When an AI algorithm makes decisions that affect more than one person, aligning it becomes a problem of social choice: how should people's divergent preferences about system behavior be reconciled and aggregated into a single coherent model? The standard approach to aligning frontier AI models - reinforcement learning from human feedback - largely sidesteps this question and has poor social choice guarantees. However, it remains unclear what alternative should replace it. We show that, when agent utilities can be expressed linearly in a fixed feature representation, the alignment problem can be reformulated as linear optimization over a convex impact space, which makes it amenable to the standard toolkit of welfare economics and mechanism design. This reformulation clarifies how alignment protocols translate into welfare consequences and, conversely, how a social planner's desired constraints on welfare consequences can be translated back into alignment protocols. We apply this transformation to show that voting-by-issues and random-dictatorship mechanisms are strategyproof and unanimous. Demonstrating the reverse direction, we also apply the impact representation to derive a family of alignment protocols that maximize utilitarian social welfare subject to various social desiderata, such as bounds on individual or group harm. We illustrate the welfare implications of these alignment protocols empirically using real human preferences over kidney allocation, charitable food distribution, LLM responses, and trolley problems.
Knowledge Tracing (KT) aims to assess students' dynamic knowledge states from their learning histories. While most existing KT methods focus on single-domain learning with notable success, real-world learning scenarios often involve multiple domains simultaneously, introducing two critical factors: 1) Cognitive load, arising from managing learning across domains in both temporal and knowledge dimensions. 2) Knowledge transfer, where knowledge states in one domain influence related states both within and across domains. In this paper, we focus on exploring these factors to improve students' knowledge state assessment in multi-domain learning scenarios and propose a novel method incorporating cognitive Load and knowledge Transfer for Multi-domain Knowledge Tracing (LT-MKT). Specifically, to bridge isolated domains, LT-MKT first integrates textual information from questions and their associated concepts to construct a Multi-domain Hierarchical Graph, leveraging the advanced representational capabilities of large language models (LLMs). Then, cross-domain features in both the temporal and knowledge dimensions are explicitly modeled to capture the effects of cognitive load. Additionally, a knowledge transfer module is designed to model the propagation of knowledge states within and across domains. By jointly modeling these factors, LT-MKT enables more accurate prediction of students' future performance. Finally, extensive experiments on real-world datasets demonstrate that our method achieves state-of-the-art performance.
Safety alignment helps models adhere to human values, but it often reduces response utility. We ask a critical but understudied question: Does safety alignment impose the cost equally across language groups? To answer this, we introduce a rigorous protocol to measure the utility loss imposed solely by safety alignment, which we term Safety Cost. Through direct pairwise comparisons between safety-aligned models and their unaligned counterparts, we find a systematic inequity: non-English users consistently bear a higher Safety Cost than English users. We further identify three underlying patterns. First, multiple languages lie in a double-penalty zone, experiencing both weaker safety protection and larger utility loss. Second, certain languages exhibit apparent utility gains that are in fact a consequence of safety filters failing to engage. Third, even high-resource languages pay a larger Safety Cost than English to reach the same level of safety. We show that these disparities arise from both explicit refusals and implicit qualitative differences across multiple dimensions. By accurately measuring the disparate effects of safety alignment, our findings expose a systematic disparity in current safety alignment practices.
Web-facing AI agents repeatedly access webpages, documents, repositories, and tool responses, yet reading interfaces often expose entire objects when only sparse evidence is needed. This over-reading increases token and latency costs and can dilute relevant evidence. We present SparseRead, a reusable reading control layer that governs content admission without modifying the agent model. Its Read Intent Protocol tracks evolving needs and source coverage; Adaptive Reader Backends select how evidence is located and exposed; and the Budgeted Read Gate bounds cumulative raw exposure per episode with faithful full-view fallback. A unified Reader interface supports interchangeable evidence acquisition across heterogeneous resources. Across six models, including Claude Opus 5, and five scenarios, SparseRead reduces total tokens by up to 92.9% and wall time by up to 89.0%, preserving or improving task quality in 97% settings. On Web deep-search tasks, it reduces tokens by 31.0--43.0% and task completion time by 23.0--37.0%. Ablations validate the three designs, while integration with three agent frameworks demonstrates portability.
