Papers

Large language models offer a promising interface for translating natural-language protocol descriptions into formal security models, but their outputs remain difficult to trust without expert validation. In this paper, we present a human-in-the-loop framework for generating Tamarin-verifiable formal models of security protocols. Our key observation is that the main correctness bottleneck is the semantic accuracy rather than the syntactic validity of the intermediate protocol representation. To address this problem, we introduce a protocol intermediate representation (IR) that serves as a human-auditable semantic checkpoint between natural-language parsing and formal model generation. The IR explicitly captures protocol participants, message flows, value provenance, cryptographic operations, proof targets, and compromise assumptions. We further design an interactive interface that highlights uncertain fields and guides users to inspect the most critical semantic decisions based on model confidence before model generation. Rather than replacing formal-methods experts, our approach uses LLMs to produce auditable semantic drafts while leveraging verification tools to check the resulting formal models. Code and verification artifacts are available at https://github.com/laplace1002/TamarinAgent.git.

In a companion paper, we found that near the untwisted point of a four-boundary AdS$_3$ wormhole, whenever the complete bipartite Ryu-Takayanagi entropy vector is insensitive to a Fenchel-Nielsen twist, the holographic $\mathtt q=4$ multi-entropy is insensitive as well. Here we show that the situation changes at finite twist. The twist lengthens the crossing geodesics and thereby enlarges the region in which all bipartite RT entropies remain invariant. In part of this enlarged region, a twist-sensitive $\mathtt q=4$ network becomes globally minimal. We thus find an open region of wormhole moduli space in which the complete bipartite RT entropy vector is exactly unchanged while the $\mathtt q=4$ multi-entropy detects the twist. Our result provides an explicit example in which the $\mathtt q=4$ genuine multi-entropy probes bulk geometric information invisible to all bipartite entanglement entropies.

We ask whether multipartite entanglement can probe bulk moduli invisible to bipartite entanglement. We study this question in an explicit family of four-boundary AdS$_3$ wormholes related by a Fenchel-Nielsen twist $\tau$, asking whether the complete set of bipartite RT entropies can remain exactly $\tau$-independent while the holographic $\mathtt q=4$ multi-entropy is sensitive to the twist. Within the class of configurations analyzed here, for sufficiently small $|\tau|$, we find no such regime along the symmetric line $\ell=m$: the globally minimal $\mathtt q=4$ network is itself exactly $\tau$-independent throughout the parameter range analyzed. This conclusion is not a priori obvious: it emerges from a competition among several admissible configurations whose relative ordering changes multiple times as the geometry is varied, and persists up to the point where the bipartite RT entropies themselves become twist-sensitive.

Interpretation of compact binary signals observed by third-generation gravitational-wave detectors is complicated by their high signal-to-noise ratios, often reaching hundreds to thousands, and by time- and frequency-dependent detector responses arising from Earth's rotation and the detectors' large physical extent. Building on prior work for LISA, we implement generic, robust extensions of the RIFT parameter estimation pipeline that enable interpretation of these sources using any existing waveform model, without recourse to intermediate pretrained data products like reduced-order approximations or normalizing flows. We demonstrate end-to-end recovery of finite-size binary-neutron-star injections in a three-site next-generation network to $\mathrm{SNR}=1000$, localizing the source's extrinsic parameters to $0.8\times10^{-3}\,\mathrm{deg}^2$ in minutes on a single GPU.

Modeling the temporal evolution of macroscopic properties of complex systems is an important scientific task. To predict this evolution without full microscopic simulation, a common approach encodes microstates into compact latent states, learns their evolution, and reads out macroscopic predictions from the latent trajectory. These latent states are often learned through microstate reconstruction. However, with limited latent capacity, reconstruction can favor high-variance microscopic details over information needed for macroscopic prediction. Yet jointly learning latent states and their transition without reconstruction often fails to obtain latent dynamics that support accurate macroscopic prediction. We show that this failure can arise from latent scale collapse: shrinking the latent state scale reduces training loss while macroscopic evolution error remains large. Here, we propose a reconstruction-free framework to learn latent states with their dynamics for prescribed macroscopic prediction. Training alternates between updating the latent representation with the transition and next-state latent targets fixed, and updating the transition with the latent representation fixed. At inference, the trained model predicts macroscopic states recursively from an initial microstate. Our theoretical analysis characterizes reconstruction misalignment and scale collapse under joint training, and gives a sufficient condition for local convergence to correct latent dynamics for our method. Experiments on epidemic spreading on a lattice, mixing of two particle species, and polymer stretching demonstrate that the proposed method achieves substantially better macroscopic prediction over baselines.

