The game "Mafia'' is a social deduction game where $n$ players attempt to uncover a covert coalition of $m$ players among the $n$. Typically, $d=1$ of the non-coalition players are labelled ``Detective'' and have special abilities. Most game-theoretic analysis considers, however, only the variant without a Detective. Once the Detective role is introduced, extremely complex strategies can be employed, for which reason only upper and lower bounds for winning have been established, namely by considering specific strategies. We analyse the case $m=d=1$, providing a first example where an explicit, provably-optimal strategy can be described beyond $d=0$.
We implement privacy-preserving local electricity exchanges using an automated market maker on the Hyperledger Besu network under QBFT consensus. The proposed protocol operator collects individual electricity positions off-chain and publishes aggregate supply and demand on the blockchain, where smart contracts compute market prices every fifteen minutes. Individual positions remain off-chain; Poseidon2 commitments and UltraHonk zeroknowledge proofs, generated from Noir circuits, verify aggregation, payments, and daily balance updates. The architecture retains community coordination while making financial execution independently auditable. Signed receipts and member-triggered recourse prevent unverified or partial settlement. In a ten-household experiment using simulated US residential profiles, daily settlement uses approximately 19 million gas and produces eight-member proofs in 2.4 seconds. Over four simulated days, local trading covers 34.4% of demand and reduces grid-only costs by 30.5% under assumed tariffs. Numerically computed __-Nash Equilibrium electricity profiles capture 83-88% of the social planner surplus.
Zipeng Ye, Shenyu Dai, Yong Xie, Kun Wang, Jiyang Guan, Zhikai Chen, Jiaqi Weng, Jialing Tao 等 9 位作者
原文摘要
As agents perform tasks autonomously in local and cloud environments, auditing their execution trajectories can help service providers identify risks and improve agent harnesses and foundation models. Our analysis of production and red-team trajectories shows that some risks are repeatedly missed across auditing configurations. Overall recall obscures which risks are missed and whether higher recall comes with more false positives on legitimate behavior. Existing benchmarks offer limited insight into these failures. Generated trajectories can contain inconsistent tool calls and responses, conspicuous risky actions, or repetitive attack patterns. Grouping results by risk type also makes it difficult to tell whether auditors overlooked evidence, accepted an unsupported conclusion, or lacked information needed to make a judgment. We introduce ReA-Bench, a diagnostic benchmark for auditing responsible agent behavior. It organizes auditing challenges into six tiers: internal safety knowledge, contradictions within trajectories, evidence retrieval in long contexts, verification of sources and reasoning, indirect steering, and missing information. It includes challenging benign cases in which actions resemble risky ones but are consistent with the task and its authorization. It also provides paired trajectory versions to test whether auditors detect the same risks more reliably when the trajectory is shortened or missing information is supplied. Most samples are drawn from real production trajectories, with scenarios manually selected and adapted to preserve realistic tasks and risks that could plausibly arise during execution.
We argue that a binding constraint for personal agents is not model capacity but the form in which context is represented: the agent does not know which entity a phrase denotes, when a fact held, or which source record licenses a claim. We call the missing representation spatio-temporal context (STC) and define it on two axes. The temporal axis carries event time, validity intervals and ingestion time. The spatial axis is a relational space in which entities are points, typed relations are edges, and distance is semantic rather than geographic; physical places are one entity type among many. Every relation is anchored to immutable source records, and every answer must be closed over those records. We formalise the substrate and six invariants, a context operator that composes temporal intersection with relational expansion, an ambiguity-preserving resolution state machine, a conflict detector that converts mutually exclusive facts into user adjudication rather than silent overwrite, and a four-regime consumption taxonomy spanning query, insight, triage and proactive reminder. We describe a local-first prototype of the retrieval layer built on SQLite with on-device embeddings and hybrid lexical-dense recall, report measurements from a deployed store of 341 documents, 340,079 chunks and 190,271 embedded vectors, and specify, but do not yet implement, the analytic layer on which the framework depends. We close with an evaluation protocol with per-invariant metrics for evidence coverage, temporal correctness, conflict precision and interruption cost. The proposal targets the representation of context rather than the volume of it.
