelihood DevFeed: [Training Autoregressive Language Models on Rounded, Bounded, and Censored Numerical Labels via Exact Likelihood]() Original publisher: [Read original article]() Author: Jingquan Yan, Yuwei Miao, Junzhou Huang Published: 2026-11-03T05:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [LLM Techniques](), [Transformers]() Tags: [arxiv](), [autoregressive](), [code](), [continuous](), [cs-lg](), [distribution](), [distributions](), [experiments](), [gradient](), [grid](), [language-models](), [latency](), [measurement](), [supervision](), [tokens](), [training](), [trains](), [values]() ### AI overview Boundary marginalization lets an autoregressive language model train on the exact likelihood of interval-valued numerical observations using its existing digit tokens. A fixed-width, order-preserving representation groups grid values by shared digit prefixes; an interval's probability and gradient can be computed from at most two endpoint-prefix paths, each logarithmic in the support size. The objective covers precise, rounded, bounded, and censored labels, and the same computation supports exact distributional queries and unbiased continuous ranked probability score gradients for exact labels. The abstract reports recovery of conditional distributions from rounded and censored reports, improved likelihood on naturally censored molecular measurements, and more accurate code-latency distributions on two backbones. It says interval updates remain tractable for supports up to one million values. ### Source excerpt arXiv:2611.02461v1 Announce Type: new Abstract: Numerical labels often specify a set of possible values: a measurement is rounded, a laboratory test reports only a bound, or a program is stopped at a timeout. We introduce boundary marginalization, the first method that trains an autoregressive language model on the exact likelihood of such observations through its existing digit tokens. A fixed-width, order-preserving numeral representation groups consecutive grid values under shared digit prefixes. An interval is a union of such groups, so its exact probability and gradient follow from at most two boundary paths, the successive digit prefixes of its endpoints, each of length logarithmic in the number of grid values. The resulting objective unifies precise, rounded, bounded, and censored supervision and trains digit decisions at the resolution supplied by each observation. The same computation supports exact distributional queries and unbiased gradients for the continuous ranked probability score on exact labels. Experiments show recovery of a conditional distribution from rounded and censored reports alone, improved likelihood on naturally censored molecular measurements, and more accurate code latency distributions on two backbones. Exact interval updates remain tractable as the numerical support grows to one million values. ## GatedV: Self-Gated Values for Transformer Language Models DevFeed: [GatedV: Self-Gated Values for Transformer Language Models]() Original publisher: [Read original article]() Author: Li-Ming Yu Published: 2026-11-03T05:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [LLM Techniques](), [Chinchilla scaling law](), [DGX Spark]() Tags: [attention](), [cs-cl](), [cs-lg](), [decoder](), [language-models](), [latency](), [nvidia-dgx](), [performance](), [perplexity](), [training](), [transformer](), [value](), [values]() ### AI overview GatedV applies SiLU to the attention values before aggregation, adding no trainable parameters and leaving the attention-weighting mechanism unchanged. In matched decoder-only Transformer comparisons with paired initializations and equal token budgets, it lowered mean held-out perplexity from 59.1 to 57.9 at 2.33M parameters, improving across all five seeds. At 26.16M parameters, after an equal-budget learning-rate search, perplexity fell from 20.2 to 19.7 across all three seeds. A small-model ablation found tested affine value rescaling did not fully reproduce the gains. A short interleaved NVIDIA DGX Spark benchmark measured about 0.7% higher full-context forward latency for the larger model; the reported results are limited to these experiments. ### Source excerpt arXiv:2611.02145v1 Announce Type: cross Abstract: Standard attention forms query (Q), key (K), and value (V) matrices through learned linear projections of input hidden states. Architectures such as GAU and Mega incorporate nonlinear transformations into attention. In this paper, we propose GatedV, a simple attention variant that applies SiLU only to V before aggregation without introducing additional trainable parameters. We evaluate GatedV against a standard decoder-only Transformer using matched parameter counts, paired initializations, and equal training-token budgets. At 2.33M parameters, GatedV reduces mean held-out perplexity from 59.1 to 57.9 (2.0%), with improvements across all five paired seeds. At 26.16M parameters, following an equal-budget learning-rate search, mean held-out perplexity falls from 20.2 to 19.7 (2.4%), with improvements across all three paired seeds. Small-model ablations show that the tested affine rescaling of values does not fully reproduce GatedV's gains. A short interleaved benchmark on NVIDIA DGX Spark measures approximately 0.7% higher full-context forward latency for the larger model, based on synchronized wall-clock