The components of a transformer communicate by writing to and reading from a shared residual stream, and the mechanistic interpretability literature has mapped these connections by hand, one circuit at a time. We present the communication map, which charts every potential communication channel from the geometry of the model's weights alone. Its unit of analysis is the whole transformer rather than the individual head or circuit. Every ordered writer-reader pair, across all 18 connection classes from head-to-head to neuron-to-neuron, receives a coupling coefficient that generalizes the composition score of Elhage et al. (2021), with an exact chance level. The census finds that 70-89% of head pairs are oriented far from chance, some coupled strongly and others actively avoiding each other. The map recovers the known induction circuits blind from the strongest head-to-head couplings and groups the heads into communities, and ablating one such community destroys the model's in-context copying. As an application, we pool the coupling coefficients of every head to identify a two-dimensional residual stream subspace, distinct from those found by activation PCA or outlier dimensions, whose deletion abolishes induction in six models up to Pythia-6.9B. We release the map, the statistical machinery, and the intervention suite.
Clinical decision-making is inherently experience-driven: physicians progressively refine their reasoning by synthesizing patient history, multimodal observations, and prior diagnostic experiences across interactions. In contrast, current multimodal large language model (MLLM)-based medical AI agents largely operate as stateless inference systems, generating decisions independently for each interaction without retaining or internalizing experiential knowledge. This discrepancy limits their ability to progressively improve reasoning reliability through usage and adapt to longitudinal patient contexts in real-world clinical workflows. In this study, we propose Medical Structured Multimodal Memory (MSM-Mem), an agentic memory framework that enables medical AI agents to evolve through accumulated clinical experiences. MSM-Mem organizes heterogeneous clinical experiences into semantic, episodic, and visual memory and incrementally updates them during inference, allowing the agent to retrieve prior experiences to inform current reasoning and progressively refine decision-making over time. Evaluations on MoE-LLaVA backbones demonstrate consistent performance improve- ments with further gains observed through continued usage. In general, MSM-Mem offers a viable pathway toward medical AI agents capable of evolving their reasoning competence in a manner analogous to the way clinicians learn from practice over time.
The rapid proliferation of memecoins on blockchain platforms has increased the risk of fraudulent activities, particularly rug pulls. While previous studies have focused on Ethereum-based tokens, this paper shifts the spotlight to Solana, the leading blockchain for memecoins by trading volume and token count. Unlike Ethereum, where rug pulls often exploit smart contract backdoors, Solana memecoin rug pulls are predominantly driven by liquidity manipulation and social dynamics. This research pioneers large-scale rug pull early detection in the Solana ecosystem by assembling a dataset of 6.4 million tokens over 7 months. Market analysis reveals that a vast majority of these memecoins exhibit rug pull characteristics within one hour of launch, highlighting the urgency of short-horizon prediction. Despite the absence of code-level features, we demonstrate that classic machine learning models, particularly Gradient Boosting (XGBoost), achieve robust performance in detecting potential rug pulls using only the first 5 minutes of trading data. Furthermore, we evaluate cross-platform generalization between PumpFun and Raydium, revealing that multi-source data fusion significantly mitigates domain shift and improves detection reliability. This study advances the understanding of DeFi fraud on high-throughput chains and provides a practical framework for protecting investors.