Diffusion language models (DLMs) enable parallel generation by predicting and committing multiple tokens at each denoising step, yet they can generate individually plausible but mutually inconsistent tokens. Recent work shows that soft tokens can mitigate this issue by representing uncertain positions with continuous embeddings built from the model's predictive distribution at the previous decoding step. However, although soft tokens are commonly understood as preserving predictive uncertainty, how soft-token feedback improves parallel decoding has not been systematically examined. In this paper, we investigate this question in frozen pretrained DLMs to examine soft-token feedback without the effects of additional training. To construct soft-token inputs in a training-free setting, we identify a geometric mismatch between conventional soft-token construction and the pretrained embedding space. Based on this observation, we propose a training-free, geometry-aware construction of soft tokens. Our analysis of soft-token feedback suggests that uncertainty preservation alone does not fully explain how it reshapes subsequent predictions. To better explain how soft-token feedback improves parallel decoding, we provide empirical evidence that it favors coherent token sequences. Across four pretrained DLMs and four math and code benchmarks, our method outperforms standard parallel decoding and a training-free Euclidean soft-token baseline. Code: https://github.com/kodaikawamura/rethinking-soft-tokens

We present a determinantal evaluation formula for the very-well-poised ${}_8\Phi_7$ basic hypergeometric Askey-Wilson function in terms of elementary functions. The formula in question is valid for discrete values of the four permutation-symmetric Askey-Wilson parameters; specifically, these consist of all quadruples of parameter values that are given, up to a sign, by powers $q^k$ with $k$ a positive integer or half-integer, in such a way that each of the four types (positive/negative sign, integer/half-integer power) occurs exactly once. The proof hinges on a product formula for an associated Cauchy-type determinant, which is of independent interest and is established here by elementary means using Krattenthaler's `identification of factors' method for determinant evaluations.

For an integer $b\neq0,1$ let $\theta$ be the unique real root of $f_b(x)=x^3-3bx-b^3$ and let $K_b=\mathbb{Q}(\theta)$. We exhibit an explicit set of $b$ of positive density for which $\eta_b=-1/(\theta-(b+1))$ is the fundamental unit of $K_b$, and we determine the set of $b$ for which $\eta_b$ is the square of a unit: it is parametrized by the Pell equation $D^2-3E^2=1$, hence infinite, and under an explicit mild condition on the field discriminant it accounts for all $b$ for which $\eta_b$ is not the fundamental unit. That condition fails for only four $b$ with $|b|\le3000$, and at one of them, $b=3$, the conclusion itself fails. In the order $\mathbb{Z}[\eta_b]$ generated by $\eta_b$, by contrast, $\eta_b$ is the fundamental unit for every $b$, so these exceptions are a phenomenon of the maximal order. As an application we construct infinitely many biquadratic fields whose $3$-class field tower has length greater than $1$.

We study the Assouad dimension, lower dimension, and Assouad spectrum of graphs of the functions \(f(x)=\sum_{n=1}^{\infty}a_n\phi(b_nx+\theta_n)\), where \(\phi\) is a nonconstant \(1\)-periodic \(C^2\) function, \(a_n,b_n>0\), \((\theta_n)\) is an arbitrary real sequence, \(\sum_{n=1}^{\infty}a_n<\infty\), and \(b_{n+1}/b_n\to\infty\). We give sufficient conditions for the graph over every nondegenerate compact interval to have Assouad dimension two or lower dimension one. For \(a_n=b_n^{-\alpha}\), \(0<\alpha<1\), an explicit frequency-gap condition ensures that the Assouad spectrum equals two for \(\alpha\leq\vartheta<1\). Writing \(\beta_*=\limsup_{n\to\infty}\log b_{n+1}/\log b_n\), we show that, if \(\beta_*>(2-\alpha)/\alpha\), the spectrum attains the general upper bound for \(\alpha\)-Hölder graphs whenever \((2-\alpha)/\beta_*\leq\vartheta<\alpha\) and \(\vartheta>0\). When \(\beta_*=\infty\), this determines the complete spectrum. Together with Barański's formulas, our results yield graphs whose lower, Hausdorff, upper box, and Assouad dimensions are four distinct numbers.