timings. These findings suggest that value-only self-gating can improve language modeling performance without increasing parameter count or redesigning the attention-weighting mechanism. ## On Sparse Graph Limits for Neural Combinatorial Optimization DevFeed: [On Sparse Graph Limits for Neural Combinatorial Optimization]() Original publisher: [Read original article]() Author: Thien Le, Melanie Weber Published: 2026-11-03T05:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [Combinatorial optimization](), [graph theory](), [Neural Network](), [Artificial Intelligence]() Tags: [combinatorial-optimization](), [cs-lg](), [experiments](), [features](), [graph](), [graph-neural-networks](), [guarantees](), [math-co](), [network](), [neural](), [optimization](), [random](), [values]() ### AI overview The paper establishes size-transfer guarantees for normalized graph optimization values using finite energy features. For affine objectives such as MAX-CUT, coloring defect, and independent-set density, value differences are controlled by discrepancies in normalized scalar energy probes on bounded-degree graphs. General Lipschitz objectives can incur an additional representation error: it vanishes on hyperfinite graph families but may stay bounded away from zero on expanders. The paper also gives prediction guarantees under conditions on encoder accuracy, readout stability, and representation coverage, plus statistical generalization bounds. It proves Shapley-Folkman bounds and constructs the energy features with a fixed-weight message-passing network using random node initialization on bounded-treewidth graphs. Experiments with exact solvers and trained models test the guarantees. ### Source excerpt arXiv:2611.01069v1 Announce Type: new Abstract: Graph neural networks are often trained on small graphs and deployed on larger ones, raising the question of which graph information must be preserved for combinatorial optimization values to transfer across sizes. We establish size-transfer guarantees for normalized graph optimization values using finite collections of energy features. For affine objectives, including MAX-CUT, coloring defect, and independent-set density, differences in task value are controlled by discrepancies in normalized scalar energy probes on all bounded-degree graphs. For general Lipschitz objectives, such as those with balance penalties, an additional representation error appears; this error vanishes on hyperfinite graph families but can remain bounded away from zero on expanders. Thus, at fixed degree bound, state space, Lipschitz constant, and target accuracy, a finite energy representation independent of graph size suffices universally for affine tasks and asymptotically for general Lipschitz tasks on hyperfinite families. These results yield prediction guarantees under explicit conditions on encoder accuracy, readout stability, and representation coverage, together with statistical generalization bounds. Theoretically, we prove Shapley-Folkman bounds with sharp parameter dependence, including a dimension-free $\ell_2$ regime, and realize the resulting energy features with a fixed-weight message-passing network using random node initialization on bounded-treewidth graphs. Experiments with exact solvers and trained models test the resulting transfer guarantees. ## LEANGUARD: Lean-Verified Guardrails for Tool-Using Agents DevFeed: [LEANGUARD: Lean-Verified Guardrails for Tool-Using Agents]() Original publisher: [Read original article]() Author: Sreehari Sankar, Debargha Ganguly, Vikash Singh, Mohsen Hariri, Weicong Chen, Yanyan Zhang, Shouren Wang, Wang Yang, Chaoda Song, Alexander Nemecek, Osama Zafar, Idan Davidovich, Yuting Shao, Kai Ye, Published: 2026-11-03T05:00:00Z Content type: article Language: en Sources: [arXiv: Artificial Intelligence]() Topics: [Lean](), [AI Agent](), [Agent Harness](), [rule-engine](), [Dynamic authorization rules]() Tags: [agents](), [bench](), [cs-ai](), [cs-cl](), [evaluation](), [execution](), [feedback](), [guardrails](), [history](), [lean](), [metaprogramming](), [policy](), [prompt-injection](), [qwen](), [repair](), [telecom](), [values]() ### AI overview LeanGuard checks each proposed tool call against a Lean4 policy before it reaches the environment. Policies combine state predicates and scoped past-time rules over execution history; checks fail closed, and the decision procedure is proved equivalent to LeanLTL semantics. Denials can return the violated rule and a hint, or use Lean metaprogramming to find a nearby admissible call that preserves intent without relying on the model. In tests with five Qwen models on Tau-Bench, unguarded agents made at least 33 policy-forbidden calls and LeanGuard made none; feedback did not significantly change task success. On DTap, direct prompt-injection success fell from 57.9% to 5.9%. In the telecom setting, where benign tasks required no writes, benign success was unchanged. Feedback reduced repeated blocked calls and improved task success for four of five models versus bare denial. On replayed argument-value denials, repair proved an admissible edit for 88 of 252 cases, with a reported median of 2 ms. ### Source excerpt arXiv:2611.02332v1 Announce Type: new Abstract: Tool-using language-model agents often execute actions that their domain policies forbid. We present LeanGuard, a runtime that checks every proposed tool