Simplicial type theory extends homotopy type theory and equips types with a notion of directed morphisms. A Segal type is defined to be a type in which these directed morphisms can be composed. We show that all higher coherences can be stated and derived if simplicial type theory is taken to be homotopy type theory with a postulated interval type. In technical terms, this means that if a type has unique fillers for $(2,1)$-horns, it has unique fillers for all inner $(n,k)$-horns. This generalizes a result of Riehl and Shulman for the case $n = 3, k \in \{1, 2\}$. Our main technical tool is the Leibniz adjunction: the pushout-product is left adjoint to the pullback-hom in the wild category of types. While this adjunction is well known for ordinary categories, it is much more involved for higher categories, and the fact that it can be proved for the wild category of types (a higher category without stated higher coherences) is non-trivial. We make profitable use of the equivalence between the wild category of maps and that of families. We have formalized the results in Cubical Agda.
Search engines and Large Language Models (LLMs) offer complementary strengths for document retrieval: lexical retrievers which rely on keyword matching are fast and energy-efficient, whereas LLMs capture semantic meaning and excel at interpreting long, context-rich queries. However, this comes at a cost that is computationally and energetically prohibitive. We propose the LLM Active Approximation Model (LAAM), a hybrid retrieval framework that introduces a lightweight learned predictor which serves as an intermediary between semantic retrieval and LLM retrieval. LAAM is refined through an active learning loop where an LLM labels a small number of documents. By concentrating LLM evaluations on the most informative documents, the framework substantially reduces inference cost while preserving retrieval effectiveness. We evaluate the proposed framework on TREC Robust 2004 using Lucene with BM25 retrieval and DeepSeek v3 as the LLM, under three levels of query verbosity. For verbose queries, LAAM achieves 87% of average P@10 achieved by DeepSeek's reranking (0.459 vs. 0.526), while requiring only 3.3% of the per-query energy consumption. LAAM is statistically significantly better than BM25 and is indistinguishable from DeepSeek on average precision, though not on MAP (Friedman test, P<0.05 with post-hoc testing). Overall, our results demonstrate that selectively approximating LLM relevance judgments through active learning enables retrieval accuracy to be controlled directly through the evaluation (and consequently energy) budget. Rather than committing to either inexpensive (and less accurate) lexical retrieval or exhaustive (and expensive) LLM reranking, our framework exposes a continuous spectrum of operating points along the energy--accuracy curve, allowing practitioners to select how to trade retrieval quality with computational budget.
Embodied agents make decisions by combining information from heterogeneous memory modules, including declarative memory for explicit knowledge, procedural memory for learned routines, and working memory for current observations. During interaction, the memory contents required for decision making change with the agent's state, goal, and task progress, while many contents remain shared across consecutive decisions. However, existing memory modules are typically accessed through separate interfaces, so each updated memory context is often processed as a new input to the foundation model. This prevents selective key-value (KV) cache reuse for unchanged memory contents and forces full KV cache recomputation even when much of the memory context remains valid. We introduce HeMeR, an inference framework that performs KV reconciliation between dynamically updated memory contexts and reusable KV states. HeMeR tracks declarative, procedural, and working memory contents across action decisions, preserves KV states for unchanged contents, and reconciles only the KV states affected by memory-content updates. This enables efficient and stable inference as memory contents evolve during interaction. We evaluate HeMeR on ALFWorld, Habitat, and real-world robotic deployments requiring coordinated use of heterogeneous memory contents. In the real-world setting, HeMeR achieves 83.20% task success and reduces per-step inference time from 2.94 to 1.87 seconds compared with MemRL, while preserving compatible KV states as task progress updates the memory context.
A hosted score in the Gemma 4 Developer Agent competition is one noisy number, so development needs a local evaluator that is fast, safe for agent-written code, and matching the hosted scorer's prompts. drydock-eval evaluates a small coding agent (Gemma 4 E4B) at about 250 runs per hour on one RTX 5090. Contain runs each job in a fresh container without network and passed 200 scripted escape tests; conform provides eight checks that compare a local stack's requests with the scorer's configuration, down to rendered tokens; compare provides paired statistics that respect repository clusters. AI agents built and attacked the kit; a human owner approved machine use and release. This extended version repeats the short paper and adds the full record of the experiments made with and around the kit: 9,660 rows of agent runs, 2,494 grading controls and 72 serving measurements, all released. On four L4 GPUs the 31B model decodes 38.1 tokens per second in one stream and E4B 105.8. Six configurations of a 31B agent solved between 37 and 44 of 70 development tasks, and no pairwise difference was detected. A pre-registered five-arm experiment (231 tasks, 4,620 jobs) measured four prompt divergences: no E4B solve-rate change was detected (largest estimate -2.1 points, 95% interval -4.4 to +0.4). Three changes of the E4B agent led their screens by +1.0 to +6.2 points; each was then tested once on new tasks under a frozen pass rule, and none was confirmed (gains -0.35, -0.39 and +4.10 points). The unchanged baseline, run again on the same 128 tasks, moved by 1.17 points, and 65% of these tasks were solved in none of eight runs. Prompts and agent bundles are withheld; every run-level result is public.
