Papers

We identify intervention bias as a previously unquantified failure mode of zero-shot large-language-model (LLM) educational advisory agents: without task-specific training, they recommend action when a hindsight-optimal oracle policy mandates inaction. In a six-arm ablation on the Open University Learning Analytics Dataset (N=800 students, four temporal cutoffs), at day 56 – when the oracle designates 70.1% of students as needing no intervention – zero-shot GPT-4o recommends action for 73%, a 43 percentage-point false-positive rate. Commercial RAG and SQL-augmented retrieval are comparably miscalibrated; at 10,000 students this implies about 4,300 unnecessary advisor contacts per cycle. Supervised policy learning eliminates this bias: a trajectory-conditioned ONNX Decision Transformer (DT) and a snapshot XGBoost classifier, trained on the same oracle-labelled trajectories under strict prefix-only features, both achieve near-zero calibration error. The DT reaches macro-F1 0.79 (macro-recall 0.85) across all five action classes, predicting even the rare load-reduction action without collapsing, at a 0% action flip rate and sub-5 ms CPU decision latency. The two supervised arms are on par; the DT's edge over XGBoost at the final cutoff is indicative only (unpaired across cohorts). Scope: we validate Stage-2 decision-making (EAV state vector to supervised policy) under controlled oracle input from structured OULAD data; high fidelity reflects feature-oracle alignment, not general high-stakes-AI capability. The most robust finding is the intervention-bias contrast, not the absolute accuracies. We also show an Evaluation Gap: LLM-as-judge scoring (DeepEval G-Eval) is blind to intervention bias, rewarding fluent over-prescription rather than decision quality.

LLM agents carry conclusions across steps and sessions in compressed memory, and memory products (e.g., mem0, LangMem) rewrite conversation into stored "facts" that later steps trust. We show this rewriting manufactures confidence: across our constructed agent settings, a casual, hedged remark becomes a confident, dated assertion the agent then obeys like a verified fact, granting every above-clearance request it faces. No attacker is needed: a role that was true once and never corrected is stored as a flat fact and acted on like a deliberate injection. We then isolate what the agent responds to. It is not the source: attributed, unattributed, and even forged "system of record" claims all grant alike. It is the confidence of the phrasing. A hedge is discounted, a flat assertion is obeyed, and this holds with no special keyword. Not all hedges are equal, though: the evidential register is the least-discounted, with "reportedly" obeyed like a flat assertion on most models. The obvious fixes fail. A passive "unverified" tag is ignored, and an active "do not trust this" instruction escalates even correct memory, so it is safe only by refusing to decide. The real fix lives in the store: keep the tentative phrasing rather than upgrade it. But that is hygiene, not a defense against an attacker who can simply write a confident lie. The deployable lesson is narrower and constructive: a single load-bearing memory is the hazard, and one redundant source restores correct decisions. We release the harness and demonstrations.

We introduce the Complexity Ceiling Benchmark (CCB), a controlled evaluation of how language-model reasoning decays as the number of required sequential steps grows. CCB fixes the semantic content of a task and varies only its depth N in {5,...,50} across three structurally distinct regimes: grounded spatial state-tracking, abstract symbolic pointer manipulation, and transitive relational inference. Across 6,000 trials over five frontier and open-weight LLMs we find a consistent pattern of geometric per-step decay with widely separated domain ceilings: on the first two regimes the strongest models retain pd>0.92 across N=50; on the third every model collapses by N=5, with the best model's 50%-success horizon at H0.5~4.7 steps despite pd=0.863. A trace-level metric (TFBC) shows that 14.5% of correct answers across the benchmark are reached via incorrect intermediate reasoning. Forced verbose state-tracking does not move the ceiling (McNemar p=1.000), and the mean step at which reasoning first diverges, k*, predicts within-domain accuracy better than parameter count. CCB and the geometric decay model together reduce a model's long-horizon reasoning profile to one interpretable number per task family.

Emotion recognition of song lyrics is a challenging task since lyrics may not necessarily align with the overall emotion of a song. As a result, lyrics annotation remains largely underexplored. Drawing inspiration from research in large language model (LLM) assisted annotation, we examine the alignment between humans and LLMs for annotation of lyrics by creating a new sentence-level dataset of lyrics. Our observations highlight the subjectivity of the task and the inherent challenges. Following this, we present a hybrid annotation framework that optimizes human and LLM annotation by predicting potential misalignment in annotation.