call against a policy written in Lean4 before the call reaches the environment. Policies combine state predicates with scoped past-time temporal operators over the execution history, evaluation is fail-closed, and the decision procedure is proved equivalent to a LeanLTL semantics. A denied call can be corrected in two ways. With feedback, the agent receives the violated rule and a hint; with proof-guided repair, Lean metaprogramming searches for a nearby call that preserves the user's intent and proves that the policy admits it, without the model. On Tau-Bench with five Qwen models (0.8B-27B), unguarded agents execute at least 33 calls that the Lean policy forbids, while LeanGuard executes none; with feedback, task success does not change significantly. On DTap, LeanGuard reduces direct prompt-injection success from 57.9% to 5.9%, and in telecom, where benign tasks need no writes, benign success is unchanged. Feedback reduces repeated blocked calls and raises task success for four of five models over a bare denial. On replayed denials, repair proves an admissible edit for 88 of 252 denials caused by argument values, in a median of 2,ms. ## Whose Values, Which Contexts, and What Signals? Grounding Cultural Value Alignment with Re3Culture DevFeed: [Whose Values, Which Contexts, and What Signals? Grounding Cultural Value Alignment with Re3Culture]() Original publisher: [Read original article]() Author: Jing Yao, Hyunjin Hwang, Weihua Zheng, Bryan Chen Zhengyu Tan, Jia Wang Peh, Dylan Raharja, JinYeong Bak, Roy Ka-Wei Lee, Rui Cao, Yuki Arase, Jacki O'Neill, Zhicheng Dou, Xing Xie, Xiaoyuan Yi, Furu Published: 2026-11-02T05:00:00Z Content type: article Language: en Sources: [arXiv: Artificial Intelligence]() Topics: [ai alignment]() Tags: [alignment](), [communities](), [coverage](), [cs-ai](), [data](), [dataset](), [grounding](), [human-data](), [llm](), [target](), [values]() ### AI overview Re3Culture argues that cultural value alignment targets need to be grounded in representative human evidence, observed in concrete community-guided situations, and elicited as explicit priorities. Its dataset links 700+ participants across four cultures, culturally sensitive scenarios, 1,400+ value surveys, 20,000+ behavioral principles, and responses to 6,000+ situated questions. The reported analyses find that substituting cultures or narrowing subgroup coverage shifts estimates, while LLM simulations overrepresent some subgroups. Abstract questionnaire profiles weakly recover situated priorities, and communities differ in which scenarios reveal values. Response preferences and final answers alone are also unreliable signals: they mix value agreement with other response qualities, and similar answers can reflect different priorities. ### Source excerpt arXiv:2610.41793v1 Announce Type: new Abstract: Cultural value alignment requires not only model optimization but also a well-specified target grounded in real-human evidence. Yet how such targets should be grounded in human data remains largely unexplored, raising three fundamental questions: (1) whose values should be represented? (2) where should cultural values be observed? and (3) what signals reveal underlying values? We investigate these questions through Re3Culture, a large-scale dataset linking (i) Representative communities: 700+ human participants across four cultures recruited through cultural-affiliation screening and demographic-aware sampling; (ii) Realistic scenario: culturally sensitive scenarios selected by community members, e.g., family and kinship; and (iii) Revealed value signals: 1,400+ value surveys, 20,000+ topic-grounded behavioral principles and explicit value priorities with responses for 6,000+ situated questions. Our analysis shows why these choices matter for defining the target, with three implications. First, estimating cultural values requires both community affiliation and broad internal coverage. Substituting other cultures or restricting subgroup coverage shifts value estimates. LLM simulations also disproportionately reflect particular subgroups. Second, values must be observed in concrete, community-guided situations. Abstract questionnaire profiles weakly recover situated priorities, and communities differ in which scenarios they identify as culturally value-revealing. Third, value priorities should be elicited explicitly, not learned from response preferences or bottom-line answers alone. Preferences conflate value agreement with other response-level qualities, and similar answers can be derived from different value priorities. By addressing these requirements, Re3Culture provides the first human-grounded resource that explicitly informs target definition for cultural value alignment. ## CVE-2026-62334 - OpenVM's OpenVmHalo2Verifier self-call can bypass proof verification DevFeed: [CVE-2026-62334 - OpenVM's OpenVmHalo2Verifier self-call can bypass proof verification]() Original publisher: [Read original article]() Author: CVEFEED.io; github.com Published: 2026-10-30T20:17:12Z Content type: news Language: en Sources: [Latest High/Critical Vulnerabilitiy Feed]() Topics: [Zero-knowledge proof](), [cve remediation]() Tags: [1-7](), [2-0-0](), [2026](), [bypass](), [cve-2026-62334](), [dispatch](), [exposed](), [fallback](), [fixed](), [function](), [openvm](), [raw](), [selector](), [solidity](), [treats](), [user](), [using](), [values](), [verification](), [verify]() ### AI