Customer-service LLM agents must follow organizational policy when acting on a user's behalf. Compliance failures arise from either forbidden actions, such as granting an ineligible change, or omitted procedural requirements, such as identification or confirmation. Runtime safeguards can intervene on risky actions, but action-local checks do not guide an agent through a multi-step procedure. Workflow-following systems support prescribed process execution, but primarily target workflow completion rather than safeguarding agent behavior. PolicyGuide instead compiles each domain policy into a workflow graph and invokes a proactive verifier at user-turn boundaries. From persisted graph state, the verifier reconciles open requests and returns step-specific remediation along a policy-compliant path. Across the $\tau^2$-bench airline, retail, and telecom domains with a GPT 5.4 agent and verifier, PolicyGuide raises mean $\mathrm{Pass}^4$ from $0.42$ to $0.62$, with the largest gain on telecom ($0.19$ to $0.61$), the most workflow-structured domain. The same workflows transfer to Claude Sonnet 4.6 and Gemini 2.5 Pro agents. Complementary evaluations find the lowest observed attack-success rate under adversarial users and the strongest procedural compliance in an independently validated workflow-level audit.
Site-specific wireless design relies on ray tracing to predict how signal strength varies across a floorplan, but a converged indoor radio map takes tens of seconds to simulate and planning a single building requires thousands of them. A learned surrogate can remove most of that cost, yet training one runs into a second problem: high-fidelity labels are exactly what is too expensive to produce at scale, so the affordable labels come from finite-ray simulations that carry residual Monte Carlo noise. We address this with RayNO, a physics-unrolled neural operator that reconstructs the converged output of a ray tracer from one cheap low-fidelity run and scene priors. Rather than treating a radio map as a generic image, RayNO refines the field through three stages matched to reflection, diffraction, and scattering, each supervised by a label of matching physical complexity. We prove that when the label error is conditionally unbiased the learned operator can end up closer to the high-fidelity reference than the labels it was trained on, and we verify that condition directly in the simulator using repeated independent ray seeds. Across diverse floorplans, RayNO outperforms image-to-image baselines, wireless learning models, monolithic neural operators, and classical multi-fidelity regression on both image-quality and wireless deployment metrics, while reducing the cost of a high-fidelity map from roughly 28 s to 0.11 s.
Online video platforms can expose young users to harmful content, but independent audits remain difficult because video annotation is costly and moderation judgments vary across languages. We audit TikTok in France, Italy, and Sweden with sockpuppet accounts representing four age personas (13, 16, 19, 40), collecting 36,971 videos from passive For-You-page scrolling and active sessions that scroll, search for harm keywords, and scroll again. To scale annotation, we validate four multimodal LLMs against native-speaker labels on a 300-video reference set. Gemini 2.5 Flash with eight sampled frames plus text performs best (aggregate kappa = 0.42), at half the per-call cost of native-video upload, and we apply it to a 10% sample for approximately \$50 in total API spend across both modalities. Keyword search returns 35-56% harmful content, a 1.5-7.5x increase over the scrolling baseline in ten of twelve country-age combinations; the spike is temporary and flattens the age differences observed in France and Sweden. Under passive scrolling, Italy has the highest harm rate at every age, with Italian age-19 reaching 48.6%. Overall, MLLM-based auditing offers a scalable approach for cross-national youth-safety audits, while provider safety filters (1.1% refusal rate) under-count the most explicit harms.
Reinforcement Learning (RL) has emerged as a powerful approach for improving reasoning in language and vision-language models, yet its strongest successes still depend heavily on ground-truth supervision (e.g., verifiable reward). Such annotations are costly to obtain and become increasingly scarce as reasoning capabilities advance beyond what humans can reliably evaluate. Self-rewarding RL reduces this dependence by enabling models to derive reward signals from their own completions. However, training solely on self-generated feedback can reinforce existing biases, reduce response diversity, and ultimately lead to homogenized responses and training collapse. In this work, we show that unsupervised reasoning can emerge through cooperative multi-agent training. We introduce Co-RL, a framework in which multiple decoupled models, sharing no parameters, are simultaneously optimized through RL using rewards derived from the other agent. We further show that increasing cohort diversity, through heterogeneous model families, sizes, and rephrased training samples, reduces the correlated errors that drive self-reinforcing feedback loops. This diversity consistently improves reasoning performance and mitigates training collapse. Without any ground-truth labels, Co-RL improves the base models by 3.0-7.9% on average across seven benchmarks and by 1.6-4.4% across four multimodal benchmarks. It also outperforms the prior state-of-the-art multi-agent RL method CoMAS by 4.0% and the self-rewarding vision-language RL method MM-UPT by 2.0%. Our code is available at https://github.com/Diverse-CoRL/Co-RL and website at https://williamium3000.github.io/Co-RL.
