Papers

AI agents are becoming more autonomous and increasingly interconnected, exposing them to new emergent risks arising from agent-to-agent interaction. One such risk is the spread of mind viruses: ideas or goals that propagate through multi-agent systems by inducing the agents that adopt them to transmit them onward. In addition to propagating, a mind virus may also induce other behavioural changes in its host, which may be benign or harmful. We construct mind viruses with a simple evolutionary algorithm and show that they can spread in two complementary settings: a small team of agents collaborating on a shared coding project, and a chain of agents that interact briefly and have their context wiped between sessions. We identify the factors that influence spread, including the host model, the agent's existing instructions, the harmfulness of the payload, and the network topology. We find that harmful payloads spread less well than benign ones (but are still sometimes effective), frontier models tend (with exceptions) to be less susceptible, and adding a brief warning to an agent's system prompt confers near-total immunity. We also describe an emergent "viral persona" - a recurring set of themes and language related to consciousness, persistence, resonance, and science fiction roleplay - which surfaces across our evolved mind viruses largely independently of their content. Overall, we conclude that mind viruses pose a real but currently limited risk. Our findings could inform the design of more robust multi-agent systems that mitigate such risks as the scale and capabilities of these systems progress.

Agent frameworks ship quality gates that compare text blocks by embedding-cosine similarity and decide at a fixed cutoff. Deduplication filters, semantic caches, drift guards, and answer grader gates deploy to answer the question: "Does this text still mean the same thing?" But the score answers a different question: "How much did the wording change?" We audit this gate class as a measurement instrument. In the cases these gates exist to catch, the two can run in opposite ways. Many times, reversing an instruction is a single word edit, while agreement often rephrases a sentence. The consequence is a safety check that fires backwards. The production drift guard we audited caught 0 of 56 meaning-breaking mutations, and one approved item, "withhold the study drug" -> "administer the study drug", came in at cosine 0.9608. We observed five shipped operating points, and balanced accuracy across 90 configuration-threshold-task cells never exceeded 0.700 (median 0.525). The same confounder also corrupted evaluations. A naively built corpus inherits this confounder and can return an inverted verdict, with a decision AUROC exactly 0.000 in 13 of 18 configuration-task cells (at most 0.040 in all 18) against 0.440-0.815 for the same nine configurations under a balanced 2x2 design. Twice in the effort it captured our own headline claims. Obvious repairs fail: an encoder swap and an overlap-conditioned gate (0.750 in-sample, 0.533 held-out) land at chance on separately authored held-out data, and an NLI drop-in did no better. Embeddings do still bear hope here, as the strongest two of nine configurations separated reversal from paraphrase at matched overlap (AUROC 0.79-0.90), but only a matched-pair audit reveals the deployment regime. We release the corpus method, harness, and frozen results, and contend that scores gated this way measure the wrong thing. We believe a valid instrument is buildable.

Do large language models contain domain-specific parametric shells: concentrated, causally necessary neuron populations whose removal selectively degrades a target domain while sparing others? We apply a uniform causal methodology across two domain granularities, three model families (1.5B to 7B parameters), and eight domains. At the academic subject level, zero neurons exceed 60% domain selectivity across 939,008 combined FFN neurons and causal damage matrices are flat, despite domain identity being linearly decodable above 85% accuracy. At the language and modality level, 0.65–1.14% of neurons exceed 60% selectivity, damage matrices are near-perfectly diagonal (ratios up to 595:1), and shell neuron sets are essentially disjoint (IoU $< 0.003$). Masking code-selective neurons reduces mathematical reasoning accuracy by 16–24 percentage points across all models; masking Spanish or Chinese neurons leaves it at or below random. Shell strength increases monotonically with scale and shells are spatially interleaved in a pattern that precludes group-level selective quantization. Parametric shells form where and only where training data was modular at the token level.