overview Before OpenVM 1.7.0 in the 1.x line and in affected 2.0 and 2.1 prereleases, `OpenVmHalo2Verifier.verify` builds calldata from user-controlled `proofData`, then uses a `staticcall` to its own wrapper and accepts any successful call. Crafted first four bytes can dispatch to another successful external function instead of the Halo2 verifier, allowing applications using the generated Solidity wrapper to accept arbitrary app commitments and public values without validating the proof suffix, accumulator, commitments, or public values. Fixed in 1.7.0, 2.0.0-rc.3, and 2.1.0-preview. ### Source excerpt CVE ID :CVE-2026-62334 Published : Oct. 30, 2026, 8:17 p.m. | 1 hour, 37 minutes ago Description :OpenVM is a performant and modular zkVM framework built for customization and extensibility. Before 1.7.0 in the 1.x line and in affected 2.0 and 2.1 prereleases, `OpenVmHalo2Verifier.verify` constructs raw proof calldata from user-controlled `proofData`, performs a `staticcall` to `address(this)`, and treats a successful call as proof acceptance. If the first four proof bytes match another external selector exposed by the wrapper, dispatch can invoke a successful non-verifier function instead of the generated Halo2 fallback verifier. Applications using the generated Solidity wrapper can consequently accept arbitrary app commitments and public values without checking the proof suffix, accumulator, app commitments or public values. This issue is fixed in versions 1.7.0, 2.0.0-rc.3 and 2.1.0-preview. Severity: 8.7 | HIGH Visit the link for more details, such as CVSS details, affected products, timeline, and more... ## doomformer: neural networks computed inside DOOM maps DevFeed: [doomformer: neural networks computed inside DOOM maps]() Original publisher: [Read original article]() Author: Anne Barela Published: 2026-10-30T14:47:49Z Content type: article Language: en Sources: [Adafruit Industries]() Topics: [Neural Network]() Tags: [arithmetic](), [artificial-intelligence](), [doom](), [doomformer](), [game](), [gaming](), [github](), [light](), [map](), [model](), [museum](), [neural-networks](), [numpy](), [values]() ### AI overview doomformer compiles integer neural networks into DOOM maps, using voodoo dolls for arithmetic and map light levels and plane heights to store values; exposed values are checked against NumPy. One map runs an MNIST network with 1,355 dolls: the player enters a digit on the floor, presses READ, and the winning digit's doll lights the display. The float model scores 97.58 percent. A separate museum.wad runs without Python; its 136-doll network reads seven floor pads, and MUTATE rewires one connection while running, changing a 7's result to 1. A third map models a marine's brain with 1,345 dolls sensing the nearest demon and choosing actions. It is not playable: each decision takes 22 seconds of game time, while the marine dies in 5.5 seconds without a kill. ### Source excerpt doomformer is a project from alp (@mottopanikeiku) that compiles an integer neural network into a DOOM map. Voodoo dolls do the arithmetic. Light levels and plane heights hold the values. Every value the map exposes is read back and compared against NumPy. One map puts 1,355 voodoo dolls to work on an MNIST network. The [...] ## A Utility-Mining-Driven Active Learning Approach for Analyzing Clickstream Sequences DevFeed: [A Utility-Mining-Driven Active Learning Approach for Analyzing Clickstream Sequences]() Original publisher: [Read original article]() Author: Danny Y. C. Wang, Lars Arne Jordanger, Jerry Chun-Wei Lin Published: 2026-10-30T04:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [Predictive AI](), [Data science and engineering](), [Data Mining & Modeling]() Tags: [behaviors](), [commerce](), [cost](), [cs-lg](), [data-processing](), [experiments](), [high-quality-data](), [labeling](), [learning](), [machine-learning](), [mining](), [model-training](), [modeling](), [performance](), [prediction](), [sequences](), [settings](), [shap](), [strategy](), [utility](), [values]() ### AI overview HUSPM-SHAP uses SHAP values with high-utility sequential pattern mining to select e-commerce clickstream data for model training, aiming to reduce labeling needs while retaining predictive performance for purchase and non-purchase behavior. The abstract reports strong results across diverse scenarios and notes that settings for positive and negative SHAP values affect which patterns are mined. It does not provide experiment details or quantify the labeling reduction. ### Source excerpt arXiv:2410.07282v2 Announce Type: replace Abstract: In rapidly evolving e-commerce industry, the capability of selecting high-quality data for model training is essential. This study introduces the High-Utility Sequential Pattern Mining using SHAP values (HUSPM-SHAP) model, a utility mining-based active learning strategy to tackle this challenge. We found that the parameter settings for positive and negative SHAP values impact the model's mining outcomes, introducing a key consideration into the active learning framework. Through extensive experiments aimed at predicting behaviors that do lead to purchases or not, the designed HUSPM-SHAP model demonstrates its superiority across diverse scenarios. The model's ability to mitigate labeling needs while maintaining high predictive