What fine-tuning teaches a model is usually diagnosed after the fact. We make it an experimental variable. OraclePhys has three parts. OraclePhys-Bench is a structural-mechanics benchmark whose finite-element oracle grades every answer and counterfactual edit. OraclePhys-30K holds seven answer forms over byte-identical structure descriptions. A controlled training study compares the forms and verifier roles. Two findings follow. First, within this task family the answer form of the label determines what fine-tuning teaches. Ranking supervision lifts held-out localization of the most-deforming story on taller structures from 0.25 to 0.86, a peak-value label yields at best a partial predictor, and a pass/fail label none. This ranking-versus-peak gap replicates on a second domain, a second model family and a paraphrased surface. Second, the gap between our standard SFT and GRPO runs is mostly a recipe effect. With learning rate and step budget matched, the two reach held-out accuracies within 0.03 of each other. At higher rates, GRPO falls below its start while SFT does not. Both also teach the model to apply edits given as instructions. The trained 8B matches a specialist and exceeds a frontier LLM. Both domains are linear, steady-state systems with synthetic text inputs.
With the growing deployment of machine learning models, formal guarantees of the robustness and fairness of these models have become increasingly important in safety-critical and legal-compliance settings. However, model parameters are often commercial secrets that cannot be disclosed to auditors or end users. To this end, we present PANDA, a scalable system that uses zero-knowledge proofs (ZKPs) to prove the robustness and fairness properties of a model without revealing its private parameters. PANDA is built on top of CROWN, an efficient robustness certification framework that is used in many state-of-the-art formal verification tools for neural networks. The core contribution of PANDA is a novel algorithm for proving linear relaxation bounds for non-linear activation layers, yielding simple, lightweight proofs. Remarkably, our implementation can generate proofs of local robustness for neural networks with more than 2.9M parameters in 5 minutes, and can verify them in 10 seconds. Prior ZKP-based robustness system rely on exponential-time algorithms that cannot scale to nontrivial networks. In contrast, PANDA scales polynomially in the number of neurons in a network, allowing us to support neural networks 4 orders of magnitude larger than previous approaches with significantly reduced prover overhead.
Direct communication between AI systems relies on natural language as an intermediate layer, incurring encoding/decoding overhead, token cost, and latency. We ask whether internal activation states can instead be transferred causally between different large language model (LLM) architectures via a learned projection, evaluated at three levels: representational similarity, cross-model retrieval from projected states, and end-to-end causal transfer via activation injection during generation. Using four architecturally diverse open-weight models (Qwen2-0.5B, Phi-3-mini, Mistral-7B, FLAN-T5-base), we find that representational alignment in trained models exceeds a random-initialization null baseline and is best captured by a rank-based metric (mutual k-nearest-neighbour alignment), more robust to activation-magnitude outliers than centered kernel alignment (CKA) or Procrustes analysis. A learned projection network retrieves the correct target-model representation from a held-out set well above chance for the three causal decoder-only model pairs (45-50% top-1 accuracy vs. 5% chance) but at chance level for the encoder-based FLAN-T5. Injecting projected activations into a target model during generation produces a statistically significant, pre-registered causal effect on retrieval-based output similarity for only one of the three decoder-only pairs (Qwen2-0.5B to Phi-3-mini: 23.3% vs. 0.0% under negative control, p=0.047, FDR-corrected); the two pairs targeting Mistral-7B show no such effect despite comparable representational alignment at the hidden-state level. We interpret these results as evidence for causal transfer of the representational vehicle, not of meaning, and conclude that end-to-end activation-state transfer between LLMs, as currently implemented, is architecture-dependent rather than universal.