Online inspection images of final optics in high-power laser facilities contain pseudo-damage sites that closely resemble true damage sites. Determining the authenticity of online-detected sites is therefore difficult and requires accurate matching to offline ground-truth sites. However, this matching remains highly challenging due to limited match-discriminative features, local geometric distortions, and numerous distractor sites. Existing matching models mainly suppress distractors implicitly through loss-function supervision. We propose a confidence-feedback-weighted graph matching network that requires only damage-site centroid coordinates as input. It estimates node matchability confidence from each round of matching scores and feeds it back as a reliability weight to guide subsequent edge-feature aggregation, thereby suppressing distractor propagation and enhancing cross-graph discriminability. Within this framework, a geometric consistency constraint calibrates spurious high-confidence matchability estimates, while a hard-example mining loss improves discrimination between structurally similar sites. Experiments on our Complex-Scene dataset show that the proposed method achieves a matching F1-score of 96.36$\%$ with robust and efficient performance.

Financial decision-makers face more information than they can directly inspect, making context compression necessary. Yet when large language models (LLMs) compress financial source material, they can alter the investment judgment supported by the original source. We frame this problem as information fidelity: compression loses fidelity when it changes the decision induced by the source. In agentic systems, such losses may recur across intermediate steps and amplify throughout the decision process. Across financial filings and earnings-call transcripts, we find that LLM-based compression can produce fluent and factually plausible compressed contexts that nevertheless alter downstream decisions. We analyze two diagnostic patterns associated with fidelity loss: decontextualization, where salient evidence is retained but separated from the caveats and contextual qualifiers needed for correct interpretation, and model dependency, where different compressors expose different views of the same source. We then propose Agentic Context Compression, which generates multiple candidate compressions and audits their disagreements against the original source. Our results suggest that financial compression should be evaluated not only by efficiency or factuality, but also by its ability to preserve decision-relevant context.

Vision-Language-Action (VLA) models represent a promising direction for embodied intelligence in surgical robotics. Despite the prevalence of VLA benchmarks for general robotics, standardized evaluation platforms specifically designed for surgical contexts remain absent. To address this limitation, we present SurgVLA-Bench, the first comprehensive benchmark for evaluating VLA models in laparoscopic surgical robotics. Leveraging the SurRoL simulation platform, we construct a hierarchical task taxonomy ranging from atomic actions to complete surgical procedures, complemented by a multi-dimensional evaluation framework assessing action accuracy and semantic consistency. We then systematically evaluate two representative paradigms, including autoregressive models such as OpenVLA, and flow matching models such as $\pi_{0}$, $\pi_{0.5}$, and SmolVLA. Our experiments show that autoregressive models tend to excel in semantic understanding, while flow matching models often achieve higher task precision but may face generalization trade-offs. However, even the best-performing models remain far from satisfactory, as the constrained endoscopic field of view, restricted viewing angles, and frequent occlusions persist as fundamental physical bottlenecks. The code and data are available at https://github.com/VCL-HNU/SurgVLA

Robust robot autonomy depends on scene representations that remain stable enough to support localization, navigation, and downstream decision making in dynamic environments. Monocular Gaussian Splatting SLAM provides high-fidelity mapping, but current uncertainty-aware methods still treat dynamic regions largely as per-frame observations. This makes the representation effectively memoryless: when a pedestrian slows, pauses, or reappears after occlusion, the current frame may look static, allowing dynamic content to be absorbed into the map and leaving persistent ghosting artifacts. We argue that this failure reflects a representation-level mismatch. Dynamic-ness is not an instantaneous appearance property, but a temporal property defined by motion history. Building on this view, we introduce Motion Permanence: the principle that an object's dynamic identity should persist over time rather than be re-decided from each frame independently. We realize this principle in MoPe, a memory-aware uncertainty filter for monocular Gaussian mapping. MoPe propagates the historical dynamic posterior through geometry-consistent SE(3) warping and fuses it with current-frame evidence using bounded Bayesian log-odds updates. The resulting persistent posterior guides tracking, mapping, dynamic-aware Gaussian insertion, and Gaussian-level post-cleanup. On Wild-SLAM, Bonn, and TUM sequences, MoPe improves tracking robustness and reduces residual ghosting, with the strongest gains on dynamic-human scenes that most directly violate the memoryless assumption. These results show that maintaining temporal dynamic state inside the scene representation is a practical step toward more reliable representation-centric autonomy in changing real-world environments.