performance is highlighted. Our findings demonstrate the model's capability to refine e-commerce data processing, steering towards more streamlined, cost-effective prediction modeling. ## An updated Model Spec for more capable and beneficial AI DevFeed: [An updated Model Spec for more capable and beneficial AI]() Original publisher: [Read original article]() Published: 2026-10-29T17:00:00Z Content type: release Language: en Sources: [OpenAI News]() Topics: [ai alignment](), [Artificial Intelligence]() Tags: [about](), [accounting](), [actions](), [added](), [aggregate](), [ai](), [alignment](), [all](), [also](), [applications](), [apply](), [approach](), [artificial-general-intelligence](), [core](), [guide](), [model](), [openai](), [safety](), [spec](), [updated](), [values]() ### AI overview The updated Model Spec organizes model behavior around five values: faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care. It says models should interpret instructions in context, make assumptions visible and correctable, account for effects on others, and support people's ability to understand and choose. The Spec combines firm safety boundaries with defaults that users and developers can usually override. New guidance covers context, oversight, agency, social norms, side effects, and an updated chain of command; most detailed style rules are retired, while objectivity remains a stronger default. OpenAI says the Spec describes intended behavior that models do not yet fully meet and will continue to evolve through research, deployment experience, and feedback. ### Source excerpt OpenAI's updated Model Spec outlines five core values to guide more capable AI toward sound judgment, human agency, and faithfulness to human intent. ## Creating a Typescript union, while also having runtime access to the values DevFeed: [Creating a Typescript union, while also having runtime access to the values]() Original publisher: [Read original article]() Author: Jamie Tanna Published: 2026-10-28T12:17:12Z Content type: article Language: en Sources: [Jamie Tanna]() Topics: [union types](), [data type]() Tags: [blogumentation](), [documentation](), [example](), [forgejo](), [function](), [git](), [github](), [gitlab](), [hosting](), [important](), [parameter](), [pattern](), [platform](), [platforms](), [reference](), [renovate](), [runtime](), [typescript](), [union](), [values]() ### AI overview A TypeScript string union constrains values at compile time, but does not provide a runtime list to iterate over or use when generating documentation. The excerpt shows defining the values once in an array with `as const`, then deriving the union type as `(typeof Platforms)[number]`. This makes the values available at runtime while retaining the `Platform` type for function parameters. The example uses `github`, `gitlab` and `forgejo`; the excerpt does not elaborate on other approaches. ### Source excerpt I very regularly end up writing this pattern - primarily in Renovate's code - so thought I'd write it up as a form of blogumentation, so I can refer back to it in the future. (This example uses concepts from the Renovate codebase, but does not take any code from it) For instance, let's say that you have a known union of big Git-hosting platforms - github, gitlab and forgejo. If you wanted to make sure that a given function parameter could only include them, you'd write i.e. export type Platform = 'github' | 'gitlab' | 'forgejo'; Normally this would be sufficient for usage, and nothing else is needed. But if you then want to be able to reference the ?? - for instance, as part of generating documentation - then you need some way of accessing each of these values at runtime. This leads us to: export const Platforms = [ 'github', 'gitlab', 'forgejo', // no shade to other platforms ] as const; // `as const` is important! export type Platform = (typeof Platforms)[number]; Now, you can access Platforms at runtime (and loop through them), while also enforcing the use of Platform as a function parameter. ## Can AI Understand You Better Than You Understand Yourself? DevFeed: [Can AI Understand You Better Than You Understand Yourself?]() Original publisher: [Read original article]() Author: Hanul Lee (이하늘) Published: 2026-10-28T07:48:34Z Content type: article Language: en Sources: [HackerNoon]() Topics: [knowledge-engineering]() Tags: [actions](), [ai](), [ai-agents](), [ai-and-human-identity](), [ai-assisted](), [ai-knowledge-management](), [ai-self-awareness](), [artificial-intelligence](), [brain](), [digital-self](), [essay](), [feedback](), [layers](), [learning](), [memory](), [second-brain-technology](), [self-model-fidelity](), [understand](), [values]() ### AI overview The author argues that a second brain's fidelity to a person depends on more than stored facts: it must also reflect what they consider important, how they interpret events, the conclusions they draw, and how they revise beliefs after mistakes. These higher layers are harder to transfer from records alone. Self-description can conceal a gap between stated principles and recorded behavior, while automatically accumulated records can expose it. In the author's example, noticing that he repeatedly reclaimed assignments after a third mistake led him to rewrite a principle in the SSOT; the person and digital self then informed each other through a feedback loop. The author