Papers

Conversational AI systems can pose safety risks to their users such as hallucination, sycophancy, overconfidence, and anthropomorphism, but these risks are difficult for users to detect during everyday use. We introduce Safety Nudges, a browser-based tool that provides lightweight, in situ flags when concerning behavior is detected in chatbot conversations. We evaluated Safety Nudges in a two-week field study with 45 frequent chatbot users, collecting interaction logs, surveys, and feedback on individual nudges. Participants found the tool useful, clear, and minimally disruptive, with nearly all users reporting an increased awareness of potential AI harms, though we found that this improved awareness alone did not necessarily lead to discernible behavioral changes. Our results suggest that user facing safety nudges can complement model-level safeguards by helping people critically evaluate AI responses in context, while highlighting the importance of relevance, calibration, and user control in nudge design for conversational AI safety. The code for our Safety Nudges extension is publicly available at https://github.com/jtbwedgwood/safety-nudges.

Long-running agents accumulate reusable Skills, but a Skill that is semantically relevant to a task is not necessarily worth loading in the current state. We study the applicability question that arises once a candidate Skill is known: should it be loaded in the current state? We propose SkillApt, which uses matched WITH/WITHOUT Skill executions on the same task state as persistent evidence, estimates the conditional marginal utility of the Skill, and chooses LOAD or ABSTAIN accordingly. The base model, agent architecture, and Skill contents stay fixed; only the external deployment policy changes. We call this constrained setting strategy-level recursive self-improvement (Strategy-Level RSI). On 20 Skills and 160 held-out states, as paired evidence accumulates, SkillApt's task success rises from 81.9% under a cold start to 91.3%, matching a strong zero-shot LLM controller; yet SkillApt activates Skills on only 26.3% of states, versus 98.8% for the zero-shot controller. A hard-candidate study shows that non-optimal Skills mostly leave correctness unchanged while raising execution cost, and occasionally cause correctness harm. An ablation shows that a history recording only WITH success makes the policy load almost everywhere, whereas paired evidence substantially improves selectivity. These results indicate that relevance is not applicability: the main effect of execution evidence is not to make the model stronger but to change how existing Skills are deployed, moving the system from near-always loading to selective reuse.

Web applications are increasingly targeted by cyberattacks that exploit HTTP requests to evade security mechanisms. Traditional web application firewalls (WAFs) rely on rule-based approaches that often exhibit high false positive rates and limited adaptability. Recent studies have explored machine learning techniques and word embedding models to improve anomaly detection in HTTP traffic. This paper presents a benchmark for static embedding models, specifically Word2Vec, FastText, and Doc2Vec, within a unified, single-class classification framework. We propose HEDA (HTTP Embedding-Based Detection Architecture), a modular detection pipeline that combines static embedding representations with single-class anomaly detection models to detect anomalies at the request level. The approach operates in an unsupervised environment, where both the embedding models and detectors are trained exclusively on benign HTTP traffic. The proposed methodology is evaluated on three datasets with heterogeneous characteristics, including both synthetic and real traffic. The experimental results show that the choice of embedding representation significantly affects detection performance, and that FastText-based embeds produce the most consistent results across all datasets, achieving high detection rates while keeping false positive rates under control.

Navigation in off-road conditions is challenging due to the lack of structure. There is no fixed vocabulary for what is traversable. The traversability depends on both the environment and the embodiment's dynamics. Neither of these two variables can be hand-labeled at scale. Thus, traversability has to be learned by the embodiment's own experience. Modern platforms tend to use multiple sensors to estimate traversability and navigate: RGBD cameras, lidar, radar, IMU, with computationally intensive platforms to run inference on neural networks. Against this trend, we propose FLINT, a lightweight traversability estimator: a 21.6M-parameter backbone, 38\times smaller than a comparable foundation-model backbone, that scores higher on held-out terrain probes and runs at 14.7 FPS on CPU alone using a RGB camera has the only sensor. Despite that gap in scale, FLINT produces a cheaper, more accurate costmap than a deployed foundation-model system (WildOS) on 23 of 24 replayed field logs. We compare different self-supervised learning signals and deploy the resulting models on a real platform in closed-loop field trials: the best self-supervised head reaches 99% autonomy over the route, outperforming a human-label-trained baseline deployed live on the same course. Our results show that heavy sensing and computing are not necessary for traversability estimation.