Enlarged perivascular spaces (PVS) visible in brain magnetic resonance imaging (MRI) are increasingly thought to be linked to poor brain health. PVS are elongated structures of less than 3 mm in diameter and can be numerous. To reflect the incidence of PVS, radiologists visually score their burden following a clinical grading scale - a task that would benefit from automation to accelerate analyses and overcome the influence of inter-observer differences. We developed and evaluated methods for training machine learning models to score PVS incidence in the basal ganglia (BG) and centrum semiovale (CSO) leveraging the Potters/Wardlaw scale. The novelty in our work lies in the use of imperfect, semi-automatically generated "silver-standard" PVS segmentation masks during training, in addition to PVS radiological scores. We comparatively evaluated a conditional convolutional neural network (CNN) which accepts PVS masks as an extra input channel, a multi-task CNN which performs both PVS segmentation and scoring, and a logistic regression model which utilises features derived from PVS masks to predict PVS scores. Multi-task learning was the most effective method, achieving a mean average precision of 64.08% compared to 60.22% for the conditional CNN, 52.11% for a baseline CNN trained only to predict PVS scores, and 49.32% for the logistic regression model. The multi-task model showed an ability to localise individual PVS not shown by the other CNNs, and behaved in a probabilistically sensible way, predicting with lower confidence on inherently harder classes. Age, sex, hypertension status, white matter hyperintensity volume, and ischaemic stroke lesion status were shown to be associated with the multi-task model's PVS score predictions and the ground truth in a similar way.

Pre-operative planning for bronchoscopy is critical for the diagnosis of lung lesions. Current accessibility assessment relies on subjective manual inspection of CT scans, which is time-consuming and prone to inter-observer variability. In this paper, we formalize bronchoscopy accessibility prediction as a novel supervised learning task and present the first end-to-end framework to address it. We propose an Anatomy-Aware Mixture-of-Experts (MoE) model that integrates specialized modules: a CT Expert for local morphological features, a Lobe Expert for anatomical priors, and a Path Geometry Expert that encodes the sequential constraints of the bronchial tree. To support this task, we curated the first clinical dataset of 438 cases with pre-operative CT scans and documented procedural outcomes. Experimental results demonstrate that our method achieves an AUROC of 0.8052, significantly outperforming both state-of-the-art baselines and experienced human experts. This work establishes a new benchmark for computer-aided interventional planning in pulmonary medicine. Our data and code will be publicly available at https://nubagcilab.github.io/BronchoAccess/.

We investigate fragmentation and molecular growth induced by keV ions in initially neutral gas-phase adamantane clusters, the simplest diamondoid molecule (Ad, C$_{10}$H$_{16}$). Using time-of-flight mass spectrometry in combination with quantum-chemical calculations, we find that intact cluster cations undergo extensive dehydrogenation. We also observe a distinct distribution of singly charged products with masses predominantly between those of the adamantane monomer and dimer cations. We attribute these species to ion-induced growth products and interpret their formation considering the smallest cluster cation, [Ad]$_2^+$, using quantum-chemical calculations. Our results demonstrate that ionization strongly reorganizes the dimer geometry, facilitating H$_2$ elimination and intermolecular covalent C–C bond formation. Weakly bound adamantane clusters are thereby transformed into covalently bound carbonaceous species with a strong preference for specific molecular structures. These findings provide new insight into ionization-driven molecular growth in saturated hydrocarbon systems contributing to the understanding of molecular complexification, with implications for chemical physics and astrochemistry.