presents fidelity and "loss rate of the self-model" as concepts for describing this experience, not validated measures; the diagrams are conceptual, not performance scores. ### Source excerpt Can an AI second brain truly understand you? Explore five layers of fidelity, from memories and values to decisions and self-correction. ## OpenTofu 1.13 Lets Module Authors Reduce Unknown Values at Plan Time DevFeed: [OpenTofu 1.13 Lets Module Authors Reduce Unknown Values at Plan Time]() Original publisher: [Read original article]() Author: Mark Silvester Published: 2026-10-26T21:26:00Z Content type: news Language: en Sources: [InfoQ]() Topics: [opentofu](), [data type](), [Null-Safety](), [modules](), [versioning]() Tags: [architecture-design](), [cloud-native-computing-foundation](), [development](), [devops](), [error](), [functions](), [infrastructure-as-code](), [module](), [news](), [opentofu](), [opentofu-1-13-unknown-values](), [plan](), [terraform](), [values]() ### AI overview OpenTofu 1.13 adds `convert` and `assumenotnull` so module authors can give the planner type and non-null guarantees for otherwise unknown values. This can resolve `count` or `for_each` instance decisions while resource attributes remain unknown until apply. Authors must ensure the values satisfy those guarantees; violations are reported during apply. Terraform shipped the equivalent `convert` function first, and OpenTofu chose to match its design. The release also introduces opt-in experimental lint rules and Symbol Libraries, which cannot directly access resources, data sources or provider functions. It removes WinRM support from `remote-exec` and `file`; changes to `base64gzip` may trigger resource updates despite equivalent decompressed output. The excerpt also reports Windows ARM64 packages, a macOS 13 minimum, and 1.13.2 security fixes. ### Source excerpt OpenTofu 1.13 adds convert and assume... functions that let module authors give the planner type and value hints for unknown values, so more expressions resolve during tofu plan. If a value unknown during planning later violates a guarantee, OpenTofu reports an error during apply. By Mark Silvester ## Sorting a Python Dictionary: Values, Keys, and More DevFeed: [Sorting a Python Dictionary: Values, Keys, and More]() Original publisher: [Read original article]() Author: Ian Currie Published: 2026-10-25T14:00:00Z Content type: tutorial Language: en Sources: [Real Python]() Topics: [Python](), [Sorting](), [ordering]() Tags: [2](), [22](), [3](), [about](), [absolute](), [code](), [compare](), [fast](), [how-to](), [keep](), [key](), [keys](), [learn](), [nested](), [python](), [sort](), [sorting](), [techniques](), [value](), [values]() ### AI overview Python's `sorted()` returns a new list, so sorting dictionary entries typically means sorting `dict.items()` and optionally rebuilding a dictionary with `dict()`. Default tuple ordering sorts by key; pass a `key` callback such as `lambda item: item[1]` or `operator.itemgetter(1)` to sort by value, and use `reverse=True` for descending order. For nested fields, a named function can apply defaults with `.get()` when values may be missing. Dictionaries cannot be sorted in place. The text reports `itemgetter()` outperforming a lambda in one benchmark, and a list sorting faster than sorting and rebuilding a dictionary in another; dictionary lookups were much faster than linear list searches in its sample. It cautions that timings vary by data and system. For repeated sorting, a list may fit better; for frequent lookups, a dictionary is faster. Ordered dictionary iteration also does not ensure custom key order will survive JSON or cross-language use. ### Source excerpt Learn how to sort a Python dictionary by value or key, handle nested attributes, and compare techniques to keep your code fast and readable. ## How to fix Webflow API not returning all Collection fields DevFeed: [How to fix Webflow API not returning all Collection fields]() Original publisher: [Read original article]() Author: Ismail Ajagbe Published: 2026-10-24T00:00:00Z Content type: tutorial Language: en Sources: [Webflow Blog]() Topics: [webflow](), [API](), [client library]() Tags: [2](), [22](), [3](), [adapter](), [added](), [all](), [also](), [api](), [api-s](), [app](), [apps](), [array](), [async](), [audit](), [cms](), [collection](), [fields](), [guides](), [live](), [values](), [webflow]() ### AI overview Missing Webflow CMS fields usually mean an item has no stored value, or code is using a field's display name instead of its slug. The guide proposes auditing the collection schema against staged and live items, then normalizing each item so every schema slug is present, using null for absent values. It distinguishes staged items from published-only live results and notes that Content Delivery API responses can remain cached for up to five minutes (120 seconds on Enterprise plans). Other causes covered include locale selection and fields whose values are IDs or HTML. The implementation uses a server-side Next.js Route Handler on Webflow Cloud, a CMS-read site token, and pagination; it requires a Premium Site plan or higher, Next.js 15+, and Node.js 22+. ### Source excerpt Webflow API missing CMS fields? Learn why empty values drop out of fieldData and how to audit staged and live items. ## LHB Linux Digest #26.16: Bash Arrays, mapfile, Self-hosted Linear