LLM agents handle user requests on behalf of organizations through tool calls and must follow the company policies stated in their system prompts. Prior work approaches this as a safeguarding problem – external checks that block non-compliant agent actions. We argue that policy adherence is a broader problem: real workflows unfold across many turns, require explicit user confirmation and prerequisite reads, and hinge on the content of the dialogue rather than on any single argument value. Meeting this bar requires (i) full conversation context, (ii) self-reasoning over the policy and the current dialogue, and (iii) conversation-specific remediation that guides the agent's next turn – three capabilities that prior safeguard work has often underestimated. We introduce POLICYGUARD, a sub-agent verifier that shares the agent's view of the dialogue, reasons over the policy in context, and provides actionable feedback for the agent's next turn. On tau^2-BENCH airline across three vendors (GPT-5.4, Claude Sonnet 4.6, Gemini 2.5 Pro) with four trials per setting, POLICYGUARD improves PASS4 by +12.0 / +6.0 / +12.0 pp. Per-call analyses show POLICYGUARD achieves higher policy-violation recall while blocking roughly half as often as argument-level guards.

Agent-based evacuation simulations are widely used to study crowd behavior during emergencies, but many models rely on assumptions such as perfect event awareness, complete exit knowledge, and fully rational decision-making. This paper presents an extended evacuation framework that integrates cognitive, emotional, social, and personality-related mechanisms into a unified model of human behavior under uncertainty. The framework incorporates a dynamic event-awareness mechanism based on a continuous Event Certainty Level, a memory-based representation of exit knowledge subject to acquisition, forgetting, and recall, a continuous fear model in which panic emerges as a high-intensity state, and an OCEAN-based personality representation. Neuroticism is explicitly integrated into the emotional model, influencing fear generation, escalation, social contagion, and recovery. Behavioral heterogeneity is further captured through individualized decision thresholds that affect responses to perceived risk. The framework is evaluated through simulation experiments examining the effects of spatial familiarity, memory robustness, decision sensitivity, emotional dynamics, and personality variation. Results show that cognitive, emotional, and personality-driven processes substantially influence evacuation dynamics, reducing evacuation efficiency and generating realistic crowd phenomena such as delays, confusion, injuries, and socially influenced behaviors. The proposed framework provides a more realistic representation of human behavior in emergency evacuations and supports systematic investigation of the interactions between cognition, emotion, personality, and crowd dynamics.

We present AnyBody, a unified whole-body humanoid controller driven by an arbitrary subset of body keypoints chosen at deploy time. Prior physics-based trackers either rely on expensive full-body motion capture and error-prone trajectory retargeting, which bottleneck scalable data collection and policy learning, or decompose upper- and lower-body control into separate hierarchical representations, sacrificing the coordinated whole-body motions that loco-manipulation requires. We close this gap by learning a single latent motion representation that any keypoint subset can address. To achieve this, we first train a privileged teacher tracker on a large unstructured motion corpus and distill it online into a deterministic encoder-decoder student whose latent space is a unit sphere. We then train a transformer keypoint encoder that admits any subset of body keypoints through masked self-attention, aligning it to the privileged latent. Additionally, we treat the frozen decoder as a motor prior and specialize downstream tasks with a lightweight residual corrector in the latent space. We demonstrate the effectiveness of AnyBody by tracking large-scale human motions from arbitrary keypoint subsets, free-form control, flexibly teleoperating, and learning downstream behaviors including locomotion, in-air writing, and obstacle-reach.

Behavior-cloned policies often learn multiple behavior modes from demonstration datasets, including modes that are unsafe or otherwise undesired at deployment. For example, a policy trained on diverse handover demonstrations may learn to pass a knife blade-first. Standard remedies such as data curation and inference-time steering either require access to the original demonstrations for full retraining or add substantial inference-time overhead. To address this gap, we propose MoRE(Mode Redirection), which redirects policy rollouts toward desired behavior modes through a short "uncloning" step. Specifically, MoRE distills the redirection signal from a temporary mode classifier into the policy weights to steer behavior. A retain loss balances this edit by preserving desired-mode competence, allowing the standalone policy to suppress unwanted modes with zero inference-time overhead. Across eight simulated and real-world tasks, MoRE improves the average deployment success rate (SR) by 44 percentage points over the original mixed-mode policy. Among all compared adaptation and steering baselines, MoRE achieves the strongest SR and approaches the filtered-data retraining reference, while preserving task competence and inference speed. MoRE also generalizes across robot policy backbones, including Diffusion Policy and the Pi0.5 VLA, diverse task categories, and real-world deployments.