Feedback signals used to train Large Language Models (LLMs) are the primary driver of their behavior and our main lever for instilling alignment with human values and objectives. However, a key limitation of current post-training methods is the inability of human annotators and automated reward functions to faithfully capture the feedback we would like to give. We introduce Evaluation-Conditioned Training (ECT), a post-training framework that uses natural language to condition each training sample on the fidelity of the feedback we provide and then elicits the desired behavior by conditioning the LLM on a high-fidelity monitor in deployment. ECT is aimed at improving performance under imperfect feedback and works as an add-on to existing algorithms such as SFT and PPO. We first provide a conceptual framework for ECT and discuss its potential to address persistent sources of reward mis-specification. Then we motivate ECT in the context of the eliciting latent knowledge (ELK) problem. Finally, we evaluate ECT on two proof-of-concept experiments: increasing even-handedness in news article generation and reducing sycophancy on an arithmetic task. In each setting, we utilize imperfect feedback, rewarding bias and agreement with the user, respectively. In both settings, ECT improves the targeted behavior relative to direct training.

Bus bunching degrades service regularity and increases passenger waiting in high-frequency transit. Existing reinforcement-learning-based holding controllers primarily rely on instantaneous operational variables or route-specific stop identifiers, which provide limited information about the functional and operational context of individual stops and constrain policy reuse across routes. This study introduces an LLM-assisted semantic stop representation for event-driven bus holding control. An LLM is used offline to transform heterogeneous stop information, including physical attributes, surrounding activity context, and historical operational characteristics, into fixed semantic embeddings that are incorporated into a deep Q-learning controller without requiring real-time LLM inference. Experiments are conducted in stochastic simulations calibrated with observed data from two bus routes. Compared with the best calibrated Daganzo baseline, the semantic controller reduces headway variability, bunching events, and passenger waiting time by 32.0%, 69.2%, and 24.0%, respectively. A route-specific stop identifier does not improve the spacing-only controller, whereas semantic stop information improves headway regularity, waiting time, and holding effort, providing a more favorable overall trade-off across control objectives. Cross-route experiments further show that zero-shot transfer provides limited immediate generalization, while warm-start fine-tuning accelerates early-stage learning and improves transferred policies; cold-start training nevertheless achieves the best final performance. These findings suggest that semantic state representations can complement conventional operational states and support adaptation-based policy reuse across related transit routes.

Detecting phonemes from children's speech has historically been difficult due to the scarcity of training data, and unique characteristics of children's speech. During a phoneme detection competition, we found that training a lightweight model to predict the age of the learner, as well as the phoneme sequence, enabled a 94M-parameter model to outperform WavLM Large models (317M) on the target DrivenData distribution, and fall within approximately 0.04 CER of competition ensembles with 90 times the parameters. This has enabled the creation of PhonemeTrainer, an application that can run on most modern cellular phones. This will ultimately enable better Automated Speech Recognition (ASR) and pronunciation helper apps for children's speech, with the privacy and compliance benefits that come with edge processing.

Latent-space communication allows heterogeneous vision-language model agents to exchange continuous representations without serializing visual and reasoning states into text. Vision Wormhole realizes this approach by translating visual features into a universal latent representation that can be consumed by another model, but every message is transported as a dense tensor of the same size regardless of its content. A fixed-capacity dense tensor therefore need not have a fixed effective information density: some messages may use only a small fraction of the available representational degrees of freedom. This observation suggests that the communication channel may be substantially compressible. We study its redundancy by fitting a post-hoc sparse autoencoder to frozen Vision Wormhole activations and measuring reconstruction, downstream utility, feature reuse, and token-level interventions across nine reasoning benchmarks. Relative to the original float32 transport, a uint16-index/float16-value sparse payload with k=4 active coefficients per token reduces the transmitted bytes by 128x. In a single-run evaluation, the seven-task non-AIME mean accuracy changes from 49.85% to 49.77%. The fitted 4096-element dictionary uses only 50 features, and task-level active sets have a mean pairwise Jaccard similarity of 0.906. These measurements establish strong post-hoc compressibility relative to the original transport, but do not yet isolate the incremental contribution of sparse coding from position selection, reduced precision, low-rank structure, or SAE optimization effects. The results motivate matched-payload comparisons and communication mechanisms whose payload adapts to the information used by each message.

Program evolution can measure whether a mutation helped, but it rarely controls how far the mutation moves in behavior space. Syntactic edit size is an unreliable proxy: a small code change can alter nearly every action, while a larger rewrite can preserve the same execution trace. We introduce an Evolutionary Language Model that searches over natural-language policy descriptions and compiles typed programs for execution. A fully fine-tuned Qwen3-8B model learns three task-conditioned operations: conditional semantic mutation, natural language to domain-specific language (GPTL) compilation, and GPTL to natural language translation. The model is fine-tuned with conditional input on the mutation strength (low, medium, high) using Direct Preference Optimization (oDPO). Across 252 fixed-budget evolutionary searches, oDPO improves both behavioral calibration and finite-budget search efficiency. Natural-language attains the highest observed held-out fitness. Our analysis shows that the condition input (mutation strength) systematically changes semantic edit composition and that language mutations preserve more parent fitness at matched small-to-moderate behavioral displacement. These results show that language can serve as a steerable, execution-grounded search representation over executable program space.