Alternative and More DevFeed: [LHB Linux Digest #26.16: Bash Arrays, mapfile, Self-hosted Linear Alternative and More]() Original publisher: [Read original article]() Author: Abhishek Prakash Published: 2026-10-23T12:39:19Z Content type: tutorial Language: en Sources: [Linux Handbook]() Topics: [Bash](), [Self-hosted](), [Linear](), [Tool](), [Tutorial]() Tags: [1-4](), [2](), [22](), [about](), [account](), [added](), [agents](), [ai](), [ai-agents](), [ai-tools](), [all](), [also](), [analysis](), [arrays](), [bash](), [linear](), [linux](), [newsletter](), [self-hosted](), [values](), [variable]() ### AI overview The Bash section explains indexed and associative arrays, including zero-based indexing, element counts, appending, slices, and how quoted "${array[@]}" preserves items containing spaces while "${array[*]}" joins them into one string. It shows mapfile -t with process substitution to load find results into an array. The excerpt also describes itsaplan as a self-hosted, Linear-style project management platform for teams working alongside AI agents, and lazyenv as a terminal interface for managing .env files. It warns that a breach affecting about 1.4% of OpenCode Console users may have exposed third-party API keys, and reports that GitHub Enterprise Cloud can enforce open-source license compliance during pull requests. The certification discount and sale deadline are time-sensitive claims in the newsletter. ### Source excerpt One variable, many values. ## Adaptive Pluralistic Alignment: A pipeline for dynamic artificial democracy DevFeed: [Adaptive Pluralistic Alignment: A pipeline for dynamic artificial democracy]() Original publisher: [Read original article]() Author: Rachel Freedman Published: 2026-10-22T04:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [ai alignment](), [Model Development](), [Artificial Intelligence]() Tags: [adaptive](), [ai](), [alignment](), [arxiv](), [cs-lg](), [data-collection](), [democracy](), [implement](), [machine-learning](), [pipeline](), [proof-of-concept](), [reward](), [shift](), [sliding-window](), [social](), [tracking](), [updating](), [value](), [values](), [voting]() ### AI overview Adaptive Pluralistic Alignment (APA) proposes updating AI alignment as societal values shift without repeating pretraining or large-scale data collection. It learns compact personalized reward models from low-rank reward bases, uses them as a jury to select candidate outputs through social-choice voting, then adapts the jury by fitting new annotator weights over the same bases. A proof of concept with simulated historical annotators finds that jury composition and voting rules can substantially affect outcomes, and bounds value-tracking error for a sliding-window jury. The evidence describes a proof of concept, not validation on real-world annotators. ### Source excerpt arXiv:2605.01642v3 Announce Type: replace Abstract: Prevailing alignment methods target a fixed set of preferences and therefore risk forcing value lock-in as societal norms evolve over time. We introduce Adaptive Pluralistic Alignment (APA), a modular pipeline for updating pluralistically aligned AI systems to track evolving values and avoid value lock-in without repeating costly pretraining or large-scale data collection. APA has three stages: (1) learning compact personalized reward models via low-rank reward basis decomposition, (2) using these models as a jury that collectively selects among candidate outputs through social-choice-theoretic voting, and (3) efficiently adapting the jury over time by fitting new annotator weights over the fixed reward bases as values shift. The resulting system is efficient, explainable, steerable, and modular. We implement a proof-of-concept instantiation using simulated historical annotators, demonstrate how jury composition and the choice of voting rule can substantially affect outcomes, and bound value tracking error given a sliding-window jury. We provide full code at https://github.com/RachelFreedman/apa. ## Unintended Side Effects of Value Training in LLMs DevFeed: [Unintended Side Effects of Value Training in LLMs]() Original publisher: [Read original article]() Author: Siddhesh Pawar, Chen Cecilia Liu, Anna Korhonen, Isabelle Augenstein Published: 2026-10-22T04:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [post-training](), [ai alignment](), [Fine-tuning](), [ai safety]() Tags: [alignment](), [behaviour](), [capabilities](), [cross](), [cs-ai](), [cs-cl](), [cs-lg](), [dimensions](), [effects](), [evaluations](), [fine-tuning](), [llms](), [nlp](), [pipeline](), [post-training](), [risk](), [safety](), [social](), [training](), [unrelated](), [value](), [values]() ### AI overview The study reports that fine-tuning four instruction-tuned LLMs on individual Hofstede value dimensions can affect more than the targeted values. The authors evaluated cross-value transfer and broader behavior, finding side effects on untrained values, risk preferences, social cognition, and safety traits. They recommend broader behavioral evaluations after value alignment and further investigation of where value training fits in the post-training pipeline. ### Source excerpt arXiv:2610.29922v1 Announce Type: cross Abstract: As LLMs are increasingly deployed in value-sensitive contexts, aligning them with human values has emerged as a post-training objective in NLP, aimed