Neural information retrieval for Italian is served almost entirely by multilingual models. Several multi-vector (late-interaction) retrievers include Italian among dozens of languages, and several strong Italian dense embedders exist, but as of August 2026 no late-interaction retriever specialised on Italian had been released. We present ItColBERT, a 135M-parameter Italian ColBERT trained with PyLate following the ColBERT-Zero recipe: initialise from a checkpoint that already retrieves, then apply supervised contrastive training followed by single-teacher distillation, for a total of roughly 14.5 GPU-hours on one RTX 3090. Across four Italian retrieval benchmarks it outperforms every general-purpose late-interaction baseline we tested except one (mLateOn), at 2-4.4x fewer parameters than every baseline but one of comparable size. Our principal empirical finding is methodological and partly negative. On the only cleanly out-of-domain benchmark (MLDR-it), an inference-time chunking recipe applied to an unchanged checkpoint yields +0.0602 nDCG@10 (p = 0.0225), a larger effect than anything two further rounds of training produced. Self-mined hard negatives and native 1024-token training were both evaluated against pre-registered decision gates and both failed. We report every comparison with paired bootstrap tests against an empirically measured noise floor of 0.0030 nDCG@10, and we release the weights, the training and evaluation code, and the complete experimental record including the rejected rounds.

Relational Deep Learning (RDL) models multi-table databases as heterogeneous temporal graphs, and graph transformers currently achieve state-of-the-art performance on benchmarks like RelBench. However, the current leading model, RelGT, suffers from two key limitations: its random local sampler yields loosely connected subgraphs that hinder message passing, and its global attention module relies on a single, seed-feature-based memory that ignores broader macro-level dynamics. To overcome these limitations, we introduce QUARTET, an expressive graph transformer architecture that applies full self-attention on local subgraphs while enriching global context through cross-attention branches. Specifically, QUARTET employs a Causal Random Walk (CRW) sampler based on recency-truncated Personalized PageRank (PPR) to extract compact, hub-robust, and densely connected local subgraphs without temporal leakage. Concurrently, a quad-branch cross-attention module integrates global context from four complementary perspectives: seed feature, seed topology, temporal dynamics, and collaborative dynamics. Across the RelBench v1 classification tasks, QUARTET consistently matches or outperforms the current state-of-the-art graph transformer baselines (HGT and RelGT). Ablation studies confirm that the CRW sampler significantly enriches local neighborhood quality, while the global branches provide essential, task-specific predictive gains.

Sound Scene Generation is about the automatic synthesis of artificial sound scenes. We introduce SsgCaps, a publicly available dataset of human-engineered sound scenes wherein each scene matches a precisely structured prompt that guides the sampling process. The corresponding prompts are sampled from a predefined action-based typology that allows extensive sampling while retaining plausibility. SsgCaps is a sound scene dataset derived from the unpublished reference dataset for Task 7 of the 2024 DCASE Challenge edition, which contained private-and public-domain audio samples. In contrast, SsgCaps contains only public-domain audio samples, allowing us to open this dataset to the community. To make this dataset useful to the community, we first elaborate on the rationale for the prompt and dataset structure. We then perform a comparative quantitative analysis of the 2 versions of the dataset. To do so, we compare both versions to the audio synthesized by the SSG algorithms submitted to the challenge using Fréchet Audio Distance (FAD) and Kernel Audio Distance (KAD) as well as perceptual ratings. This analysis shows only small differences, which enables us to recommend the open version for further benchmarking of SSG algorithms.

While Large Language Models (LLMs) have achieved remarkable results across various benchmarks, their alignment with normative values often results in homogenized responses that fail to address diverse user preferences. Existing training-free methods often occupy valuable context windows through prompt engineering, while training-based methods typically remain static post-training, failing to support the continual optimization required in real-world settings. To address these challenges, we propose COPE (Continual Optimization with Personalized embedding and self-Evaluation), a novel optimization framework tailored for real-world-motivated interaction settings with sparse user feedback. Our framework assigns learnable personalized embeddings to each user and synergistically integrates preference capture, self-evaluation calibration, and personalized response optimization within a single update step. A key innovation of our method is the use of self-evaluation to generate proxy rewards, enabling continuous model updates even when explicit user feedback is unavailable. Experiments show that COPE consistently outperforms strong training-free and training-based baselines under sparse feedback, and remains complementary to Retrieval-Augmented Prompting (RAP). Further analyses confirm COPE's reliable self-evaluation, meaningful preference patterns, stable general capabilities, and robustness under shifting preferences and alternative evaluators.