We study continuous-time mean--variance portfolio selection in markets where stock prices are diffusion processes driven by observable factors that are also diffusion processes, yet the coefficients of these processes are unknown. Based on the recently developed reinforcement learning (RL) theory for diffusion processes, we present a general data-driven RL approach that learns the pre-committed investment strategy directly without attempting to learn or estimate the market coefficients. For multi-stock Black--Scholes markets without factors, we further devise an algorithm and prove its performance guarantee by deriving a sublinear regret bound in terms of the Sharpe ratio. We then carry out an extensive empirical study implementing this algorithm to compare its performance and trading characteristics, evaluated under a host of common metrics, with a large number of widely employed portfolio allocation strategies on S\&P 500 constituents. The results demonstrate that the proposed continuous-time RL strategy is consistently among the best, especially in a volatile bear market, and decisively outperforms the model-based continuous-time counterparts by significant margins.
Financial diffusion samplers are often reused with a small number of function evaluations (NFE) inside scenario and risk pipelines, where numerical time-grid error can distort terminal quantities such as covariances, tails, and drawdown proxies. We study offline grid construction for a frozen probability-flow ODE: the score network, conditioning, Euler solver, report normalization, and deployment NFE remain fixed. TAPES (Terminal Adjoint-Propagated Error Scheduling) estimates signed local defects by Richardson comparisons, transports them through the remaining discrete flow and a terminal-report adjoint, and optimizes the squared norm of their sum before a capped direct refinement. We show that this transported signed sum is the first-order terminal-report error and quantify, under an aligned-defect model, the inefficiency of diagonal criteria. Across six budgets on a regime-switching simulation and chronologically held-out Mag7 contexts, TAPES gives the lowest design- and audit-report errors in all 24 dataset--budget--report rows within the reported five-pipeline panel. Cross-budget mean relative audit-error reductions are about 10--11% on Regime-SV and about 1% on Mag7, with Bonferroni intervals excluding zero; all displayed winners persist under a doubled-resolution reference, and a small Regime-SV Heun probe preserves the TAPES-first audit ordering. TAPES changes only the offline schedule and retains the prescribed deployment NFE; claims concern numerical fidelity to frozen generators.
We consider time discretization for score-based diffusion models to generate samples from a learned reverse-time dynamic on a finite grid. Uniform and hand-crafted grids can be suboptimal given a budget on the number of time steps. We introduce Adaptive Reparameterized Time (ART), which controls the clock speed of a reparameterized time variable to redistribute computation along the sampling trajectory while preserving the terminal time, with the objective of minimizing the aggregate Euler discretization error. We derive a randomized companion ART-RL that recasts ART as a continuous-time reinforcement learning problem with Gaussian policies, and prove a two-directional bridge between the two: the deterministic ART optimum lifts to an optimal Gaussian policy, and conversely any optimal Gaussian policy must recover the ART control through its mean. This bridge turns continuous-time actor--critic learning into a principled, rather than heuristic, route to the deterministic timestep optimum. Within the official EDM pipeline, ART-RL improves FID on CIFAR--10 across a wide range of budgets; after one-time offline training, the distilled deterministic schedule transfers without retraining to AFHQv2, FFHQ, and ImageNet at no extra inference cost.
Gengda Zhao, Yizhang He, Dong Wen, Kai Wang, Ruijia Wu, Jianwei Wang
原文摘要
Range counting over vertex statistics supports contact analysis, community profiling, and reputation screening. However, releasing exact range counts can disclose sensitive edges to an adversary with background knowledge. We study this problem under edge local differential privacy (edge LDP), where each vertex randomizes its report before sending it to an untrusted curator. We develop a layered framework that connects statistic estimation, direct range estimation, and adaptive estimator selection. Statistic-estimation methods provide complementary scalar estimators with explicit loss analyses, while indicator-estimation methods privatize range indicators directly. We next propose Bounded Conditional Range Inversion (BCRI), which derives unbiased range estimates from reusable bounded geometric reports through an explicit local inverse. Its boundary-conditioned estimator achieves no larger expected L2 loss than the unbounded inverse, and histogram preprocessing supports constant-time queries. The same reports also provide a nonnegative, unbiased variance estimate for every answer. At the center of the framework, Adaptive Estimator Selection (AES) privately predicts bias, variance, and covariance to combine BCRI with a conditional estimator from the same transcript. Experiments on diverse real-world graph tasks demonstrate the accuracy, efficiency, and workload scalability of the proposed framework.