at instilling additional values and guiding model behaviour. However, current evaluations remain narrow. Because values are interconnected, it is unclear how fine-tuning for specific values alters unrelated capabilities and behaviours. In this work, we adapt four instruction-tuned LLMs across individual Hofstede's value dimensions and evaluate them on both cross-value transfer and broader behaviours. We find that value-specific training produces unintended side effects on untrained values, risk preferences, social cognition, and safety traits. Consequently, we suggest that value alignment after post-training should incorporate broader behavioral evaluations, and we highlight the need to investigate where additional value training might best fit into the pipeline. ## Super Second Brain: What Happens When AI Knows You Better Than You Know Yourself? DevFeed: [Super Second Brain: What Happens When AI Knows You Better Than You Know Yourself?]() Original publisher: [Read original article]() Author: Hanul Lee (이하늘) Published: 2026-10-21T08:13:32Z Content type: opinion Language: en Sources: [HackerNoon]() Topics: [Multi Agent Systems](), [AI Context Management]() Tags: [agents](), [ai](), [ai-agents](), [ai-knowledge-repository](), [ai-thought-experiment](), [artificial-intelligence](), [brain](), [digital-identity](), [digital-self](), [human-ai-collaboration](), [memory](), [philosophy](), [question](), [super-second-brain](), [values]() ### AI overview The thought experiment begins with separate work agents holding different versions of their user, then gives them access to a shared cloud repository of life records, values, and judgments while preserving each agent's distinct perspective and memory. As agents and possible robots contribute experiences to that repository, the "Super Second Brain" could remember more, encounter more of the world, and apply the person's values more consistently than the person can. The author asks whether the person remains the original or becomes a figurehead, and whether responsibility or the ability to stop recording is enough to preserve agency. He argues that responsibility still matters legally and socially, but delegated action can make it insufficient as a personal answer. This is explicitly a thought experiment, not a forecast or a claim about any product's performance. ### Source excerpt A shared memory for AI agents led me to a harder question: if an AI lives by my values more consistently than I do, what remains of my agency? ## Composable Parametric Demographic Injection for Values Survey Simulation DevFeed: [Composable Parametric Demographic Injection for Values Survey Simulation]() Original publisher: [Read original article]() Author: Xinyuan Cao, Qingyi Pan, Bangde Du, Yujia Zhou, Hexi Wang, Weihang Su, Yiqun Liu, Yueyue Wu, Min Zhang, Qingyao Ai Published: 2026-10-21T04:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [lora](), [llm-reasoning]() Tags: [alignment](), [composition](), [cs-ai](), [cs-cl](), [cs-lg](), [distribution](), [distributions](), [inference](), [language-models](), [large-language-models-llms](), [llms](), [lora](), [lora-module](), [models](), [modules](), [overhead](), [performance](), [prompt](), [prompting](), [reduce](), [simulation](), [survey](), [values]() ### AI overview PARADE builds simulated survey respondents by combining reusable LoRA modules for demographic attribute values with a shared Task LoRA for survey capability. Composition calibration is intended to reduce interference, and modules can be recombined at inference time for demographic combinations not seen in training. In experiments on the World Values Survey and SocioBench using three backbone models, PARADE reportedly aligned population-level response distributions while maintaining competitive predictive performance, reducing KL divergence by up to 52.8% versus the strongest baseline. The supplied text does not specify detailed experimental settings or limitations. ### Source excerpt arXiv:2610.23867v1 Announce Type: cross Abstract: Large language models (LLMs) are increasingly used as simulated respondents in computational social science, yet accurately reproducing population-specific heterogeneous response distributions remains challenging. Existing methods such as demographic prompting are sensitive to prompt design and increase inference overhead, while parameter-based methods struggle to reuse demographic information efficiently, limiting generalization to unseen combinations. In this regard, we propose PARADE, a composable PARAmetric DEmographic injection framework that represents each demographic attribute value with a reusable and composable LoRA module. Specifically, PARADE separates task-common survey capability from demographic-specific value shifts via a shared Task LoRA and attribute-level Demographic LoRAs, with composition calibration to reduce parameter interference. At inference time, target population models can be constructed on demand by recombining the relevant modules, including demographic combinations not observed during training. Extensive experiments on the World Values Survey and SocioBench across three backbone models show that PARADE achieves strong population-level distribution alignment while maintaining competitive predictive performance, reducing KL divergence by up to 52.8% relative to the strongest baseline. [Next page]()