We study representations for relation algebras corresponding to certain edge colourings of complete graphs. Previously suitable colourings were obtained for the number of colours $n$ up to $2000$, with two exceptions: $n = 8$ and $n = 13$. Using a method that we have called the FBF (Fusion Beautiful Fusions) method, we find colourings for $8$ and $13$ colours. Our method is a new guess-and-check approach that we describe as an adaptation of Comer's finite-field method. Using FBF we construct novel colourings that are non-isomorphic to existing published colourings for 5, 6, 7, 9, 10, 11 and 12 colours.

Postoperative acute kidney injury (AKI) after major non-cardiac surgery carries substantial morbidity, yet early intraoperative risk stratification remains difficult. In this retrospective cohort study, we propose SynerT, a waveform-only hybrid temporal backbone that combines a causal dilated TCN with a hierarchy of dilated recurrent layers to encode early intraoperative physiologic trajectories for AKI risk prediction. Building on SynerT, we further design two model variants that extend the backbone with structured clinical context: SynerT-MM, a late-fusion multimodal extension that integrates hemodynamic burden summaries and preoperative covariates, and SynerTStack, a leakage-safe stacked ensemble that combines cross-validated predictions from SynerT-MM with strong tabular baselines at the meta-learning stage. All models are evaluated under a strict leakage-aware framework on VitalDB, a high-fidelity perioperative database, with prediction restricted to information available within the first 60 intraoperative minutes. Among 2,413 waveform-usable cases (180 AKI-positive; 7.46% prevalence), SynerT fell well below strong structured-data baselines, demonstrating that waveform-only temporal modeling is insufficient under strict early constraints. SynerTMM recovered discrimination by incorporating hemodynamic burden summaries and preoperative covariates, and SynerT-Stack achieved the best overall performance across AUROC, AUPRC, and F1-max. Cross-fitted Platt recalibration substantially corrected calibration defects in both multimodal variants, and decision-curve analysis confirmed the recalibrated stacked model delivered the strongest net clinical benefit across low-to-intermediate thresholds.

AI coding agents now author a large share of pull requests (PRs) merged into popular open-source projects. A merged agent PR is usually considered finished work; yet, prior studies have reported issues in agent code after the merge (e.g., code smells and static-analysis issues). However, little is known about how often a merged agent PR is fixed afterward, and who actually authors the fixing. In this paper, we follow 6,774 merged agent PRs across five AI coding agents (OpenAI Codex, GitHub Copilot, Devin, Cursor, and Claude Code) from the AIDev-pop dataset (open-source repositories with at least 500 stars) into their follow-up fixes, against a baseline of 5,044 contemporaneous human PRs from the same repositories. We link each merge to its candidate fixes, verify every candidate with human annotators and an LLM judge that matches human-level agreement (binary Cohen's Kappa=0.78 against a human-human K=0.77, Direct-fix precision 90%), and attribute the fixing work at the PR and the commit level. Our findings show that (1) merged agent PRs attract verified fixes at 1.62 times the odds of merged human PRs in the same repositories over the same period of time; (2) 69.6% of verified fixes in agent merges come from the same agent; and (3) 76.4% of the verified fix PRs are agent-authored throughout all commits. These results show that agents currently largely finish their own job, but their merges still require fixing more often than human merges.

We study the reconstruction of a directional albedo field from confocal non-line-of-sight measurements. For mirror-symmetric vector fields in Sobolev spaces, radial preprocessing reduces the data to spherical means of the divergence with centers on the relay wall. Full data determine this divergence uniquely, with divergence-free fields forming the entire ambiguity. An explicit Fourier–sine formula recovers the irrotational Helmholtz component and reproduces the data. The Fourier–sine transforms of the model data form a weighted Hilbert space that we characterize exactly, with norm equal to the Sobolev norm of the reconstructed field. This weighted range remains well defined even when the preprocessed data fail to belong to standard Sobolev spaces. For Schwartz fields in the model class, such Sobolev regularity holds exactly when the divergence has zero depth integral. Measurements on an open subset of the relay wall, for all radii $0<r<R_{\max}$, uniquely determine the divergence in the union of the corresponding balls. An FFT-based algorithm with Stolt interpolation implements the reconstruction in $O(N^3\log N)$ operations on an $N^3$ grid. Tests on explicitly defined fields assess reconstruction accuracy and invariance under divergence-free perturbations. We present imaging examples of potential and vector-field reconstruction from model-generated, rendered, and measured transients.