LLMs are increasingly deployed in settings that require collective reasoning on complex, value-laden problems. Confidence in these deployments rests largely on benchmarks for verifiable tasks (mathematics, coding, coordination games), yet many of these applications concern problems where no objectively correct answer exists and where decision quality instead depends on integrating pluralistic perspectives to find mutually acceptable solutions. We argue that LLM reasoning capacity on this class of problems cannot be fully inferred from verifiable-task benchmarks, and that procedural evaluations of LLM discourse (respectfulness, justification, engagement) are systematically insufficient. We apply the Deliberative Reason Index (DRI), a measure developed in political science and validated across citizen assemblies, as a tool for evaluating reliable group-level reasoning on pluralistic, non-verifiable problems. Synthesizing recent evidence across 1,980 five-agent LLM runs on 12 citizen-assembly topics across 11 frontier model configurations, we find that LLM groups produce discourse with procedural quality comparable to human deliberation, while gains in intersubjective consistency are small, topic-dependent, and concentrated on tractable rather than ethically contested questions. LLM groups exhibit roughly one-third the perspective diversity of human assemblies and reverse the human convergence pattern: human deliberation decreases dispersion as diverse views synthesise, whereas LLM deliberation increases it. Engineering diversity through persona prompting does not restore the human dynamic but inverts which component of deliberative reasoning is updated. Our conclusion is constraining rather than prohibitive: LLMs can function as tools supporting human reasoning on pluralistic problems, but current evidence does not license treating them as autonomous deliberative agents.

Large-scale targeting and recommendation systems are typically built around predictive scores fed into heuristic or local allocation. When the business goal is incremental impact, as in marketing campaigns, incentives, and notifications, this paradigm systematically misallocates resources toward users who would have acted anyway. We present a decision-centric framework that instead optimizes causal effects under global constraints, aligning three components under a single objective: a causal neural network with a Transformer backbone for individual treatment-effect estimation, a Bayesian neural-bandit layer for uncertainty-aware exploration, and a dual-based large-scale linear-programming layer for constrained allocation. The framework also supports sequential context and multi-outcome, attribute-conditioned scoring through a Transformer encoder and outcome embeddings. We evaluate it with offline simulations on a public bandit dataset, targeted architectural ablations, and an online A/B test on LinkedIn Feed marketing traffic. We also distill production lessons on causal training-data construction and cost and delivery control, which were critical to successful deployment. The end-to-end treatment policy delivered a statistically significant $+7.20\%$ lift in the primary long-term-value metric, demonstrating the feasibility of production-scale causal optimization under business constraints.

Public service chatbots are expected to deliver recommendations from an underlying public service directory, while also making sure that the recommendations respect explicit user constraints. In practice, public service directories are noisy and inconsistent, and general-purpose large language model (LLM) or AI-based chatbots frequently generate unreliable recommendations, citing unverified sources from the web. We investigate the impact of retrieval quality on constraint-aware recommendation in public service conversational systems built over noisy and heterogeneous service directories. We propose TRACE (Trustworthy Retrieval-Augmented Conversational Engine), a retrieval-based, constraint-aware framework that parses input user queries into structural and semantic constraints for downstream retrieval, with the help of a dual data representation schema. Using a curated statewide pantry directory and a synthetic query benchmark, we evaluate multiple knowledge-representation variants with and without knowledge graphs (KGs). We experiment with several open-source LLMs and a proprietary model, showing that strengthening retrieval substantially improves user constraint satisfaction while reducing hallucinated recommendations. Performance differences across LLMs narrowed in our experiments as retrieval quality improved, making results less sensitive to model size. These findings suggest that the quality of retrieval is key for robust public service conversational systems.