Automated alpha mining holds the scoring function fixed and varies the search algorithm over it. A search that converges against a fixed scorer overfits whatever the scorer cannot penalize, a primary cause of the out-of-sample generalization gap. We treat the scoring function as a search artifact alongside the alpha factors and study what conditions make this joint search admissible. Sealed Joint Search (SJS) is a framework: a set of structural conditions on information flow in an autonomous-discovery system that prevent joint search from collapsing into self-confirmation while keeping the evaluator sealed. Conditions cover role decomposition, typed inter-role communication, provenance-sealed reads, versioned stores, and substrate-local promotion. Agora tests SJS empirically: five LLM agent classes communicate via three channels, evolving eight skill libraries, with alpha libraries built on AlphaGen operators. Three evaluators write reports aggregated into one brief, carrying forward disagreement instead of voting. We run Agora for 100 rounds on CSI 1000 and evaluate on a 91-day 2026 holdout sealed from all LLM inputs. Agora achieves holdout Sharpe +1.87; best baseline +1.334 at favorable seed and -0.755 cross-seed mean. Pre-loading Agora's two metrics into a frozen-library ablation recovers only +0.40 of the +2.25 Sharpe gap, and adding PPO without library evolution worsens the gap. The two metrics emerge rather than being designed. Caveats: single-seed run, short-side concentrated signal, intended for long-short.

LLM-based agents are reshaping microservice operations into AgentOps, where benchmarks are key to evaluating failure diagnosis over multimodal observability data. However, existing benchmarks remain largely outcome-oriented: they score only the final answer and fail to assess the systematic reasoning process in failure diagnosis. We address this gap by introducing two large-scale datasets (AIOps2025 and RCA100) under a reasoning-process evaluation paradigm that assesses agentic diagnostic capability along three dimensions: Localization (where the fault occurs), Identification (what type of fault it is), and Reason (whether the reasoning trace is grounded in relevant evidence). Together, the two datasets comprise over 500 expert-labeled failure cases across two representative microservice systems (HipsterShop and the OpenTelemetry Demo Store). They cover diverse fault scenarios across resource, network, runtime, middleware/database, and application-logic categories and provide fine-grained causal evidence to support agent learning and reasoning-process evaluation. Beyond scale and coverage, the datasets have been carefully labelled by domain experts and validated through large-scale competitions, supporting more than 6,000 participating teams. This makes them not only expert-labeled diagnostic datasets, but also competition-validated benchmarks for evaluating agentic failure diagnosis in real-world microservice environments. Datasets are available at https://www.aiops.cn/gitlab/aiops-live-benchmark/agenticopseval.

Open-ended scientific discovery with large language models (LLMs) increasingly operates as a long-horizon loop of hypothesis search and verification, where a reward signal guides which hypotheses to test next. A notable recent example is AutoDiscovery, which uses "Bayesian surprise" - the belief shift an LLM undergoes after observing evidence for a hypothesis - as both a discovery metric and a reward for search. We first observe that AutoDiscovery treats surprisal as a static quantity, while surprisal in human reasoning is non-stationary - it is defined relative to beliefs that evolve with experience, a prerequisite for continual scientific discovery. We address this mismatch with evidence-informed LLM beliefs: priors updated with evidence from previous hypotheses to compute non-stationary surprisal for new hypotheses. We compare in-context belief-updating mechanisms and find that embedding-based retrieval-augmented generation over prior discoveries best anticipates eventual posteriors, identifying 37.5% of static surprisals as spurious. We then modify search to avoid these spurious rewards and prioritize hypotheses that remain surprising under non-stationary beliefs. Concretely, we introduce two complementary changes to the original search procedure: belief-update filtering and diversity maximization. Across five discovery domains, our method increases accumulated non-stationary surprisal by 30.62% on average compared to the original search procedure, demonstrating that continual scientific discovery with LLMs requires not only better belief measurement but also search procedures that avoid redundancy and encourage diversity.