Residual-based adaptive GMsFEM for high-contrast elliptic problems repeatedly evaluates local weighted $H^{-1}$ indicators on every coarse neighborhood, making indicator evaluation a recurring cost in repeated-query settings. We introduce a non-intrusive Gaussian-process (GP) surrogate for the indicator scores used in Dörfler marking. The operational predictor is the GP posterior mean, algebraically equivalent to a kernel ridge regression (KRR) estimator under the stated convention; it uses compressed local solution and spectral features without changing the multiscale solve, local spectral construction, or basis enrichment. A nonuniform perturbed-marking result quantifies how pointwise score errors affect the exact indicator mass captured by surrogate-selected neighborhoods, while a conditional bounded-discrepancy KRR pathway identifies sufficient assumptions for such score bounds. In controlled held-out in-distribution tests, the surrogate-guided method gives error-versus-DoF trends comparable with classical $H^{-1}$-residual offline adaptivity and evaluates the online indicator component 2.0-2.1 times faster, excluding offline data generation and GP training.

High-capacity neural network models have achieved state-of-the-art performance across diverse classification tasks, yet they frequently operate as black-box models, lacking the transparency necessary for critical decision-making. Such opacity creates a persistent trade-off between performance and explainability. This paper proposes a solution to address this gap: the NeuroRule knowledge distillation framework that results in explainable rule-sets from neural network models. NeuroRule adapts the EVOTER rule-set evolution infrastructure to treat neural networks as targets for the evolution process, distilling their performance into concise sets of propositional logic expressions. There are three primary contributions: (1) an evolutionary method for distilling black-box neural network models into explicit rule-set models; (2) a method for making rule sets more explainable by including a conciseness objective to evolution; and (3) a demonstration that the distillation is viable even without access to the original neural network training data. The paper thus establishes that black-box neural network models can be made explainable and therefore useful in real-world applications where trustworthiness is paramount.

Post-hoc probability calibration is usually evaluated under an optimistic assumption: the held-out calibration labels are clean. In many AI deployment settings, however, labels come from weak annotators, historical decisions, heuristics, or distant supervision, so the same label noise that corrupts training also corrupts calibration. We study this overlooked failure mode for tabular classifiers and propose LWCal, a CPU-only post-hoc calibrator that down-weights calibration examples whose noisy labels are contradicted by the base model's held-out probability. LWCal requires no clean validation labels, no noise-rate estimate, and no retraining of the base classifier. A second variant, Gated-LWCal, adds a conservative disagreement gate that backs off toward the raw score when the calibration split appears extremely inconsistent. On nine local binary tabular tasks, six random seeds, symmetric and asymmetric label corruption, and three tree-based base learners, LWCal obtains the lowest average calibration error while Gated-LWCal obtains the best average proper-score tradeoff. In the main random-forest study over 432 noisy cells, Gated-LWCal reduces expected calibration error from 0.188 to 0.122 and negative log likelihood from 0.438 to 0.396 relative to the raw classifier. Paired bootstrap intervals for Gated-LWCal versus raw, Platt, isotonic, and beta calibration exclude zero on ECE, Brier score, and NLL. The artifact contains all scripts, result tables, figures, and the compiled paper.

Diffusion Large Language Models (dLLMs) have recently emerged as a promising alternative to autoregressive LLMs by enabling non-autoregressive text generation. However, their practical deployment remains limited by inefficient inference, largely due to the absence of effective Key-Value (KV) caching and scalable parallel decoding mechanisms. Existing acceleration methods typically study KV caching and parallel decoding in isolation, overlooking the I/O bottlenecks that arise when cache reuse and parallel token verification are jointly applied. In this work, we introduce $\textbf{Flash-dLLM}$, a training-free inference acceleration framework for fast and memory-efficient dLLMs. Flash-dLLM first identifies GPU memory I/O as a dominant bottleneck in KV-cache-enabled dLLM inference and addresses it with an I/O-aware fused KV-cache kernel that reduces redundant memory movement. Building on this optimized cache mechanism, Flash-dLLM further proposes an efficient KV-cache-driven draft-and-verify decoding strategy, where the dLLM itself serves as both drafter and verifier without requiring an auxiliary model. This unified design enables faster decoding while preserving generation quality and improving scalability to longer sequences and larger batch size. Extensive experiments on mathematical reasoning and code-generation benchmarks demonstrate that Flash-dLLM consistently outperforms existing state-of-the-art dLLM acceleration methods in both inference speed and memory efficiency. In particular, it achieves $5.1\times$ and $11.0\times$ speedups over prior strongest baseline Elastic-Cache on GSM8K and HumanEval, respectively.