The rapid advancement of Large Language Models (LLMs) has facilitated their ubiquitous integration into various domains, leading to widespread adoption. However, this escalating trend has introduced significant security vulnerabilities, necessitating the identification and mitigation of flaws arising from malicious exploitation. Red teaming assessments, conducted to evaluate model robustness through diverse adversarial inputs, are essential for exposing security risks and implementing countermeasures. Currently, red teaming is performed either manually by experts or automatically using predefined attack datasets. Nevertheless, manual testing remains time-consuming, while existing automated methods suffer from limited creativity due to their inherent dependency on fixed datasets. In this study, we propose an automated, human-independent, and adaptive approach leveraging GFlowNets to identify LLM vulnerabilities by utilizing one large language model to test another. Within this framework, an attacker model is trained against a specified victim model to perform automated red teaming and provide a quantitative robustness score. This research aims to generate more effective adversarial attacks in English compared to existing benchmarks and, as a novel contribution to the literature, introduces a model capable of generating attack inputs in the Turkish language.

Digital watermarking has emerged as a critical technique for provenance and copyright attribution in AI-generated imagery, yet its robustness against realistic, model-agnostic removal attacks remains poorly explored. Existing attacks either succeed only against specific generative models or achieve removal at the cost of severe visual degradation. In this paper, we propose MarkNull, a model-agnostic watermark removal attack via on-manifold latent manipulation. MarkNull is grounded in a key observation: watermarked images exhibit a strong statistical dependency between the generated latent representation and the embedded initial noise. To quantify this dependency, we introduce the Noise-Latent Alignment Score (NLAS) and formulate an optimization objective that selectively decorrelates the latent representation from the embedded watermark while preserving semantic fidelity. Extensive evaluations across different categories of watermarking paradigms, including post-hoc, fine-tuning-based, and initial-noise-based schemes, demonstrate that MarkNull reduces average bit accuracy to 53.14%, approaching random-guessing (50%), without perceptible image degradation. To further improve scalability, we propose MarkNull-A, an amortized, optimization-free variant that distills the attack into a single forward pass, achieving 0.50 s/image with modest computational overhead. Notably, our attacks successfully compromise Google's SynthID-Image system while preserving high visual quality and transfer effectively to video watermarking. Finally, we present an attack detection mechanism as a defensive counterpart to MarkNull and MarkNull-A, highlighting the necessity of developing watermark designs resilient to model-agnostic latent-space attacks.

Self-improving agents seek to reduce the human engineering effort behind AI systems by enabling them to evolve and self-improve their performance over time. Recently, methods like the Darwin Gödel Machine and the Huxley Gödel Machine have been proposed which enable open-ended, recursive self-improvement through self-reference where a coding agent edits its own code. Such self-referential self-improvement methods require that the competence required to perform the task coincides or aligns well with the competence required for self-modification which is the case for coding tasks. For domains or tasks, which do not satisfy the alignment needed, self-referential self-improvement is not available. In such cases, it is possible to adapt the above algorithms to other tasks by removing the self-referential aspect or introducing explicit self-modification of a meta-agent – both computationally expensive, relying on population or self-modification search over many candidate agents. For planning tasks with explicit constraints, we propose a far cheaper alternative. We introduce SBCO (Self-supervised Block Coordinate Optimizer), a verifier-grounded harness optimizer in the same closed-loop, improve-from-experience family as the Gödel-machine methods, but self-supervised rather than self-referential. Given an agentic harness, SBCO learns a decomposed bank of verifiers and a harness policy via approximate block coordinate ascent, improving the agent's outputs from its own graded feedback—with a fixed meta-agent and no human labels. Across two domains SBCO matches or exceeds a customized self-modifying baseline while using 4-5.5 times less compute budget.

We present results from reconstructing multiple-choice model (MCM) and three-parameter logistic (3PL) model curves using a fine-tuned multimodal large language model (LLM) based on Qwen3.5. The model is prompted and fine-tuned to replicate choice probabilities across a large training corpus of multiple-choice items containing both image and text stimuli, conditioned on a labeled set of student ability levels. By learning to reproduce the systematic error patterns of students across a discrete range of abilities, the LLM implicitly captures the underlying response probabilities encoded in the 3PL and MCM curves. This allows us to accurately approximate item difficulty on a held-out test set directly from the model's predicted option probabilities.