Molecular representation learning is central to computer-aided drug discovery. Molecular graphs, SMILES strings, and 3D conformations provide complementary structural information, yet many multimodal approaches encode these views independently and align them only at a later stage, limiting fine-grained cross-modal interaction and substructure-level interpretability. To address these limitations, we introduce MoTIF-X, a motif-centered framework that uses graph-grounded chemical motifs as shared anchors for multimodal integration and interpretation. Its first pretraining stage learns motif representations through hierarchical contrastive learning across atomic, motif, and molecular scales. The second stage contextualizes these representations with SMILES and torsion-angle tokens through multimodal masked token modeling. After pretraining on drug-like molecules with multiple conformers, MoTIF-X achieved the lowest mean absolute error on all nine OpenADMET ExpansionRx endpoints and the best overall performance among the evaluated methods. Significance analyses supported its advantage in the vast majority of endpoint-baseline comparisons after multiple-testing correction. Ablation studies supported the complementary contributions of motif-token contextualization, multimodal integration, and two-stage pretraining. Beyond molecular properties, the framework extended to drug-target interaction prediction, achieving the best average classification performance across the evaluated benchmarks and generalizing to an external drug-cold-start dataset without additional fine-tuning. Its motif-centered design also enabled substructure-level interpretation: higher motif attribution scores were associated with larger experimentally measured activity shifts. Together, these findings support MoTIF-X as a transferable and interpretable framework for molecular modeling.

We prove combinatorial rigidity for the full renormalization attractor of neutral quadratic polynomials. More precisely, any two bi-infinite renormalization towers with the same combinatorics are conformally conjugate on neighborhoods of their Mother Hedgehogs. The rigidity theorem includes arbitrary irrational combinatorics and respective parabolic enrichments. The proof relies on a comprehensive analysis of neutral cascades, i.e. transcendental dynamical systems arising from the rescaled limits of the first return maps of neutral quadratic polynomials. We establish uniform butterfly bounds for neutral cascades and prove that any two combinatorially equivalent neutral cascades are affinely conjugate. We also prove rigidity for parabolic towers with equivalent backward combinatorics.

In factorial randomized experiments with $K$ binary treatment components, we study which baseline covariates modify component main effects and interactions. We define effect-modifier cells $(j,S)$, $j\in[p]$ and $\emptyset\neq S\subseteq[K]$, through the coefficients $\theta^\ast_{jS}$ of population linear projections of conditional Walsh–Hadamard contrasts $\tau_{S}$ onto standardized covariates. Using baseline-robust direct scores, we develop frequentist inference for all $p(2^K-1)$ cells. First, we prove a debiased central limit theorem with global insensitivity: the residualization baseline enters the cross-fitted score through an exactly conditionally mean-zero perturbation. The influence-function remainder therefore requires only foldwise weighted $L_2$ consistency of the baseline estimator, with no prescribed polynomial rate or nuisance product-rate condition. A separate, explicit penalty-scale condition controls the baseline's effect on the score Lasso. Second, under explicitly stated uniform nuisance, baseline, studentization, and influence-array conditions derived from bounded sparse-regression primitives, we establish a high-dimensional Gaussian approximation over the full grid, calibrated by a Walsh-spectral multiplier bootstrap. This yields simultaneous confidence bands and Romano–Wolf step-down selection of modified cells with strong familywise error control, without effect-heredity assumptions. The factorial design is fixed in the asymptotic analysis, while the covariate dimension may grow; all $2^K-1$ contrasts are covered jointly.

Neural simulation-based inference (SBI) has been widely successful in inferring a relatively small number of interpretable parameters from potentially high-dimensional observations, such as images or time series. Accordingly, representation learning in SBI has focused almost exclusively on compressing the observations used to condition the posterior. More recently, however, SBI has begun to target increasingly high-dimensional parameter spaces, raising the complementary question of whether the inference target itself should be compressed. Our answer is a practical merger of SBI and latent generative modeling, which learns a low-dimensional representation of the simulator parameters, performs posterior inference directly in this latent space, and maps posterior samples back to the original parameter space. We characterize the conditions under which latent-space inference recovers the desired target posterior and systematically study its empirical trade-offs. Across four case studies and three generative families, we compare latent and standard estimators while controlling for network capacity, regularization, optimization, and training compute. At matched training compute, latent-space inference achieves accuracy and marginal calibration comparable to direct target-space inference while sampling up to more than an order of magnitude faster.