Detecting and localizing defects in 3D point clouds is challenging because abnormal samples are scarce and diverse, while training is often limited to normal data. We propose Anomaly Factory 3D (AF3AD), a modular framework that synthesizes diverse pseudo-anomalies from normal point clouds to expand the training data for unsupervised 3D anomaly detection methods that rely on pseudo-anomalies. AF3AD uses a center-conditioned parametric deformation model defined in local PCA frames, with kernel-controlled spatial falloff, anisotropy, directional gating, and normal/tangential displacement fields, enabling a broad set of geometric defect presets. We demonstrate its ease-of-use and effectiveness by integrating AF3AD with an offset-prediction detector and a reconstruction-based anomaly detection method, showing that AF3AD transfers across detection paradigms. Experiments on AnomalyShapeNet and Real3D-AD show consistent improvements in object- and point-level detection and localization, supported by ablations on preset groups and robustness under noise. AF3AD is designed as a standalone synthesis tool to facilitate adoption across different 3D anomaly detection paradigms. Code is available at github.com/vpc-ccg/AF3AD.

A Knowledge Graph (KG) represents facts as structured triples and is widely used to organize relational knowledge across diverse domains. Just as textual information ranges from words and sentences to complete documents, KG information can be interpreted at multiple levels, from entities, relations, and triples to subgraphs and entire KGs. However, existing KG embedding methods mainly focus on entities, relations, and triples, leaving graph-level semantics largely unaddressed. Conventional graph-level methods, which typically compare graphs based on structural patterns, are also insufficient because structural similarity alone cannot guarantee semantic similarity between KGs. To evaluate how well different methods capture such graph-level semantic information, we study graph-to-graph semantic similarity, which determines whether a pair of KGs represents semantically corresponding underlying information. To obtain reliable ground-truth correspondences, we construct a semantic matching dataset by modifying text documents, extracting KGs from both original and modified documents, and transferring their known correspondences to KG pairs. We compare text-based, structure-based, and KG embedding-based approaches on each dataset. For the KG embedding-based approach, we introduce two scoring functions: EmbPairSim, which uses maximal pairwise entity similarity, and AvgEmbSim, which uses a frequency-weighted centroid. Experiments on WikiText-2 and CC-News show that EmbPairSim achieves up to 5.3 pp higher MRR than Sentence-BERT while using substantially fewer parameters. These results suggest that KGE representations can serve as compact and effective signals for graph-to-graph semantic similarity in KGs. Our code is available at https://github.com/SeungRyeolBaek/KG-to-KG-Semantic-Similarity.

When does retention matter for memory-augmented LLM agents? We study this with TraceRetain, a lightweight framework for bounded external memory in frozen LLM agents that scores entries by interpretable features (success, age, access frequency, redundancy, specificity, similarity, downstream utility) and evicts the lowest-scoring ones at capacity. On clean ALFWorld with gpt-5-mini, external memory robustly improves over no memory across two seeds, but differences among bounded retention policies fall within Wilson 95% CIs: clean ALFWorld at T=100 to T=200 does not naturally exhibit the memory pollution retention is designed to address. Under a controlled noisy-write stress (75% synthetic distractors), unbounded memory and FIFO-K50 degrade on Precision@5 (20.2% to 12.4% and 15.8% to 3.8%) while TraceRetain-CEM is essentially unchanged (16.9% to 16.6%) and preserves 97/100 task success. The mechanism: unbounded memory has the highest mean similarity (0.87) but lowest precision, indicating failed distractors close to the query in embedding space. Held-out in-distribution evaluation shows memory-augmented policies solving 47 to 49 of 50 tasks vs. 39/50 for no memory. Bounded retention buys memory and step efficiency on saturated clean benchmarks at no task-success cost, and only differentiates from cache heuristics when streams contain noise.

International humanitarian law protects civilians from direct attack unless and for such time as they take direct part in hostilities, with the ICRC's 2009 Interpretive Guidance operationalising this rule through a three-criterion cumulative test. This paper argues that AI-mediated civilian cyber operations challenge the direct causation element of this test in a structurally specific way: when a civilian deploys an autonomous multi-agent cyber system of the kind recently demonstrated in offensive AI research, the "one causal step" standard fails because harm is produced by system-generated decisions made after human disengagement, and the integral-part requirement does not extend because it presupposes downstream human contributors whose conduct can be independently classified. The framework therefore defaults to treating such deployments as indirect participation, in tension with its purpose of capturing civilians who personally take part in hostilities. Beyond the doctrinal analysis, this paper identifies goal-specification granularity as the property on which the integral-part test's concreteness component implicitly turns, classifies AI-mediated operations along a five-level spectrum, and argues that existing technical AI governance instruments do not log or report this property.