3D Gaussian Splatting (3DGS) can reconstruct a captured scene photorealistically, but the resulting representation does not by itself support physical interaction. Robot simulation instead requires object-level change, i.e., objects must move independently, make contact, and reveal previously occluded surroundings. This gap arises because object appearance may remain entangled with the background, while hidden object geometry and occluded background content may be unobserved. To address this challenge, we present \phi-RIE, a Gaussian-native pipeline that converts selected objects into movable simulator assets while preserving the remaining reconstruction. Our key observation is that asset construction and source removal should be coupled, i.e., one object identity should define the movable asset and the scene content to remove and complete. Accordingly, Scene Observation supplies shared evidence to Coupled Scene Construction, which creates registered assets and completed background Gaussians for simulator-driven rendering in an Interactive Environment. This coupling preserves unedited Gaussians while aligning visual and physical state. On 50 ScanNet++ scenes, evidence-based selection and registration retry increase matched F1 at 20\,mm from 0.336 to 0.383 at fixed retention. Further tests demonstrate asset executability, manipulation gains over a single-generator baseline, and the visual cost of conversion. Together, these results demonstrate that \name\ enables interactive scene conversion.

Compositional 3D scene reconstruction has recently been explored from two directions: agentic reasoning that provides semantic understanding of spatial relationships but lacks precise alignment with input images; and visual geometry foundation models that predict dense point maps from input images but the reconstruction quality is limited. Therefore, recovering a complete 3D scene from a single monocular image with accurate inter-object relationships and high-fidelity reconstruction quality remains challenging. In this paper, we present HARMONY, a hierarchical chain-of-thought framework that leverages both agentic reasoning and visual geometry foundation. Given an image of an indoor scene, starting from an empty 3D floorplan, HARMONY first calibrates the camera against the reference image to establish a semantically-grounded spatial frame, then uses agentic VLM reasoning to recover the 3D room layout and an initial placement order. It then places the objects in a hierarchical order, from wall-mounted elements, free-standing furniture, to dependent decorations on top of furniture. We also use depth-first traversal for furniture so each placement conditions on previously resolved structure and a reflective feedback loop to avoid error accumulation. After each object placement by VLM, we use the point cloud estimations to perform geometry-based refinement so that the rendered image aligns better with the input. HARMONY can produce 3D scenes that are semantically consistent and perceptually aligned with the reference image, extending single-image compositional reconstruction to complex indoor scene images. Experiments on synthetic and real-world images demonstrate that HARMONY outperforms the evaluated reconstruction baselines, while qualitative comparisons with GPT-6 Astra suggest more faithful object arrangements and better preservation of scene details.

Faithfully evaluating end-to-end driving policies in simulation requires observations that are not merely photo-realistic, but preserve the scene features a policy relies on to make decisions. Existing platforms, however, exhibit a sim-to-real visual gap that corrupts policy perception, undermining their ability to assess a policy's closed-loop decision-making. To this end, we propose DreamStream, a generative, closed-loop simulator that achieves policy-oriented fidelity using a simulator-grounded autoregressive video model. Our video model is distilled from a large pretrained video model via traffic layout guidance, varying visual appearance while preserving policy-relevant features such as scenario layout and the temporal consistency of dynamic objects. We further observe that perceptual metrics like FID misrank how well these features are preserved. To tackle this, we introduce FD$\pi$, a new multi-representation metric that measures the sim-to-real gap as the Fréchet distance over scene-context features from public E2E policies. Under FD$\pi$, DreamStream improves over the strongest prior closed-loop simulator by $1.6\times$ on nuScenes and $4.7\times$ on NAVSIM, and induces the least perturbation to policy's perceptual observability. Based on DreamStream, we construct Navhard-CL benchmark, which turns non-reactive real-world benchmark NAVSIM into interactive testing environments with adversarial driving behaviors and weather variations. This benchmark exposes many failure modes of driving policies, such as scorer bias and lack of recovery behaviors, that prior closed-loop benchmarks overlook. Code and data are available at https://github.com/VAIL-UCLA/DreamStream.