Learning how to replenish inventory and estimating the cost of doing so require different information. We establish information-theoretic lower bounds for offline multistage inventory control with independent bounded demand. For fixed additive error in total expected cost, policy learning requires cubically many scalar observations in the horizon in the worst case, even under full observation and adaptive sample allocation. This matches the known empirical-risk upper order. Stationarity permits a quadratic policy-learning guarantee, yet valuation with inherited stock can remain cubically difficult even when the optimal policy is known. Censoring sharpens this distinction: optimal decisions can be learnable while absolute costs remain unidentified. We give a carrysafe coverage condition under which truncation preserves inventory transitions and policy gaps exactly. Combining this reduction with a matching lower bound yields the sharp censored-learning rate, with raw sample requirements inversely proportional to the usable fraction of sales logs. The same qualified observations can check coverage and fit the policy under a joint error guarantee. The results isolate information limits for decisions and valuation under an explicit conditionally independent logging protocol.

Looped Transformers provide a parameter-efficient approach to depth scaling by repeatedly applying shared Transformer blocks. Recent reasoning models have likewise highlighted the value of scaling test-time computation through longer computation trajectories. However, the principles for designing effective loop transitions remain poorly understood. We view the looped hidden state as a fast weight that is updated throughout the depth. We formulate loop transitions as local gradient-based updates, with recurrent blocks predicting implicit targets at each depth. Our framework derives loop transitions in closed form from a projection, a local objective and an optimizer update rule. Mapping representative loop transitions into this framework reveals mismatches between their transitions and projections. We first align the input maps of existing transitions. We then derive OperLoop, which combines explicit weight decay, adaptive step size and a delta objective. The aligned variants reduce training loss and improve average commonsense accuracy. OperLoop improves average generative performance over the compared looped and non-looped baselines under matched training FLOPs. These results support the framework's usefulness for loop design. We extend the analysis to additional loop models and outline a roadmap for future loop transition design.

Latent world models are trained to predict what happens next, so nothing in their objective separates what an action caused from what merely co-occurred with it. Object-masking models such as C-JEPA intervene on what the predictor can see; we intervene on what physically happens. From one saved simulator state we run the dynamics under an action $a$ and under a reference action $a_{\varnothing}$, and train the model to predict the difference $\Delta z=z^{a}-z^{a_{\varnothing}}$ between the two latent futures. The resulting objective, Do-JEPA, has an effect loss, a support loss (where the action enters), a propagation loss (where its effect travels) and invariance losses (what must not change). In a synthetic system with object-aligned variables, support supervision finds the directly intervened object in 99.95% of test cases, where a sparse action mask sends the action to a nuisance slot in every case, and response-onset supervision recovers the ring-shaped propagation graph (edge AUROC 0.975 vs. 0.624). From pixels, the effect loss beats a control trained on exactly the same data: it lowers latent effect error by 28.4% on an end-to-end LeWM model and physical effect error by 13.5% when trained and tested on natural action sequences, and on three independently generated CausalWorld benchmarks it lowers responsive effect error by about 20% under physics shifts and the latent context sensitivity of predicted effects by 66%. Trained from scratch it costs factual accuracy; fine-tuning an existing model with it removes this cost. Together, these results show that intervening on the world, rather than on what the model sees, helps latent world models predict what their actions cause.

On-policy distillation (OPD) trains students using teacher feedback on their own sampled responses, yet how prompt choice shapes transfer across teacher-student pairs remains poorly understood. We systematically study prompt quantity, source, and selection across RL- and SFT-continuation pairs and cross-model settings. We find that OPD can be highly prompt-efficient: a few prompts can approach large-pool performance, with four DAPO prompts matching the observed mathematics score of 3,840 DeepMath prompts. However, prompt utility is relational rather than intrinsic: changing only the teacher can reverse the relative effectiveness of mathematics and code prompts. To characterize these transfer differences, we analyze parameter and functional changes across prompt supports and model pairs. Functional alignment with the teacher varies across supports and target tasks; in continuation pairs, teacher-aligned prediction changes can coexist with weak parameter alignment. Continued OPD on effective supports can restore performance after unfavorable transfer. Finally, targeted selection does not consistently outperform uniform random sampling, and filtering out a source that performs poorly alone yields no consistent gain across three paired support draws. Overall, our results distinguish prompt efficiency from prompt interchangeability and show that effective data choice depends on the teacher-student pair and target capability, with random sampling providing a competitive baseline in the studied settings.