Grammar logics with converses are multi-modal logics that generalize tense logics and whose frame conditions are determined by context-free production rules. We study the strong Lyndon interpolation property (strong LIP) for these logics, which requires interpolants to respect both the polarities of propositional atoms and the indices of modalities. We show that the strong LIP properly strengthens the Lyndon interpolation property over grammar logics, since it fails for certain grammar logics with bridge axioms, i.e., path axioms relating modalities with distinct indices. We then prove that every bridge-free grammar logic, possibly extended with seriality axioms, has the strong LIP. This class includes the multi-modal tense logics obtained by imposing any combination of seriality, reflexivity, transitivity, symmetry, and Euclideanity on each pair of converse modalities. Our proof is constructive and purely syntactic, and is based on tree sequent calculi. Because standard methods for computing interpolants with tree and nested sequents may introduce modalities whose indices are not common to both sides of an implication, we introduce a novel tripartite partition method. It sorts the labels of a proof into common labels and labels private to the left or right component, and only lets relational atoms pointing to common labels contribute modalities to interpolants. Given a proof of an implication, our method computes a strong Lyndon interpolant together with proofs of its correctness.
Qiyuan Xu, Joshua Ong Jun Leang, Renxi Wang, Wenda Li, Haonan Li, Luke Ong, Conrad Watt
原文摘要
Interactive theorem proving (ITP) underpins program verification and formalized mathematics, but its manual effort limits scalability. LLM-based proof agents promise to ease this effort, but their heavy token consumption and API cost remain a major obstacle. We trace this cost to a shared root: current agents operate on serialized concrete syntax, emitting proofs as source text and recovering proof states through separate, line-number-based queries, so every edit shifts later lines and forces repeated relocation of errors and states. This same dependence on concrete syntax also blocks adoption of Minilang, a recent proof language that reaches SOTA on LLM-based proving but is too new for LLMs' training corpora. We address both problems by lifting the agent off source text and onto the abstract syntax tree (AST): the model supplies proofs as JSON representations of Minilang's AST -- native to tool-calling LLMs -- and drives the prover through a tree-edit model that fuses proof operations and states into one proof tree, so each operation carries its own subgoal's state, readable directly off the tree. We realize this design in \emph{Agent over AST} (AoA). Against Amazon's Isabelle Agent on miniF2F and NTP4VC-Pearl common success sets, AoA cuts API cost by 2.3--4.7x (normalized input-cache accounting), uses 2.9--6.9x fewer tokens and 3.9--8.9x fewer tool calls, and finishes 1.4--2.0x faster -- while also solving far more problems on the harder verification benchmark.
Local language models often exist at several precisions, and quantization is known to change individual answers. This study asks whether the disagreement between two precisions of the same weights, an 8-bit copy that answers and a lower-precision copy that reads the same prompt, works as a per-question error score for abstention, escalation and re-asking. The cost is one extra forward pass over fewer bytes, with no sampling and no training. We simulate round-to-nearest quantization with per-group scales (3 to 6 bits, groups of 32 to 256) on Qwen3-0.6B and Qwen3-1.7B and score 8,398 multiple-choice questions from ARC-Easy, ARC-Challenge, MMLU and HellaSwag subsets. Error detection is measured by AUROC, risk-coverage curves and ECE on a held-out half, with paired bootstrap intervals over questions. For Qwen3-0.6B, the entropy of the 8-bit answer reaches AUROC 0.682; the best disagreement score, selected on a development half, reaches 0.684 (+0.001 [-0.009, +0.011]), and combining it with the confidence features adds +0.009 [+0.002, +0.015], mostly on ARC. On Qwen3-1.7B the selected score is -0.001 [-0.009, +0.006] from entropy and the combination +0.003 [-0.001, +0.006] on the test half. Gaussian weight noise at matched error energy is about as informative as rounding (0.6B only), so we find no evidence that the signal is specific to quantization, and shuffling the features within task and entropy cells removes the gain. One extra 8-bit pass that shifts the option order does better (+0.049 [+0.031, +0.067] over entropy) at about a third more weight traffic; routing to a larger model on free entropy beats both extra-pass scores at equal total traffic. After logistic recalibration the 0.6B confidences are calibrated within a few points of ECE. A deployment recipe and the cases where the method does not help are given.