We establish a fully discrete weak-to-strong (W2S) multi-entropy stability theory for arbitrarily high-order finite-volume and discontinuous Galerkin (DG) approximations of the ideal compressible magnetohydrodynamics (MHD) equations on general polytopal meshes, whereby a single numerical update simultaneously satisfies discrete entropy inequalities for any prescribed finite family of convex Harten entropy pairs. Because physical entropies are defined only for positive density and pressure, the central analytical difficulty lies in reconciling discrete entropy stability with positivity preservation, the magnetic divergence constraint, and the nonconservative Godunov–Powell coupling. Building upon the provably positivity-preserving MHD framework of Wu and Shu, we develop a weak multi-entropy analysis of the common cell-average evolution of finite-volume and DG methods. Through convex decomposition and relative entropy, we establish two sufficient criteria for this weak stability general polytopal meshes. Compatible magnetic corrections and locally divergence-free approximations cooperatively offset the nonconservative Godunov–Powell coupling and provide the discrete cancellations essential to both positivity-preservation and entropy stability. A W2S lifting then unifies positivity and entropy limiting into a single cellwise scaling limiter, achieving strong multi-entropy stability while strictly preserving cell averages and the locally divergence-free magnetic field. High-order temporal accuracy follows from variable-step strong-stability-preserving multistep methods. Numerical experiments confirm the theoretical findings and demonstrate computational robustness. This framework provides a rigorous, fully discrete foundation for unifying physical admissibility, magnetic divergence control, and multi-entropy stability in high-order MHD approximations.

We study asynchronous collective tree exploration, where $k$ agents with unrestricted communication start at the root of an unknown tree and discover edges online. At each step, an adversary chooses which agent moves. We give a deterministic algorithm that explores any tree with $n$ nodes and depth $D$ in at most \[ 2n+O\left(k\log^2(k)D\right) \] moves, matching known lower bounds up to a constant factor. As a direct consequence, we obtain a near-optimal competitive ratio of $O(\log^2 k)$ for synchronous collective tree exploration, where all agents move at each round. The proof relies on a multiscale power regularizer that may be of independent interest.

We present a distributed quantum algorithm that $3$-colors cycles in $O(1)$ rounds, with high probability. It follows that all locally checkable labeling problems (LCLs) that have round complexity $O(\log^* n)$ in the classical LOCAL model can be solved in $O(1)$ rounds in the quantum-LOCAL model, with high probability; this includes problems such as maximal independent set and maximal matching in bounded-degree graphs. This presents the first natural examples of graph problems with an asymptotic distributed quantum advantage for the LOCAL model; all prior examples that separate LOCAL and quantum-LOCAL are artificial problems constructed merely for the sake of demonstrating quantum advantage.

We study decentralized partially observable team decision problems with low-rank latent dynamics and unknown system models. The proposed framework combines team-theoretic equivalence with low-rank model representations to address cooperative decision-making in partially observable Markov decision processes without prior knowledge of the transition model. Each team member makes decisions based on local private information and delayed common information shared across the team. Using only this available information, each member learns an approximate low-rank Markov decision process and applies least-squares value iteration to compute its policy. This yields a fully decentralized learning and planning algorithm that requires neither a centralized coordinator nor centralized training. We show that the resulting member-side solutions approximate the centralized team solution: despite partial observability, unknown dynamics, and delayed common information, each member recovers the corresponding component of an approximate team-optimal policy. We further establish finite-sample performance guarantees and derive a corresponding sample-complexity bound for the proposed algorithm.