For graphs $H$ and $F$, let $\text{ex}(n,H,F)$ denote the maximum number of copies of $H$ in an $n$-vertex $F$-free graph. Gerbner, Győri, Methuku, and Vizer proved that $\text{ex}(n,C_{2\ell},C_{2k})=\Theta(n^\ell)$ for $k>\ell\ge2$. They determined the leading term for $\ell=2$, but for $k>\ell\ge3$ their general lower and upper bounds had different leading constants, leaving open the problem of closing this gap. We solve this problem by showing that, for every $k>\ell\ge2$, \[\text{ex}(n,C_{2\ell},C_{2k})=\left(\frac{(k-1)_\ell}{2\ell}+o(1)\right)n^\ell, \] where $(k-1)_\ell=(k-1)(k-2)\cdots(k-\ell)$.

The consistent histories formulation of quantum mechanics assigns classical probabilities to sequences of coarse-grained events when the corresponding interference terms vanish. In finite models these terms generically need not vanish exactly, and their smallness alone supplies no asymptotic geometry on the space of histories. We construct such a geometry for histories that accumulate quantum records. When the normalized Dowker - Halliwell interference between two histories factorizes as the cumulative fidelity of their records, its decay is controlled by the accumulated squared Bures distance. In the infinite registration limit, the Kakutani - von Neumann dichotomy gives a sharp sector alternative: summable Bures defect yields weakly equivalent record sequences, whereas nonsummable defect produces orthogonal superselection sectors and vanishing normalized interference. The convergence exponent of the Bures defect defines a pseudo-ultrametric on histories. After quotienting its zero distance degeneracy, it becomes a complete ultrametric, requiring neither compactness of the outcome space nor continuity of the record family, and coincides with the gauge invariant sector metric $\tilde d$ introduced in \cite{LesTP}. Its value, the consistency rate, has three equivalent interpretations: it is the sector separation metric, the polynomial growth exponent of the negative logarithm of normalized interference, and, by the Helstrom formula, the corresponding exponent of the negative logarithm of the minimum error probability for discriminating the cumulative records. Classical separation profiles determine the rate: positive long time mean defect gives the maximal value $\delta=1$, while the borderline profile $d_m^2\asymp m^{-1}$ produces distinct superselection sectors at zero polynomial rate, $\delta=0$.

Reinforcement learning (RL) has recently delivered substantial gains in multimodal reasoning, opening a promising route for fine-grained visual perception. Yet for multi-image reasoning grounding (MRG), reasoning over real-world multi-image contexts toward pixel-precise localization, existing RL-based approaches overlook two characteristics intrinsic to this paradigm: a coarse-to-fine hierarchical reasoning pattern, and heterogeneously distributed task–sample difficulties. In this work, we present MG-Thinker, a post-training RL framework that advances a new MRG paradigm featuring such hierarchical reasoning, supported by a curated 25K MRG dataset with task-adaptive Chain-of-Thought (CoT) annotations that elicit multi-perspective evidence before conclusion. To remedy the heterogeneous task–sample difficulties, we further propose Bi-Axial DAPO (BiA-DAPO), which decomposes rollout advantages along an intra-group signal axis and an inter-group competence axis through two complementary mechanisms, both grounded on our defined candidate pool for stable group-level statistics. Extensive experiments show that MG-Thinker achieves state-of-the-art performance on multi-image reasoning grounding while consistently improving generalization across multi-image understanding and diverse multimodal benchmarks.

We create a set of 6 axioms relating to strategic voting and study them on single transferable vote (STV), Meek STV, Bloc, Chamberlin-Courant, Monroe, and k-Borda. For Chamberlin-Courant, Monroe, and k-Borda, we implement “optimistic” and “pessimistic” models for handing incomplete ballots, thus turning each of the three into two separate voting rules. We determine which voting rules satisfy each axiom. We find that under Chamberlin-Courant and k-Borda, the optimistic models encourage burying, while the pessimistic models encourage truncation strategies, indicating that the handling of partial ballots has tremendous implications for strategic voting. We define an axiom called no skipped payment that differentiates similar rules, as Meek STV and Chamberlin-Courant satisfy it, while STV and Monroe fail. Each voting rule was found to fail favorite betrayal, revealing that none of them are completely immune to strategic voting. In general, STV variants are relatively resistant to strategic voting while Monroe is susceptible to a wide range of strategies.