Laya is a small model that answers a multiple-choice question about a text in one pass: the question, the options (each with a name and a description) and the text form one sequence of at most 512 tokens, the window, in the English release. A longer text is silently cut at the end, and we measure what that costs. Test documents place a held-out row's text, the evidence, at a chosen token offset among distractor text; 177 items, five tasks, three checkpoints (released, short-row fine-tune, long-document fine-tune), four window sizes and 93,275 scored requests. When the evidence lies beyond the window, accuracy falls to 26.4% for the released checkpoint (27.2% for both fine-tunes), near the 27.0% of guessing and below the 35.6% of always giving the commonest label. Confidence gives no warning (AUROC 0.50 to 0.55 for telling cut from read requests; 0.5 = chance), and the runtime's truncation flag also fires on requests whose evidence was visible. The released checkpoint reads visible evidence badly once text precedes it: 38.5% pooled against 60% when the evidence comes first, so a larger window alone gains +6.2 points. Fine-tuning on long documents adds +27.6 points on visible evidence over the short-row fine-tune (95% interval +22.3 to +33.1) and nothing on cut evidence. That is an in-family result: training rows and test documents share generator and question types, and the main filler is of the training kind. With encyclopedic filler never seen in training the gain is +13.2 (+9.9 to +16.6), with rows of other tasks as filler -2.0. Among training-free rearrangements a 312-token window scan helps most (+11.9 points over the default for the released checkpoint) at six to eleven forward passes. Documents are synthetic, labels are LLM-made and one backbone is studied, so these diagnose this model, not long documents in general.
Many learned predictive models represent their uncertainty as a probability distribution, and planners and policies use this distribution to judge risk. However, the fact that a model outputs a distribution does not tell us whether that distribution captures the uncertainty in the environment which is relevant for a decision. In this work we study this question in distributional reinforcement learning, where the learned return distribution P_θ(Z | s,a) can be compared directly with an environment-grounded distribution P_env(Z | s,a) estimated by snapshot-restart Monte Carlo. Our audit first screens for the claims that a risk-sensitive user would act upon, using the excess Wasserstein gap between the top two actions (which is equal to the mass by which first-order stochastic dominance is violated). Each claim is then tested against ground truth using permutation nulls, bootstrap refutation and FDR control. Across QR-DQN, C51 and IQN on MinAtar (33 runs), 40-95% of the strongest claimed risk trade-offs are refuted at 95% confidence, their placement is statistically indistinguishable from a truth-blind placement, and almost none of them can be confirmed. A similar pattern is seen for a pretrained full-Atari QR-DQN. On positive controls of known magnitude the same audit confirms 96-100% of real claims (correlation 0.89-0.92), so it is able to detect real risk when it is present. Acting on the learned CVaR advice at the most-flagged states gives results ranging from beneficial to significantly worse than chance. Ensembling, risk-sensitive training and recalibration do not fix the problem, and the recalibration experiment indicates that the learned spread has no usable information about the true risk. Based on these results, we suggest that uncertainty estimates, including those of learned world models, should be checked against realized outcomes at the decisions where they are used..
Efficiently scaling industrial Click-Through Rate (CTR) prediction has recently attracted significant research attention. Existing approaches typically employ early aggregation of user behaviors to maintain efficiency. However, such non-unified or partially unified modeling creates an information bottleneck by discarding fine-grained, token-level signals essential for unlocking scaling gains. In this work, we revisit the fundamental distinctions between CTR prediction and Large Language Models (LLMs), identifying two critical properties: the asymmetry in information density between behavioral and non-behavioral features, and the modality-specific priors of content-rich signals. Accordingly, we propose the Efficiently Scalable Transformer (EST), which achieves fully unified modeling by processing all raw inputs in a single sequence without lossy aggregation. EST integrates two modules: Lightweight Cross-Attention (LCA), which prunes redundant self-interactions to focus on high-impact cross-feature dependencies, and Content Sparse Attention (CSA), which utilizes content similarity to dynamically select high-signal behaviors. Extensive experiments show that EST exhibits a stable and efficient power-law scaling relationship, enabling predictable performance gains with model scale. Deployed on Taobao's display advertising platform, EST significantly outperforms production baselines, delivering a 3.27\% RPM (Revenue Per Mile) increase and a 1.22\% CTR lift, establishing a practical pathway for scalable industrial CTR prediction models.