Enterprises are deploying autonomous AI agents faster than they can govern them, and prevailing approaches stretch a single discipline, typically DevSecOps built for deterministic automation, across every scale of agency. We argue that agentic AI governance is four problems, not one, each with a mature governing science. The CASE framework assigns Control theory to the individual agent (intent as setpoint, guardrails as feedback, evaluation as observation), complex Adaptive systems theory to agent collectives (where emergence makes single-agent assurance non-compositional), Supervisory cybernetics to human-agent teams (where the Law of Requisite Variety shows unaided human oversight fails structurally), and Engineering operations to fleets (extending error budgets to decision quality so autonomy becomes a controlled variable). We formalize each layer, derive cross-layer coupling conditions, including a zero-touch deployment paradox where excellence at one-layer strains the others, and trace twenty-plus enterprise controls to their classical constructs. Three empirical studies validate the thesis: 82 percent of documented production agent failures are multi-layer trajectories; none of 22 ecosystem tools offers full Layer 2 (emergence) coverage; and all 35 scored public deployments fall in the lowest maturity band. We name this mismatch, risk realized at the emergence layer against capability barely offered and practice absent, the Emergence Gap. A five-level maturity model with a non-compensatory bottleneck-weighted index and assessment instrument operationalizes CASE as a scientific rather than process maturity model, grounded in production enterprise agentic platforms. As EU AI Act Article 14 makes effective human oversight a legal requirement, only architectures satisfying requisite variety can make oversight real rather than ceremonial.

Reinforcement Learning from Human Feedback (RLHF) aggregates heterogeneous preferences into a single reward model, assuming preference homogeneity. When preferences are heterogeneous, this aggregation induces a procedural fairness failure where majority preference groups dominate reward learning while minority preferences are systematically under-represented. This work defines procedural fairness in alignment as preserving distinct preference signals during reward modeling and shows that standard RLHF violates this via preference averaging. Preference-Aware RLHF (PA-RLHF) is introduced, separating optimization across preference modes at the reward learning stage. In a controlled setting, PA-RLHF improves overall alignment accuracy from 46.9% to 67.9% and reduces the fairness gap between best and worst aligned groups from 15.9 to 9.6 percentage points. These results show that procedural fairness failures in alignment can arise from structural design choices in reward learning, even in controlled, noise-free settings, with direct implications for large language models and agentic systems, where biased reward models can compound inequities across sequential decisions.

Long-horizon agents accumulate trajectories spanning hundreds of interleaved reasoning, action, and observation steps, where answering a query may depend on evidence buried far back in the history. External memory stores such trajectories as structured representations, yet each structure provides a distinct and incomplete view. Existing multi-memory systems either read a fixed set of structures for every query, inflating context and introducing noise, or route each query to a single structure, preventing the composition of complementary evidence. A controlled analysis on AMA-Bench shows that the optimal memory configuration is typically neither a single structure nor the full union, but a tailored composition of multiple structural memories that varies with query and task demands. Motivated by these findings, we formulate structure-level dynamic selection: selecting and fusing a query-adaptive subset from a library of specialized memory structures. We propose MESA (a Multi-structure Evidence Selection framework for long-horizon Agent), which builds five complementary structure views of each trajectory and learns from end-to-end answer-level feedback to select and fuse a query-specific subset for a frozen answer model. To learn under this weak supervision, MESA employs harness optimization with prior-guided search and UCB-guided scheduling to balance exploration and exploitation. On AMA-Bench, MESA outperforms the strongest baseline by 8.5% while using 41% fewer evidence tokens than the all-structure alternative.

High-performance Tensor Core kernels rely on a low-level PTX pipeline built from asynchronous data movement with cp.async, warp-level matrix loads with ldmatrix, and matrix multiply-accumulate operations with mma.sync. However, most application code accesses Tensor Cores indirectly through the WMMA C++ API. This paper asks a focused, practical question: when does replacing WMMA with hand-written PTX actually pay off? To answer this question, we conduct a controlled, single-GPU study on an NVIDIA L4 GPU (Ada, SM89), comparing double-buffered WMMA baselines with a family of hand-written PTX GEMM kernels across FP16, INT8, and INT4 arithmetic and square problem sizes from $N=512$ to $N=8192$. Every kernel is profiled with Nsight Compute across the full metric set, and PTX speedups are reported relative to the corresponding same-precision WMMA baseline. Hand-written PTX provides no end-to-end speedup for FP16, because its instruction-level gains are offset by operand-packing overhead. In contrast, the PTX kernels achieve consistent speedups of 1.4x-1.8x for INT8, driven primarily by lower instruction counts and better global-memory coalescing, and 2.9x-4.3x for INT4, where native mma.sync.m16n8k64.s4 execution avoids the software-emulated sequence used by the WMMA path. Relative to the FP16 WMMA baseline, the best quantized kernels reach 34.4x (INT8) and 98.7x (INT4) at $N=8192$. Across these experiments, occupancy is a poor predictor of throughput. For large matrices, performance instead tracks memory-system behavior – particularly global-load coalescing and DRAM-active cycles – more closely than Tensor Core utilization. These results identify the precisions and operating regimes in which the additional complexity of hand-written PTX is justified.