Visual reward models are essential for evaluating and improving visual generation models, yet existing approaches typically map task conditions and candidate outputs directly to scalar rewards, leaving implicit what should be evaluated for each individual case. We introduce Think Before You Score, a paradigm that explicitly determines what matters for each case before judging how well the candidate performs. Following this principle, we propose the Thinking Reward Model (TRM), which formulates case-adaptive rubrics, performs rubric-guided assessment, and produces fine-grained pointwise rewards. We further observe that conventional pairwise preference optimization can induce score polarization, and introduce Pairwise Dual-Group Relative Policy Optimization (PD-GRPO), which leverages pairwise supervision to improve reward discrimination while preserving fine-grained pointwise scoring. Extensive experiments on image generation and editing reward-modeling benchmarks demonstrate that TRM achieves state-of-the-art performance among open-source reward models while remaining highly competitive with proprietary alternatives. Moreover, using TRM as a reward for reinforcement learning consistently improves diverse visual generation models, demonstrating that its fine-grained, case-adaptive rewards translate into effective optimization signals for visual generation.

Model adaptation is typically governed by a fixed recipe, even though different update programs can produce substantially different behavioral outcomes. We introduce adaptation compilation, which reframes where, how, and to what extent a model should adapt as a joint prediction and decision problem. Rather than searching over candidate programs anew for each learning episode, a compiler learns from prior adaptations to predict a vector-valued counterfactual response surface over candidate programs—their expected effects on acquisition, transfer, boundedness, and preservation—and selects a program before adaptation begins. Because this predicted geometry captures multiple behavioral consequences rather than a single winner or scalar score, it can be reused under different downstream priorities without retraining. Across five learning types, preferred programs vary meaningfully across episodes, and this variation is predictable from pre-adaptation information. On Llama-3.1-8B, compiler-selected programs approach exhaustive search while outperforming global and objective-specific defaults. Replication on Gemma-2-9B preserves program heterogeneity and selection headroom, but shows that exploiting this headroom requires accounting for uncertainty when departing from strong defaults. Together, these results show that adaptation search can be amortized across related learning problems, turning prior adaptation experience into a basis for deciding how future learning should occur.

Many experiments have provided possible hints of gluon saturation, including geometric scaling in deep inelastic scattering (DIS) at HERA, the suppression of forward particle yields in d+Au collisions relative to p+p collisions at RHIC, and the suppression of away-side peaks in di-hadron correlations at forward rapidities in proton-nucleus collisions at RHIC and the LHC. Previous studies within the Color Glass Condensate (CGC) framework have primarily focused on fits to HERA deep inelastic scattering data, using the resulting dipole amplitudes to predict hadronic observables in proton-nucleus and proton-proton collisions. A simultaneous fit of DIS and forward particle production data, and a systematic study of their interplay, has so far been absent. In the work presented here, we focus on two processes: inclusive DIS and forward single inclusive hadron production in p+p collisions, incorporating data from HERA, RHIC, and the LHC. We solve the Balitsky-Kovchegov (BK) small-x evolution equation and perform a global fit to the combined datasets. Our analysis accounts for uncertainties in fragmentation functions and provides final uncertainties for the fitted parameters. The outcomes demonstrate the potential of the CGC formalism to provide a universal description of these diverse processes at the level of the dipole amplitude. In these proceedings, after very briefly summarizing the setup and our results, we highlight one particular aspect of the analysis, namely the complete uncertainty treatment. We validated the final uncertainties with two independent methods: the Hessian method and the Bayesian inference method, finding very good agreement.