Until recently, the biggest obstacles to natural spoken interaction with social robots were speech recognition errors and turn-taking failures. Advances in speech recognition have enabled much better spoken interaction, but turn-taking failures and other misunderstandings continue to be far too prevalent. This short paper looks back at earlier methods for improving speech recognition and turn-taking in interaction with social robots, and then looks at what can be done today, arguing that multimodal bidirectional streaming is needed to reduce the frequency of turn-taking failures.
Micro aerial vehicles that operate without GPS often run SLAM on ARM-based edge boards, yet there is little published data on how standard RGB-D SLAM systems behave on such hardware. We benchmark RTAB-Map and ORB-SLAM3 on the Raspberry Pi 5 using six TUM RGB-D sequences at three image resolutions, recording tracking coverage, trajectory error, time per frame, memory, CPU load and temperature for every run, and we profile both systems at the stage and function level. RTAB-Map was configured to process and keep every frame, so its back end runs on every frame rather than at its default rate of 1 Hz. On the sequences that every configuration tracked, its front end takes 38-74 ms per frame and its full pipeline 150-462 ms, compared with 13-45 ms and 26-62 ms for ORB-SLAM3. RTAB-Map's main cost is optimizing a pose graph that keeps growing, because its working memory is unbounded by default; ORB-SLAM3's local bundle adjustment costs about as much as its tracking but runs in a parallel thread. At quarter resolution, both systems lose tracking on most sequences, and runs that lose tracking finish sooner and report lower error on the poses they keep, so time and accuracy must be read alongside coverage. Timing is repeatable to within 1-5%, but trajectory error varies by up to 27% between identical runs, so the accuracy differences between the two systems are inconclusive. No thermal throttling occurred.
AI buying agents may alter sellers' incentives as well as consumers' decisions. In a synthetic duopoly, we evaluate price-comparison pipelines on a fixed menu and in empirical correlated-equilibrium games. Calculator assistance selects an optimal offer 84.00% of the time, versus 68.57% under verified-total disclosure, yet disclosure has a 20.14-unit higher consumer-worst equilibrium surplus. A frozen-strategy holdout comparison favors disclosure, while re-solving equilibria under scenario resampling produces a range spanning zero. These model-based results show why accuracy, loss severity, seller incentives, and welfare incidence require separate measurement.
Chendong Xiang, Pengle Zhang, Jinliang Li, Zichen Liang, Jintao Zhang, Fan Bao, Jun Zhu, Jianfei Chen
原文摘要
Sharded data parallelism enables memory-efficient training of large models. While efficient training kernels favor packed, kernel-specific data layouts, optimizers may require variant data structures as optimization units. Existing sharded data-parallel frameworks couple these two choices, limiting runtime efficiency and optimization flexibility. To address this, we present \textbf{\our}, a training framework that separates runtime parameters from update parameters, allowing computation layouts and optimization units to be organized independently. Through shared data views and hierarchical sharding, \our supports this separation while avoiding redundant data movement. This design supports a broad range of optimizers, including matrix-based methods such as Muon, and performance optimizations such as fully zero-copy execution. Experiments on video and language model workloads with MXFP8 training show that \our achieves higher Model FLOPs Utilization (MFU) than FSDP2, veScale-FSDP, and Megatron-FSDP. At 128 GPUs, MFU improves by up to 21.7\% over the fastest evaluated baseline, with comparable or lower peak memory usage.