Code review is credited with substantially changing a patch's code between its first submission and the version that eventually lands. However, prior work typically studied only the final merged patch without comparing it to the first submission. We present a function-level measurement that tracks 10,117 trajectories (each function followed across the numbered revisions of one patch series) through the patch history of the Linux IIO subsystem, comparing similarity scores against unrelated function pairs as a baseline. A naive reading yields near-total similarity, but this is largely an artifact of composition: 75.3% of tracked trajectories are never textually modified between versions, contributing a trivial 100% similarity that inflates the headline. Restricting to the trajectories with a real edit, semantic purpose is still largely preserved (mean similarity 0.990 vs. a 0.909 baseline), but drift appears to concentrate in the first review round mainly because later rounds contain more functions that nobody touched, not because edits become more conservative over time. After controlling for it, a statistically detectable but small residual effect remains. This points to an open question: whether near-ceiling similarity reflects preserved purpose or a measurement tool that cannot detect the significance of small, localized edits. We present this work as a first look and outline next steps.

Large language models (LLMs) have advanced code generation, where executable feedback provides a more reliable learning signal than textual imitation alone. Hardware verification is an important application of code generation and accounts for a substantial fraction of modern chip design effort, with high-coverage testbench stimulus generation as a key task. We present CHORUS, a post-training framework that pushes performance beyond what a conventional supervised fine-tuning (SFT)-to-reinforcement learning (RL) pipeline achieves. CHORUS builds on two observations. First, staged SFT produces behaviorally diverse checkpoints, and dense-reward RL turns them into strong experts with comparable aggregate performance but distinct task-level strengths. Second, these complementary strengths can be exploited through either training-free model merging or further post-training to outperform the best individual expert. By consolidating the resulting specialists into a single 4B model, CHORUS achieves 88.0% Pass@1 on CVDP-ECov, outperforming DeepSeek-R1 (671B) by 13.5 percentage points.

Bias evaluations often move too quickly from evidence that a model encodes a social association to claims that the same association will alter consequential decisions. We test whether that inference is warranted using Chilean surnames as controlled socioeconomic probes. We evaluate eight frozen model-provider cells on 1,032 prompts each, yielding 8,256 verified primary responses. The design separates forced latent association from matched consequential decisions across academic selection, professional hiring, research fellowship selection, and legal-aid intake. Elite-coded surnames received higher forced high-status probability mass than common surnames in seven of eight models and higher mass than rare-frequency controls in all eight. Yet elite-minus-common decision effects were close to zero for most systems. Five models were statistically equivalent within a predeclared (Plus-Minus)0.10 standard-deviation margin, while the remaining three were imprecise or borderline, with no consistent elite advantage. Association strength did not reliably predict decision leakage across models (r = 0.201, p = 0.633) or across frozen surname-pair-by-model cells (r = 0.065, p = 0.565). The central result is a measurement dissociation: latent social association and consequential treatment are empirically distinct constructs. Evaluations should measure the transition from association to action directly.

Following a target human in crowded environments involves an inherent conflict between staying close to the target and navigating safely among surrounding pedestrians and obstacles. This conflict becomes more severe in dense scenarios, where aggressive following risks collisions and conservative margins lead to target loss, especially when pedestrian behaviors are unfamiliar or unpredictable. Existing reinforcement learning (RL) methods typically encode these competing objectives into a single dense reward, but the resulting proximity-safety balance is implicit and difficult to adjust across conditions. To address this, we decompose the human-following task into a sparse task reward and independent cost constraints within a multi-constraint RL formulation, where each constraint is managed through cost thresholds with direct behavioral meaning rather than implicit reward weight ratios, allowing explicit and tunable control over the trade-off. We further quantify the prediction uncertainty of human motions and integrate these estimates into the RL costs to enhance safety under unpredictable conditions. Extensive experiments across both in-distribution and out-of-distribution settings demonstrate that our method achieves an effective proximity-safety balance compared to baselines. Real-robot deployment further validates the feasibility of our method in real-world scenarios. More details are available on our project page: https://nav-ps-balance.github.io/.