Natural-sounding multi-agent conversational AI is increasingly deployed, fundamentally altering human-machine interaction and human information processing. While prior work largely focuses on algorithmic failure, this study investigates the cognitive ergonomics and socio-cognitive impact of algorithmic competence. We present and evaluate FORMS (Framework for Opinion and Rhetoric in Multi-agent Simulations), a low-latency architecture for spatially mediated human-machine dialogue, driven by distinct LLM-based personas and real-time concurrency resolution. To conduct a system test and evaluation of its psychological impact, we exposed an adolescent cohort (n=120) and an adult pilot group (n=25) to a live, moderated synthetic debate. Our findings reveal that exposure to highly competent multi-agent systems triggers "Cognitive Destabilization," fragmenting users' prior strategic consensus. Concurrently, we observe a "Regulatory Awakening" driven by the "Normality Paradox": fluid human-machine interactions inherently increase the baseline demand for external regulation. Furthermore, our pilot study suggests the presence of a "Truthfulness Paradox": despite understanding the risks of generative AI, participants in the adult cohort rated the synthetic debate as significantly more sincere than equivalent human discourse (Cohen's d=2.04). Supported by robust statistical effect sizes, this paper contributes the FORMS architecture and a replicable evaluation protocol, illustrating how high-fidelity conversational systems can reshape human information processing.

A phase that reaches a fault-tolerant processor in additive pieces can be remembered until the last piece arrives, or executed on arrival: the first option costs classical memory carried across rounds, the second costs magic states committed before the phase is known. We determine the exchange rate $\alpha$, committed $T$ gates per bit of memory forgone, for ancilla-free coordinatewise Clifford+$T$ compilation. The Ramanujan bound gives $\alpha\ge2$ with explicit constants, the square-root barrier of the spectral method. An elementary determinant method, using that quaternion numerators lie on spheres in both real embeddings of $\mathbb{Q}(\sqrt2)$, counts words near any rotation coset below that barrier. With a height dichotomy for the sphere sections it produces and a fibration over rational projections, it gives every $\alpha<5/2$ unconditionally and uniformly over frames, by an exact SMT check of its continuum case analysis; $5/2$ is where this method stops. At Clifford-framed cosets the volume law holds up to subexponential factors: all but a vanishing fraction of $z$-rotations need $T$-count $(3-o(1))\log_2(1/\varepsilon)$, and processes whose committed pieces are near Clifford-framed $z$-rotations, including per-rotation pipelines, have $\alpha\ge3-o(1)$, which a fractional-passthrough family attains under the Ross-Selinger typical-cost hypothesis. Under an equidistribution conjecture supported by exhaustive enumeration to $T$-count 22, $\alpha=3$ in general and memory should be shed in whole rotations. The bounds hold even when the phases cancel to the identity; side information enters through a conditional entropy; probabilistic mixing halves the costs but not the rate. With clean ancillas and a phase-gradient catalyst, table lookups batched over coordinates drive the rate to $O(1/\log\log(1/\varepsilon))$, so the constant-rate law is specific to coordinatewise synthesis.

Decentralized large language model (LLM) fine-tuning lets organizations collaboratively train a shared LLM on data they cannot pool, without a central coordinator. In every round, each node exchanges a trainable adapter with its neighbors over a communication graph, and then aggregates them. This setting, however, is vulnerable to propagated backdoors, which is a hidden behavior that lets a model perform normally on clean inputs but produce an attacker-chosen output whenever a secret trigger appears. We show that a single node poisoning its own model can backdoor adapters of nodes that have never seen a poisoned example, making them refuse prompts that contain a secret trigger. We present Chorus, a decentralized mechanism that lets each node detect and reject backdoored adapters from its neighbors before aggregation, without requiring shared validation data or any knowledge of the attacker's trigger or target. Chorus judges each adapter by its behavior, using the receiver's own adapter as a trusted reference. Crucially, no node in Chorus judges adapters alone: the receivers of each adapter update probe it independently, pool their findings in the neighborhood, and vote to make a decision. So a backdoor that slips past one receiver is still caught by the others. We evaluate the effectiveness of Chorus using two instruction-tuning datasets and LLM architectures, and against a state-of-the-art baseline. Chorus cuts the average attack success rate (ASR) of the attacker's neighbors from 48-63% to at most 2.2%, within 0.6 percentage points of an omniscient oracle that knows the exact malicious nodes. Even the worst-affected honest node never exceeds 10% ASR, the same bound as the oracle, against up to 78% without defense. This all comes at a negligible communication overhead.