A multi-agent system can reduce latency on complex tasks by executing work concurrently. Several pioneering harness frameworks support multi-agent systems. However, the scalability of current multi-agent harnesses is often constrained by a central orchestrator's capacity to allocate tasks and coordinate workers. To address this limitation, we introduce Agensh, a scalable self-organized multi-agent harness without a central orchestrator: concurrent workers execute a multi-agent cooperation loop, continuously gathering context, claiming and self-assigning sub-tasks, taking action and sharing findings, verifying results, and merging progress in an asynchronous manner. The loop is supported by the agentic organization infrastructure comprising three components: a shared workspace holds proposed, ongoing, and completed work; a message interface lets workers communicate; and shared context retains reusable findings and work intentions. To test the scalability of Agensh, we evaluate it on the five hardest ProgramBench tasks with GPT-5.6-sol (high). Scaling from 1 to 128 agents raises the mean final test-pass rate from 19.31% to 28.78%, an approximately 49% relative improvement. Larger organizations reach comparable test-pass rates earlier. On pandoc, scaling from 1 to 1,024 agents raises the final test-pass rate from 33.89% to 55.06%. Worker trajectories further show that different forms of self-organized cooperation gradually emerges and standardizes as the organization grows. These results reveal the number of agents as a new scaling dimension for multi-agent organizations to expand the frontier of general intelligence, offering a practical solution for complex tasks under hard latency constraints or time budgets.

Long-term conversational memory in multi-party settings requires more than retrieving relevant content from long-term conversations: it must distinguish who said what, whom each statement concerns, how individuals perceive one another, what information is shared by the group, and how states change over time. Recent studies on multi-party dialogue benchmarks show that existing general-purpose LLM memory systems tend to lose person and group relations or struggle to integrate clues distributed across members, groups, and time. Together, these issues reveal two core bottlenecks: message attribution and relational understanding in multi-party dialogue, and state reconstruction from interleaved histories. To address both, we propose $\textbf{SpeakerMem-R1}$: its dual-track memory stores speaker-labeled verbatim messages and derived states organized into person-level and group-level views, then combines evidence from both tracks by entity, event, and time at query time. To reduce attribution and update errors during structured memory construction while enabling local deployment, we train Writer-R1 with SpeakerLevenshtein and speaker-conditioned GRPO. On GroupMemBench, SocialMemBench, and EverMemBench, SpeakerMem-R1 achieves binary accuracies of 47.9%, 69.2%, and 61.9%, respectively. On the publicly reported EverMemBench leaderboard from EverMind-AI, we achieves 62.33%, the best reported result among the latest state-of-the-art frameworks. It also achieves 70.85% on all 1,986 LoCoMo questions, which we use as a two-person long-term conversation boundary test. In a controlled evaluation of 305 questions, RL raises the SFT Writer's mean accuracy from 57.38% to 68.20%. We report both binary accuracy and token-F1, and ablations show that the verbatim and structured tracks, as well as person-level and group-level views, are complementary under the standardized evaluation interface.

Agents often work on complex problems that require millions of tokens of context, which necessitates compacting across sessions due to limited context windows. We develop CliffCompaction, an autocompaction technique that reduces cost by up to 50% under a bounded context while maintaining or improving performance on Terminal-Bench and achieving new levels of efficiency for test-time scaling and state-of-the-art results on KernelBench. The per-rollout savings of CliffCompaction make the performance–cost trade-off of test-time scaling more efficient, adding over 10 percentage points on Terminal-Bench for less than the cost of two full-context runs. Under parallel test-time scaling, CliffCompaction lets Kimi K2.6 match Opus 4.7, and exceed Opus 4.6 and GPT-5.3 Codex at lower cost. The key to CliffCompaction's effectiveness is that it keeps compacted information faithful by only truncating or dropping content, never rephrasing or rewriting it. We never compact a compaction—each pass operates only on original content, and prior compacted output is discarded, preventing context drift from accumulating. These properties sustain continual learning over sessions exceeding a million tokens: on KernelBench, CliffCompaction reaches CUDA kernel speedups of $2.23\times$ after 200 steps and $3.58\times$ after 400 steps, surpassing specialized search algorithms and trained agents despite being a general-purpose compaction technique. We open-source a scaffold-agnostic API-proxy implementation of CliffCompaction usable with Claude Code, Codex and other harnesses.