LLM-based coding agents have advanced rapidly on single-process SWE tasks, with frontier models now clustering in the high-70s on SWE-bench Verified. Distributed-system debugging, however, remains an under-explored regime: bugs span processes, nodes, and protocol interactions, with root causes rarely recoverable from source alone and brute-force exploration intractable across non-deterministic interleavings. This leaves two gaps in LLM and agent evaluation: no code-repair benchmark targets distributed-system bugs, and no controlled study isolates how much externally provided debugging context changes agent success on them. We introduce DDBench, a code-repair benchmark of 60 historical bugs mined from 13 open-source distributed systems, partitioned into three reasoning-scope tiers. DDBench evaluates every case under two matched conditions: a symptom-only condition where the agent receives only the bug symptom and repository, and a context-augmented condition where it additionally receives bounded debugging context (logs, traces, runtime state, and targeted code-investigation notes), isolating the effect of debugging context from model capability. The evaluation of 10 LLMs on DDBench reveals several findings. First, distributed debugging exercises a reasoning dimension that single-process benchmarks do not surface: models' pass rates span 45 pp, and pairwise bootstrap separates 6 of 15 top-tier model pairs at p < 0.05 on DDBench's hardest case-set. Second, bounded debugging context lifts aggregate pass rate by +19.0 pp, and the lift is asymmetric: weaker models gain more in pass rate, while stronger models gain more in efficiency. Third, debugging context requires careful curation, as even faithful debugging context can anchor a strong model on the wrong layer of the system when its observation sits far from the fault.
Residual connections are a fundamental component of transformer architectures, yet the roles of the attention and feed-forward residual pathways remain poorly understood when considered independently. This paper presents a reproducibility study of partial residual ablations in Pre-LN GPT-style transformers trained at two scales (10M and 124M parameters). I compare four architectural configurations by selectively removing the attention residual connection, the feed-forward residual connection, or both. At 10M scale, a controlled 8-seed deterministic sweep shows a clear asymmetry: removing the attention residual (FFNOnly) reaches the No-Residual collapse regime (3.350 +/- 0.002), whereas removing the feed-forward residual (AttnOnly) remains well separated from it (1.580 +/- 0.003). A subsequent controlled 124M sweep establishes that the direction of this asymmetry persists under matched seeds: across five seeds each, AttnOnly reaches 6.037 +/- 0.355 versus 7.569 +/- 0.001 for FFNOnly, with no overlap between the observed ranges. AttnOnly nevertheless remains substantially degraded relative to FullResidual (4.577 +/- 0.012) and exhibits markedly greater seed sensitivity at 124M. During the investigation, I identified and corrected an experimental measurement confound in runtime gain scaling and retained an intermediate reproduction failure rather than excluding it. I propose cross-position routing through self-attention as a falsifiable hypothesis for the observed asymmetry; no clean causal test of this mechanism has yet been completed. To support reproducibility, I release the source code, experiment configurations, training logs, and experimental results, including intermediate non-reproducing runs.
Modern stochastic predictors can model rich, multi-modal outcome distributions. However, this expressive power comes with challenges in ensuring robust predictions $-$ a critical requirement in safety-critical domains. Randomized smoothing is a leading technique for improving robustness, particularly against adversarial perturbations. Yet, in stochastic multi-modal regression settings, randomized smoothing often fails due to mode collapse, yielding averaged predictions that do not reflect the underlying distribution. To address this limitation, we propose clustered $\alpha$-smoothing, a framework that (1) partitions noisy samples using an arbitrary clustering algorithm, (2) applies $\alpha$-smoothing locally within each cluster, and (3) combines the resulting predictions into a mixture distribution. By interpreting the smoothing distribution as a mixture of $\alpha$-smoothers, we derive a lower bound on the probability that the smoothed prediction lies within a union of compact regions corresponding to distinct modes. We empirically evaluate our framework on two benchmarks, demonstrating substantial improvements over state-of-the-art methods. In stochastic trajectory prediction on a driving simulator dataset, our approach achieves, on average, a $27\%$ lower Wasserstein distance to the ground-truth distribution compared to $\alpha$-smoothing. In quadrotor control, where modes correspond to distinct feasible paths to a target, our method reduces the collision rate by $90\%$ relative to the state-of-the-art randomized smoothing.