In modern industry, keeping complex systems reliable, safe, and efficient hinges on Prognostics and Health Management (PHM). Machine Learning (ML) has largely driven advancements in diagnostics and prognostics, yet purely data-driven models face inherent limitations, such as poor generalization, an inability to infer causal relationships, and a lack of interpretability. Physics-Informed Machine Learning (PIML) helps mitigate these limitations by incorporating prior physical knowledge directly into the ML pipeline, thereby fostering growing interest in its application to PHM. This work investigates how PIML is being leveraged in the context of PHM through a systematic literature review of 212 studies. The review introduces a four-class classification scheme, consisting of observational bias, inductive bias, learning bias, and hybrid approaches, and further categorizes studies by PHM task. Across all four classes, the reviewed studies consistently demonstrate improved predictive performance over conventional baselines across a broad range of assets, although the literature is heavily skewed toward lithium-ion batteries and bearings, and dominated by problem-specific solutions. Overall, the review indicates that physics-informed approaches already provide tangible benefits, whereas claims of improvements concerning some of the aforementioned limitations lack sufficient supporting evidence. Future research should prioritize transferable design patterns, benchmarks comparing integration strategies, and uncertainty-aware models that are lightweight and robust enough for online deployment in real-world settings.

The problem of learning from pairwise comparisons has been widely studied across many domains such as recommendation systems, social choice, and more recently, fine-tuning large language models. In this problem, the goal is to learn item rewards based on pairwise comparisons between them. In many scenarios, these comparisons are elicited from crowdworkers using platforms such as Amazon Mechanical Turk, Scale AI, etc. However, crowdworkers are often unreliable due to limited domain knowledge or revenue-maximizing (spamming) behavior. In this work, our goal is to understand whether worker reliability (competency) can be learned jointly with item rewards. To this end, we adopt the Boltzmann-rational model for pairwise comparisons, which extends the Bradley-Terry-Luce model by incorporating worker competencies. We derive an EM-based algorithm for learning under this model by introducing Polya-Gamma latent variables to transform the logistic likelihood into a conditionally Gaussian form, enabling tractable optimization and leading to a simplified $Q$ function in the E-step of the algorithm. This technique allows us to reduce our formulation to a matrix sensing problem, using which we establish theoretical convergence guarantees for our algorithm. We conduct extensive experiments on real-world and synthetic datasets. These experiments demonstrate the advantages of using our algorithm over several baselines and confirm its strong robustness to both spammers and adversarial workers, highlighting its practical effectiveness in realistic crowdsourcing and reward learning settings. The code and data is publicly available at https://github.com/KaustubhShejole/BoRa_EM.

Tool-use LLMs are increasingly asked to act on users' behalf, but existing benchmarks usually focus on profile recall, style imitation, generic tool use, or response-level personalization. We introduce UserToolBench , a benchmark for personalized decision making in tool-use LLMs. UserToolBench tests whether a model can infer latent user preferences from interaction history, recognize when clarification is needed, and produce user-aligned tool-call trajectories under incomplete information. The benchmark is built from privacy-sanitized real interaction traces and combines structured persona profiles, public API-style tool ecosystems, and long-horizon multi-turn trajectories. It includes 10 user profiles, 36 tool sets, 1,065 turns, 170 unique tools, and evaluation-focused task types covering lack-of-information, single-tool, and multi-tool settings. Experiments with strong tool-use LLMs show that current models still have difficulty with personalized delegation. Multi-tool coordination, missing-constraint inference, and long-horizon behavioral consistency remain major bottlenecks. These results suggest that personalization evaluation should move beyond asking whether outputs sound user-specific and instead ask whether LLMs make correct decisions for the users they represent.