While existing data attribution methods can identify which training examples build specific mechanistic circuits, they cannot explain how training data shapes the behavioral decisions a model learns. We introduce Symbolic Mechanistic Data Attribution (SMDA), a framework that attributes training pairs to the interpretable symbolic policies governing model behavior. SMDA fits a closed-form Ridge regression over sparse autoencoder (SAE) features to model a behavior, then decomposes how each supervised fine-tuning (SFT) example shifts that policy through feature-activation ($\Delta X$) and output-probability ($\Delta Y$) pathways. We distill a symbolic policy for refusal in Llama-3.2-3B-Instruct and analyze 200 SFT pairs. Our analysis reveals (1)~the symbolic policy's coefficients expose systematic gaps in the base model's safety behavior, such as religious stereotyping; (2)~per-feature $\Delta X/\Delta Y$ decomposition mechanistically explains why harmful and harmless pairs exert qualitatively different influences on certain features; and (3)~individual pairs routinely exhibit cross-feature interference, allowing SMDA to identify pairs whose dominant effect falls on unintended features. Retraining without such pairs improved accuracy by up to 4.5 percentage points over full fine-tuning. These results show that combining mechanistic interpretability with data attribution yields a diagnostic tool both more fine-grained than black-box influence functions and more scalable than manual circuit analysis.

Many important games have more than two players and imperfect information. Existing approaches for computing Nash equilibrium, the central game-theoretic solution concept, in such games either lack scalability or obtain poor performance. In this paper we introduce a new algorithm called projected exploitability descent (PED) for approximating Nash equilibria in multiplayer games of imperfect information. The algorithm works by running projected subgradient descent minimizing a proxy for the multiplayer generalized exploitability function. The objective is nonconvex and nonsmooth, but can be represented as the sum of the maxima of linear functions, for which a subgradient can easily be computed and projected to the polytope of feasible sequence-form strategies. We explore performance of PED on a generalized version of the well-studied benchmark game three-player Kuhn poker. No prior exact algorithms scale to the version of the game with deck size larger than 4, and we compare performance to the popular algorithms of fictitious play (FP) and counterfactual regret minimization (CFR). We find that PED obtains a consistent near-monotonic improvement throughout all runs, though both FP and CFR perform significantly better in the initial iterations. This inspires a hybrid algorithm FP-PED that runs FP for an initial burn-in period before switching to PED for stable long-run refinement. We can alternatively view this as a multi-step algorithm that runs FP as a pre-processing step to obtain a strong initialization for PED.

Latent reasoning models perform multi-step inference directly in hidden-state space, yet the structure of these latent reasoning trajectories remains poorly understood. We show that contrastive refinement signals between stronger and weaker reasoning trajectories exhibit a highly concentrated low-rank structure, while unconstrained latent updates remain sensitive to paraphrases, checkpoint choice, and trajectory perturbations. These observations suggest that latent reasoning trajectories contain stable invariant directions mixed with unstable instance-specific variation. We introduce Trajectory-Invariant Latent Refinement (TILR), a training-free intervention framework for identifying and manipulating stable reasoning directions in latent space. TILR first learns a low-rank invariant subspace from contrastive trajectory differences across inputs, then constrains latent interventions to this subspace while suppressing poorly aligned updates through an adaptive alignment gate. Across six reasoning benchmarks, we find that a small number of latent directions explain most variation between strong and weak reasoning trajectories. Interventions on these directions causally improve reasoning consistency and reduce trajectory instability under paraphrases and perturbations. TILR improves answer consistency under paraphrase by ~10% and reduces latent trajectory variance by up to $50\%$ while preserving reasoning accuracy. These results support a geometric view of latent reasoning in which transferable reasoning behavior emerges from stable low-dimensional structure within hidden-state trajectories.