Three-variable first-order logic ($\mathrm{FO}^3$) and the calculus of relations (CoR) define the same binary queries, an equivalence that goes back to Tarski's work in the 1940s. The classical translation from $\mathrm{FO}^3$ to CoR is single-exponential, but it has remained open whether this blowup is necessary. We prove that it is, even over finite structures and even for circuits that share subterms. Our proof introduces a new technique: we modify a single finite structure in exponentially many local ways, each changing the answer of a short $\mathrm{FO}^3$ formula $\varphi$. In any CoR circuit $C$ equivalent to $\varphi$, only compositions of binary relations can notice these modifications, and each composition can notice at most two; thus, $C$ requires exponentially many compositions. The same technique gives an exponential succinctness gap between $\mathrm{FO}^3$ and $\mathrm{FO}^2$, and exponential lower bounds for matrix query languages over the Boolean and tropical semirings, where the analogous question was also open.

We introduce SWE-Serve, a benchmark for evaluating agents on production inference engineering tasks. Implementing an inference feature can require coordinating multiple changes across the serving stack, including model support, runtime execution, and public APIs. Existing benchmarks provide limited coverage of production inference engineering: repository-level software engineering benchmarks do not target inference, while general terminal-agent benchmarks include only a few inference tasks. Dedicated inference benchmarks, meanwhile, focus primarily on isolated kernel generation or performance optimization rather than repository-scale production feature implementation. SWE-Serve provides 53 repository-grounded tasks derived from recent production changes to SGLang, spanning six inference engineering families. Each task executes on either CPU or a single GPU (H100) and is evaluated with hidden functional and regression tests, including, where applicable, end-to-end (E2E) serving tests and calibrated performance gates. Executable no-op and oracle controls, adversarial verifier review, and closed-book execution support task validity and evaluation integrity. Across 11 models and 31 model-effort configurations, the best-performing configuration achieves 75% mean pass@1. SWE-Serve exposes a substantial gap between completing tasks locally and achieving production correctness. On 19 tasks with end-to-end coverage, model-serving E2E tests reject roughly one-third of patches that pass every other test (45.9% under the verifier versus 69.4% with E2E tests excluded from scoring), with pass rate increasing for each model's best-performing configuration. By making the production correctness gap directly measurable, SWE-Serve enables the field to track whether future agents move beyond completing tasks locally to achieving production correctness.

Vector Quantization (VQ) is fundamental to discrete visual tokenizers that power modern autoregressive and masked image generation models. While recent shared-projection codebook methods have substantially advanced codebook utilization, training stability remains a critical and underexplored challenge. We argue that the root cause lies in the entanglement of the Encoder–Decoder and Codebook training: because neither module can reliably fulfill its own responsibility in isolation, the system can only function when the two subsystems happen to cooperate—a fragile condition that breaks down precisely when training is most stressed. We propose StableVQ, which revisits the proper learning objective of each module and resolves the problems that arise when each is trained to fulfill its own role independently. Concretely, (1) Dynamic STE corrects the instability in the Encoder's learning objective, enabling it to robustly optimize the reconstruction space under discrete regularization even when codebook utilization is low. (2) Region VQ Loss reconceives the Codebook's learning objective so that it can independently guarantee full tracking of the encoder output distribution, without relying on encoder oscillations to drive activation. (3) Decoupled Schedule recognizes that the distinct responsibilities of the Encoder–Decoder and the Codebook demand distinct optimization dynamics, and assigns each an independent learning rate schedule to ensure robust system-level behavior. Built on top of shared-projection codebooks, StableVQ is lightweight and introduces no learnable parameters. Experiments on ImageNet demonstrate consistent improvements in training stability, codebook utilization, and reconstruction quality across diverse codebook sizes and initialization settings.

Decoherence free subspaces are a steady-state structure of the open quantum system which preserves quantum coherence between the states lying with in it and thus has found a variety of applications throughout quantum information science and technology. In this paper we study the computational complexity of deciding whether an open quantum system admits a decoherence free subspace or not. More specifically we study this problem with in the context of Markovian open quantum systems, governed by the time-independent Lindblad master equation. Along the way we introduce the $k$-Local Lindbladian problem, which captures the difficulty of computing purity decay rates under Lindbladian dynamics. We show that both problems are hard for the complexity class Quantum Merlin Arthur (QMA) when the locality $k \geq 5$, with the first under perfect completeness and the second being complete for QMA. Our hardness construction generalizes Kitaev's clock Hamiltonian construction to the open quantum system setting by encoding the execution of a quantum circuit into the steady subspace of a Lindbladian containing both pure and mixed history states. This subspace is then mixed depending on the output of the encoded circuit. Our results suggest that deciding whether a generic Markovian open quantum system admits a decoherence free subspace is intractable even for quantum computation.