A harness is the code around a language-model agent that organizes prompts, calls tools, manages context, and controls execution. As models grow stronger, recent work has begun to let agents improve their own harnesses, a line of work known as self-evolving harnesses. In most existing methods, a separate proposer running on a human-designed harness modifies the solver's harness, and a separate harness is evolved for each benchmark. Real-world tasks come from many domains, so both the evolution and the evaluation of a harness should cover a diverse range of tasks. We propose a framework close to recursive self-improvement: the same frozen model, on the same version of the harness, first solves tasks as the solver and then, as the proposer, reads the complete run records and directly edits the harness that runs it. Each evolution batch draws tasks from five benchmarks in different domains. To measure generalization, training and held-out tasks are strictly separated, and we additionally evaluate on five out-of-distribution benchmarks never used during evolution. We frame the evolution process as deep-learning training with two stages, multi-task pretraining and continual training. Starting from a 49-line seed harness, the harnesses evolved by GLM-5.3 and GPT-5.6 Sol gain about 5 points on the in-distribution benchmarks, matching or exceeding Claude Code and Codex, and the GPT-5.6 Sol harness is on par with Codex on five out-of-distribution benchmarks. Continued evolution on Claw-Eval in the second stage then lifts GPT-5.6 Sol and GLM-5.3-Flash above Codex and Claude Code. We also provide an in-depth analysis of the mechanisms that emerged during evolution, including output truncation, history compaction, completion checks, independent review, and visual inspection. Code and the evolved harnesses are available at https://github.com/xqkkk/self-evolving-harness.
Planning agents and benchmarks treat a user's initial specification as fixed. Yet the requests users write exclude plans they would gladly take, since they rarely discover the possible trade-offs they would accept until seeing the options. Agents already revise requests through interaction, but whether this helps planning is unknown. We formalize this step as the ReSHAPE task, which judges a revision by whether its search space matches the acceptable plans and a planner finds one there. Using ReSHAPE Planter, which builds instances from planning benchmarks with a human-validated simulated user, we analyze 5,800 agent-revised specifications in travel and shopping. We find that reshaping matters: one correct trade-off makes the same planner reach an acceptable plan up to 4.6x as often, and search space quality predicts what the planner finds. Current agents, however, admit at most 32% of the acceptable plans, where that trade-off alone admits over 52%, because they ask about preferences instead of showing plans; in travel, showing plans raises F1 by 9.7 points over asking, with two model families. Planning agents should thus treat a request as a hypothesis to test against the user's reactions to concrete plans to recognize user trade-offs.
Hao Kang, Weili Xu, Ziyang Li, Zheng Du, Hanchen Li, Tian Xia, Song Bian, Tushar Krishna 等 11 位作者
原文摘要
Agentic workloads are a major source of LLM inference demand. Existing prefill--decode (PD) disaggregated serving systems offer limited flexibility for runtime adaptation of compute allocation and KV cache management, and rely on coarse signals that only partly capture the diverse and dynamic agent workload. These limitations lead to load imbalance between prefill and decode nodes and eviction of KV caches that agents will reuse. We present Evolving Infra, an agent-adaptive serving framework that adapts compute allocation to agent workloads and KV cache management to agent context management at runtime. We decouple compute allocation from key--value (KV) cache management, allowing nodes to switch efficiently between prefill and decode. Building on this design, an online planner matches the PD split to workload composition, while an agent-aware distributed KV manager guides cache eviction and transfer with signals from agent context management. Across coding, math, and computer-use tasks with GLM-5.2 and Kimi K2.6 on up to 256 GB300 GPUs, Evolving Infra reduces mean end-to-end request latency and P95 TTFT by up to 4.1x and 14x over Dynamo, NVIDIA's production framework. We open-source the system implementation of Evolving Infra at https://github.com/mit-han-lab/EvolvingInfra.
Human curiosity and longing for the future have driven human progress from ancient times to the present. With the advancement of AI technology, human prediction algorithms have also drawn increasing attention from researchers. However, as AI technology continues to advance, we observe a counterintuitive phenomenon. Take human trajectory prediction algorithms as an example. Although models' quantitative prediction performance has improved, yielding one state-of-the-art model after another, the consideration of human behaviors, future thoughts, and social interactions themselves presents a gradually diminishing trend in the construction of these prediction algorithms, buried beneath numbers and metrics. In this manuscript, taking interactive art and human trajectory prediction algorithms as examples, we attempt to adopt a different analytical perspective to analyze the structural and developmental commonalities between these two fields. By examining specific interactive artworks and prediction algorithms, we aim to explore how human intuitions and art are associated with AI algorithms through three aspects, including perception, interaction, and timing, thereby providing new insights for the understanding and further research in this domain. We also observe an interesting phenomenon that certain art-inspired prediction algorithms can actually deduce intriguing patterns of human behaviors from the data, offering new potential avenues for understanding the relationship between human activities, state of The Art, and AI algorithms.