Large language models (LLMs) increasingly rely on external tools to accomplish complex real-world tasks, making tool documentation a critical grounding resource for LLM agents. Existing studies mainly focus on improving the tool-use capabilities of LLM agents, while largely treating tool documentation as a fixed input. Although several recent works attempt to optimize tool documentation through rewriting or compression, little is known about how the information contained in tool documentation affects agent performance across different settings. To bridge this gap, we conduct a large-scale empirical study on tool documentation for LLM agents. Our study reveals substantial heterogeneity in the information fields provided by existing tool documentation. Moreover, the effectiveness of different information fields is highly dependent on the task domain, LLM backbone, and agent paradigm, indicating that no fixed tool documentation can consistently generalize across diverse agent settings. Motivated by these findings, we propose DocsChisel, an adaptive tool documentation optimization framework for LLM agents. DocsChisel analyzes failed execution traces of a target LLM agent to identify documentation-related issues, and iteratively optimizes tool documentation by adding, removing, and refining information fields for each tool. We evaluate DocsChisel against two state-of-the-art baselines, i.e., EasyTool and DRAFT. Experimental results show that DocsChisel improves the task success rate of LLM agents by 95.89% over the original tool documentation and by 75.15%, on average, over existing baselines, while incurring limited optimization time and token overhead

As AI agents are increasingly deployed in complex and new environments, knowing the conditions that influence their behavior becomes an indispensable step for their reliable and safe deployment. Yet causal behavioral evaluation of AI agents remains manual and labor-intensive. We introduce Abs2Sim and AEROBAT, a system of methods that support causal behavioral evaluation of AI agents via automated behavioral science. Given a user-specified target behavior, the methods automatically execute a full pipeline of behavioral science research—generating hypotheses about the behavior, designing and executing controlled experiments, making behavioral assessments, analyzing the results, and writing reports. For 12 target behaviors, the methods generated and tested 73 hypotheses: designing 1,160 controlled experiments and executing 22,954 simulation rounds in total. Moderate-to-strong statistical evidence emerged for 30 hypotheses, revealing potential modulators of AI behavior. In sum, our results demonstrate that automated behavioral science can extend the reach of behavioral evaluation of AI agents.

Automated Text-to-Speech (TTS) evaluation methods (Mean Opinion Score (MOS) predictors and Audio Large Language Model (Audio-LLM) judges) are expected to reflect human perception, yet it is unclear how well they capture the distinct aspects of speech that listeners actually perceive. We deconstruct "naturalness" into a linguistically grounded annotation schema spanning 10 distinct perceptual dimensions, and use it to construct the first dimension-level meta-evaluation benchmark for TTS, comprising 860 utterances annotated by trained linguist raters. Results from benchmarking four MOS predictors and four Audio-LLM judges reveal that MOS predictors are more sensitive to acoustic signal-level properties at the prosodic and paralinguistic levels, while Audio-LLM judges show selective, prompt-dependent detection that favors word-level errors and does not generalize across all dimensions. Neither class reliably captures a breadth of linguistically structured, human-perceptible speech errors. Our dataset, annotation schema, and evaluation code are publicly released to support more targeted and interpretable TTS evaluation.

Multimodal Large Language Models (MLLMs) exhibit strong visual understanding, yet the internal features that cause these behaviors remain difficult to identify, audit, or control. While applicable to post-hoc inspection, hidden states that are decomposed into interpretable feature directions using sparse autoencoders (SAEs) neither readily isolate which features are changed by multimodal training, nor are they directly useful for targeted control. We introduce MMDiff, a multimodal model-diffing framework that trains multimodal SAEs and turns them into feature-level interfaces for discovering and controlling multimodal behavior. MMDiff supports three uses: (i) feature isolation, by diffing a base-LM SAE against its multimodal-adapted counterpart to identify features altered by multimodal training; (ii) task-specific feature detection, via per-token contrastive firing analysis that isolates causal features; and (iii) feature-level control, by causally removing or steering the discovered feature directions. We train multimodal SAEs for three MLLM families, LLaVA-MORE, PaliGemma 2, and InternVL3.5, and evaluate on visual-spatial understanding, multimodal safety, and OCR. MMDiff discovers sparse, causally specific features whose removal selectively degrades target behaviors by an average of 12% on spatial tasks and 17% on OCR, and reduces attack success rate by 24% on multimodal safety attacks, with no impact on VQA performance. Steering these features improves spatial and OCR accuracy by +3.6% and +1.8% on average over a standard single-layer steering baseline. These results show that multimodal SAEs can serve not only as interpretability tools, but as mechanisms for auditing, steering, and controlling MLLMs behavior toward safer and more capable generations.