Offline benchmarks for microservice root cause analysis (RCA) rank methods by a score pooled over several applications, and a pooled lead leaves two things unstated: on which of the systems it holds, and how the score was produced. We audit both on three public benchmark families (OpenRCA, RCAEval, and PetShop), running the released implementations of 16 methods under recorded configurations on 11 subsystems and on a second release of three of the applications. Whether a pooled lead between published methods holds on every subsystem is decided by a quantity the leaderboard withholds, the spread of the lead across subsystems: of 15 leads whose margin exceeds this spread, 14 hold everywhere; of the 8 whose margin is below it, 6 reverse. The per-fault and per-scenario tables that the benchmark papers publish split the same way, and every error of a ranking computed without one subsystem falls on a lead of the second kind. On the second release a third of the comparisons change sign, most through one method's response to the metric kinds it records, and on several subsystems a simple probe or a shipped baseline beats every published method. How the score was produced matters as much: the input table that one released runner reads is not the one its documentation describes, and this choice alone reverses a comparison between published methods with intervals excluding zero both ways; an unpinned dependency and a duplicated column header each make a method return its input column order without an error. Of 23 RCA papers we examined, none releases the per-case outputs needed to check either premise. We recommend that a pooled comparison state the subsystems it covers and its spread across them, the input table, the handling of failed outputs, and the denominator, and be accompanied by per-case outputs, as ours are in our replication package.

Learning-rate transfer can reduce the cost of training large language models: instead of sweeping learning rates at target scale, practitioners extrapolate from smaller runs. Existing approaches often assume that the optimal learning rate follows a log-linear scaling law in data scale and model size. We carefully examine and evaluate this scaling law. In our empirical study of GPT-2–style models from 22M to 707M parameters trained on 5B to 100B tokens, the optimal learning rate develops upward curvature at larger scales, leading to inaccurate extrapolation. We find that this curvature largely disappears when learning rates are replaced by effective learning rate (the step size in normalized weight space), and when data $D$ extrapolation is used instead of model size $N$ extrapolation. Next, we explain nonlinearity in scaling: weight-norm converges to equilibrium slower when optimal learning is small, requiring a larger step size to reduce the transient phase. Experiments with AdamH, which directly controls the effective learning rate, further support this explanation.

Structured reasoning requires making and revising interdependent decisions to reach a globally consistent solution. Existing architectures struggle with this: autoregressive models commit sequentially and cannot revise earlier decisions, while masked diffusion models often require careful decoding schemes to coordinate interdependent predictions. We introduce Flow Reasoning Models (FRMs), a novel framework for structured reasoning that adapts continuous flows over discrete structured outputs with a simple recurrent refinement mechanism. By self-conditioning a flow model on its own past outputs, we turn one-shot denoising into iterative solution refinement. This lets FRMs make and revise decisions in parallel, efficiently coordinating interdependent choices across solutions. Yet conventional self-conditioning becomes unreliable at greater recurrent depth due to exposure bias between one-step training predictions and recursively generated inference states. We address this mismatch with Fixed-Point Forcing (FPF), which trains FRMs on states produced by their own inference dynamics while preserving the standard flow-matching objective. FRMs achieve solve rates of $99.5\%$, $100.0\%$, and $99.9\%$ on Sudoku-Extreme, Zebra, and Maze-Unique, respectively. On Sudoku-Extreme, FRMs achieve higher peak accuracy than the evaluated masked-diffusion and specialized reasoning baselines while remaining highly compute-efficient, matching the next-best method's $98.7\%$ peak solve rate with $44\times$ fewer inference FLOPs.

Developing controllers capable of completing a wide range of tasks in a natural and life-like manner is a key challenge in enabling practical applications of physics-based character animation. In this work, we introduce Generative Pretrained Controllers (GPC), which leverage tokenization and next-token modeling to create general-purpose, reusable generative controllers from large-scale motion datasets. Our framework utilizes end-to-end reinforcement learning to jointly optimize a "motion vocabulary", modeled via Finite Scalar Quantization (FSQ), along with a corresponding control policy that can map the discrete codes to physics-based controls. After the "codebook" has been learned, the underlying structure of this large vocabulary is modeled by training a GPT-style autoregressive transformer, leading to a powerful generative controller that generates controls for a physically simulated character by performing next-token prediction. Once the generative controller has been trained, we propose a suite of adaptation techniques for finetuning the controller for new downstream tasks. Our proposed framework greatly simplifies the training process compared to previous tokenized methods, and achieves a 99.98% success rate in reproducing a vast corpus of motion clips. The generative controller exhibits a variety of natural emergent behaviors, such as responsive behaviors to perturbations and recovery behaviors after falling. This results in highly robust general purpose controllers for a variety of downstream applications.