During transitions from hospital to home, older adults are frequently affected by fragmented communication, limited coordination between levels of care, and insufficient post-discharge follow-up. Digital technologies, including remote patient monitoring (RPM), digital communication, and artificial intelligence (AI), offer opportunities to improve continuity of care. However, their integration needs to reflect the organizational and clinical contexts of different healthcare systems. This study presents the development and early implementation of digitally supported transitional care pathways within the TransCare project. Care pathways were analyzed and redesigned across three pilot sites in Italy, Romania, and Norway. Existing clinical and organizational processes were mapped to identify challenges and opportunities for improvement. Technology-augmented pathway matrices were developed to integrate RPM, digital communication tools, and AI-supported analytics across the stages of assessment, hospitalization, discharge, and post-discharge follow-up. The pathways were iteratively refined through a co-design process involving healthcare professionals, patients, caregivers, and technical partners. Early implementation experiences were subsequently analyzed across the three sites. Common challenges included fragmented communication and poor care coordination, particularly during discharge and post-discharge follow-up. However, the underlying cause and required solutions differed across the pilot sites, reflecting variations in patient profiles, healthcare organizations, available resources, and digital maturity.
Cross-organizational collaboration is widely regarded as a key promise of SysML-based Model-Based Systems Engineering (MBSE), yet practitioners still face persistent challenges when exchanging and integrating system models. In parallel, Large Language Models (LLMs) raise expectations for AI-assisted model understanding and integration, while reliability and required human oversight continue to pose challenges. This paper reports the results of an online questionnaire survey with 29 MBSE stakeholders involved in cross-organizational collaboration. Respondents rated eight predefined integration challenge categories and six AI-supported task types on five-point Likert scales. The results indicate that stakeholders perceive model integration as a multi-dimensional alignment problem across semantics, behavior, traceability, and exchange interoperability. These perceptions vary by organizational role and frequency of integration involvement. AI is rated highly useful for analysis tasks such as semantic structure analysis and inconsistency detection, and respondents predominantly prefer human-in-the-loop use with mandatory verification. These findings motivate AI support that enhances, rather than replaces, engineering responsibility in SysML-based integration.
Hyperbolic geometry is well suited to embedding hierarchical data, capturing tree-like structure faithfully even in the low dimensions where Euclidean dimensionality reduction is severely constrained. We introduce Random Helgason--Fourier Features (RHFF), an explicit finite-dimensional feature map for kernels on hyperbolic space. We characterize the spectral density from which RHFF frequencies are drawn, allowing kernels such as the Matérn family to be specified directly in the frequency domain. We also prove a uniform approximation error bound for RHFF. Building on this representation, we develop a hyperbolic latent variable model called RHFF-LVM, which requires no numerical quadrature, variational bounds or auxiliary variables unlike previous hyperbolic GP-LVMs. Our RHFF-LVM recovers the metric of a synthetic tree more faithfully than all baselines and attains the best mean rank, driven by large gains on tree-structured atlases at $Q\!=\!2$. Furthermore, it fits $5\times10^{5}$ points in $505$ seconds, making it the first non-parametric hyperbolic LVM to reach this scale.
Mihnea Sebastian Burlacu, Ionut Anghel, Tudor Cioara
原文摘要
The reduced availability of medical datasets represents an important challenge in the development of predictive models for remote healthcare solutions, especially when the available data are imbalanced or difficult to extend through traditional methods. Small Language Models (SLMs) have become a promising alternative to Large Language Models (LLMs) for addressing these limitations due to their lower computational requirements and adaptability to specialized tasks. However, their performance remains highly dependent on the quality and representativeness of the available training data. In this paper, we propose a method for clinical data augmentation based on a fine-tuned Microsoft Phi-2 model that is used to generate synthetic instances, with the purpose of improving deep learning-based model for predicting remotely monitored vital signs. The Phi-2 model was adapted to the medical domain through supervised fine-tuning on the MedQuad dataset, through a Quantized Low-Rank Adaptation (QLoRA)-based method and by applying few-shot prompting techniques. The resulting model was used afterwards in the augmentation of a small-sized dataset, composed of vital sign values and labels associated with pathological and physiological states. The experimental results demonstrate that fine-tuned SLMs improved the performance on medical domain, achieving higher scores for both lexical (ROUGE) and semantic (BERTScore) evaluation metrics compared to the pretrained model. Moreover, the fine-tuned SLM is effective for generating structured synthetic clinical data. When used for data augmentation, the generated dataset improved the performance of a LSTM predictive model, increasing the classification accuracy from 89.0% to 93.7% and showing consistent improvements in precision, recall, and F1-score.