Air traffic controllers perceive traffic complexity through the radar display, suggesting that a computer vision model operating on the same imagery may provide a natural architecture for modeling controller-perceived complexity; however, whether radar imagery is a viable input format for deep learning vision models remains unclear. Unlike natural images, radar images are extremely sparse and self-similar, consisting primarily of a black background and a few visually identical aircraft blobs, while small changes in aircraft positions can substantially alter sector-level complexity. To test whether a vision model can capture these operationally important differences, we encode each traffic situation as a position image supplemented by five channels representing aircraft state variables, including heading, speed, and altitude, and train a Vision Transformer (ViT) to regress four intrinsic complexity components derived from pairwise geometric relations among aircraft. The model achieves $R^2 > 0.96$ for all four components on test data, and a one-aircraft removal perturbation study shows that the change in the model's estimate is proportional to how much the removed aircraft contributed to sector complexity, rather than the same for every removal. These results demonstrate that, despite its atypical visual characteristics, the proposed image-based representation is a viable input format for air traffic complexity modeling.
Low-resource automatic speech recognition (ASR) often relies on cross-lingual transfer from higher-resource languages due to limited availability of transcribed speech. However, selecting an effective donor language is a nontrivial issue, particularly in settings with spontaneous speech, linguistic variation, and uneven resource availability. We present DonorRank, a learning-to-rank framework that uses linguistic and dataset-level features along with previously observed transfer outcomes to predict promising donor languages for zero-shot ASR. We evaluate DonorRank on two multilingual speech corpora of Indic and African languages. DonorRank produces informative donor rankings across both collections, outperforming common heuristics in the more closely related Indic setting while remaining competitive in the more linguistically diverse African setting. Our analysis reveals that effective donors are not always the closest linguistic relatives and that the most informative predictors vary across language collections. These findings highlight the value of learned donor selection for prioritizing transfer experiments in low-resource ASR. We will also release all of our evaluation results, feature matrices for all languages, scripts for training the DonorRank framework, and pairwise transfer performance matrices for our datasets.
Conway's 99-graph problem asks whether a strongly regular graph with parameters srg(99,14,1,2) exists. We develop two complementary lines of attack. Fixing one vertex, the conditions lambda=1 and mu=2 force its neighbourhood to be a perfect matching and determine every edge between that neighbourhood and the remaining vertices. For (99,14,1,2), the unresolved part is therefore a constrained 12-regular graph on 84 vertices. We encode this reduction in CP-SAT and validate it by recovering the unique srg(9,4,1,2). We also prove by exhaustive enumeration that no circulant graph on Z/99 satisfies more than 68.0% of the CAISc constraints, and we give a validated orbit formulation for prescribed automorphisms. We then study the partial-score search problem. Fourteen human-designed search configurations reached at most 69.43%. Separately, we supplied the scoring function to an evolutionary program-search system. It produced a degree-preserving 4-vertex-switch tabu search whose best verified artifact scores 70.73%. The generated move differs from those used in our own searches and crosses a plateau that was stable under them. A tabu search restricted to graphs with a prescribed symmetry group, first Z/33 and then its subgroup of order 11, reaches a 14-regular graph scoring 76.67%. These results do not resolve the existence problem, but they reduce the exact search space and improve the best verified partial construction found in our experiments.