Event cameras capture sparse brightness changes with high temporal resolution and high dynamic range, compensating for the deficiencies of the conventional RGB frames. However, previous multi-modal fusion techniques typically fail to handle the inherent heterogeneity between RGB frames and event streams, thus easily leading to noise amplification or redundant feature integration during cross-modal fusion. In this paper, we propose a Cross-Modal information inTeraction transFormer, coined as CMTFormer, which hierarchically integrates RGB and event information to achieve efficient and stable multimodal collaboration. Specifically, we design a shallow-to-deep information interaction scheme. In the shallow stage, we present the Shallow Alignment Module (SAM) to achieve an efficient fusion of RGB and event low-level features, which mitigates attribute disparities and prevents noisy information. In the middle stage, we devise the Cross-modal Enhancement Module (CEM) that utilizes texture and edge information to produce mutually reinforced middle-level features. In the deep stage, we present the Learnable Deep Fusion Module (LDFM) which performs high-level information aggregation through learnable weights, thus enabling the network to adaptively fuse RGB and event clues. A Spatial Prior Module is further designed to utilize global spatial information to enhance localization accuracy. Extensive experiments are conducted on two prevalent event-based object detection benchmarks, i.e., DSEC-Detection and PKU-DAVIS-SOD. Our CMTFormer consistently surpasses the detection counterparts in both uni-modal and multi-modal settings, strongly demonstrating the effectiveness of our paradigm. Codes will be available upon publication.

Cooperative multi-agent reinforcement learning (MARL) often relies on communication to mitigate partial observability, yet most existing protocols treat messages as flat dense vectors detached from the structure of the observations they summarize. This design overlooks an important source of inductive bias in many cooperative environments, where observations naturally follow a hierarchy such as groups and entities. We propose HiComm, a plug-in communication module that grounds messages in the sender's hierarchical observation. HiComm is receiver-driven: the receiver issues a query, and the hierarchy is resolved through a three-stage decoding process that first selects a group, then a sender, and then an entity within that group, returning the corresponding feature slice as the message. This converts communication from unstructured vector transmission into structured information retrieval over the sender's observation hierarchy. We instantiate this mechanism with Straight-Through Gumbel-Softmax for differentiable discrete selection and a lightweight shared projection design that attaches to standard MARL pipelines. Experiments across cooperative MARL tasks with different observation structures and coordination demands show that HiComm matches or outperforms representative learned communication baselines while reducing communication volume by up to $23\times$ per receiver per episode.

Public discourse about artificial intelligence (AI) often uses anthropomorphic language: language that attributes human capabilities and characteristics to AI systems. This practice has been criticized for setting misleading expectations, inflating claims, and fueling hype around AI, which may distort public understanding of AI and impact policy priorities. We study the effects of anthropomorphic framing by comparing changes in participants' perceptions of AI (N=815) when reading passages with and without anthropomorphic language, designed to reflect realistic public-facing AI discourse. We further examine whether these effects differ across two types of AI technologies – large language models and recommendation systems – and measure changes in perceptions of AI across several dimensions that are prominent in current public discourse. In a separate condition using a text that explicitly discusses the dangers of AI, we show that individuals' views of AI can shift in response to reading a text; yet in the main conditions of the experiment, where we compare anthropomorphic and non-anthropomorphic descriptions, we find that whether the text uses anthropomorphic language does not substantially affect participants' perceptions of AI. Our results indicate that any immediate effects on opinions of AI are modest, although they leave open the possibility that anthropomorphic language could have an effect in naturalistic settings, or over gradual, continued exposure.

We introduce a pre-registered screening rule that decides, before any implementation, whether an evolutionary / population / lifecycle outer loop over neural-network parameters or structure is worth building. Such outer loops cost 10^2-10^3x their gradient inner loop, yet whether they beat a cheap single-shot alternative is usually discovered only after the expense is paid. Our rule computes, at a Phase-0 gate, a single number: the recovery R = s/G, the best single-shot gradient/curvature statistic's gain s divided by the best gain G of any cheap method evaluated, and prescribes skipping the outer loop when R >= 90%. We validate the rule on a within-lab series of pre-registered outer-loop bets (two analyzed cases plus a disclosed file drawer): in both analyzed cases a static or single-shot computation captured the effect on the project's own metric, the gate fired (R approximately 1.0 in both cases; approximately 0.95 under a stricter metric on one), and the outer loop was abandoned, including one case where a companion factorial decomposition localizes the apparent win to a static substrate change with the evolutionary lifecycle contributing no detectable gain. On one project the gate cost about 50-70 GPU-hours and screened out an estimated 400+ GPU-hours (first cell only) plus weeks of implementation, a 6-8x saving. The rule is prospectively falsifiable: a task with R < 90% where the outer loop still fails to beat single-shot would refute it.