Elinor Poole-Dayan, Jillian Fisher, Jacob Andreas, Mitchell Gordon, Michiel A. Bakker
原文摘要
Much research on responsible AI aims to change the behavior of AI systems that billions of people use. Pluralistic value alignment---the goal of building AI systems that represent and serve diverse human values and perspectives---has emerged as one such active research agenda. Yet, in an audit of the major frontier labs and open-weight developers, we find no public evidence that it has shaped how their models are trained or evaluated. Pluralistic alignment requires adoption to achieve its main goals, so we identify three barriers and lay out a roadmap of research to address each: 1. The justifications for pluralistic alignment so far are mostly normative or theoretical. Studies are needed that measure the benefits and drawbacks of pluralistic AI for users and society. 2. The pluralistic alignment research community has not settled when pluralistic behavior is warranted or what it should look like in practice. Research is needed to specify this behavior concretely enough for developers to operationalize. 3. Existing pluralism evaluations and methods are not ready for deployment. Evaluations and methods are needed that report their effects on other model properties and meet the requirements of production systems. These barriers are not unique to pluralistic alignment, so our roadmap applies to other areas of responsible AI research with similar theories of change. As people hand more decisions to AI systems and AI agents act on their behalf, we hope our roadmap helps this research reach the deployed models making those decisions.
The problem of decomposing an algebraic variety into more structured components (such as equidimensional, irreducible, or connected components) is a fundamental problem in computer algebra and algebraic geometry. In this work, we study the computational complexity of computing equidimensional components of algebraic varieties, represented via Chow forms. We improve the state-of-the-art complexity bounds from PSPACE to CH, the counting hierarchy, for a number of problems related to equidimensional decomposition, including deciding (1) whether the Chow form of the $k^{\text{th}}$ equidimensional component of a given projective variety vanishes at a given point, (2) whether a given monomial is in the support of the Chow form of the $k^{\text{th}}$ equidimensional component of a given projective variety, (3) membership in a given equidimensional component of a given algebraic variety, and (4) membership in the closure of the image of an irreducible variety under a polynomial map. As an application of our algorithm to decide membership in the closure of the image of a variety under a polynomial map, we show that the approximate polynomial satisfiability problem can be decided in CH. We also show that the degree of the $k^{\text{th}}$ equidimensional component of a given projective variety can be computed in $\text{FP}^{\text{CH}}$. Our main technical result is the construction of a polylogtime-uniform family of constant-depth, exponential-size threshold circuits that compute the Chow forms of the equidimensional components of a given projective variety. We construct these circuits by combining the machinery of geometric resolutions together with the uniform constant-depth, exponential-size threshold circuits of Andrews, Garg, and Schost (FOCS 2026) for the multivariate resultant.
Fine-tuning a large language model (LLM) on harmful data in a narrow domain, such as bad medical advice, can make it broadly misaligned. How realistic this risk is remains unclear, since models are very rarely fine-tuned on explicitly harmful data in practice. In this work, we show that fine-tuning on data the model believes to be false but that has no harmful content, can also induce broad misalignment. Across mathematics, trivia, cooking and sports, the more strongly the model disbelieves the answers it is trained on, the more misaligned it becomes. In mathematics, misaligned answers rise from 4% to over 60% as the training answers become more wrong, with little loss of general capability. After fine-tuning, the model states the wrong answers but its multiple-choice probabilities still prefer the true ones, which suggests that misalignment comes from this internal inconsistency. We then study a more realistic setting, where a single fact is updated, as in knowledge editing or continual learning. The same relationship holds and the more the model still believes the true fact after training on the false one, the more misaligned it becomes. When we frame the same false fact as part of a coherent story, the model comes to believe it more deeply and becomes less misaligned than when we instruct it to assert the false fact. Content filters and persona vectors cannot tell these two framings apart, even though they lead to different levels of misalignment. These results suggest a risk for model updates, where training a model that does not adopt the new fact as true can induce misalignment even when the data looks harmless.
We study the classical multi-item monopoly setting with a single additive buyer and $m$ heterogeneous items whose values are independent but not necessarily identically distributed. We compare the revenue (REV) of the optimal truthful auction, which in general may be randomized or complicated, with the revenues of three simple deterministic auctions: selling all items separately (SREV), selling them as a single grand bundle (BREV), and choosing the better of the two. Our technical cornerstone is a general upper bound on the optimal multidimensional revenue, parameterized by freely chosen item-wise caps, and expressed in terms of optimal single-item, Myersonian revenues and basic statistics of the value distributions. We derive this inequality through Lagrangian duality for a nonlinear programming formulation of the worst-case approximation ratio of separate-selling. Combining this inequality with carefully chosen caps and a probabilistic analysis of the resulting terms, we obtain improved approximation guarantees for all three deterministic auctions, both for general independent values and for iid values. Notably, we prove that $REV\le 2.5\max\{SREV,BREV\}$, improving the long-standing $5.2$ factor of Ma and Simchi-Levi and the $6$ factor of Babaioff, Immorlica, Lucier, and Weinberg. We complement this upper bound with a construction establishing a lower bound of $2.19$, improving the bound of $2$ by Rubinstein. For two iid items, we determine the exact approximation ratio of selling separately. Our proof constructs tight Lagrangian dual certificates for the aforementioned programming formulation, when item values lie in the grid $\{0,1/K,2/K,\dots,1\}$, for any discretization parameter $K$. Taking $K\to\infty$ yields the tight ratio $1+W(1/e)\approx1.2785$ in the continuous-valued setting, where $W$ is the Lambert-W function, closing the $[1.278,1.368]$ gap left by Hart and Nisan.
Unified Multimodal Models (UMMs) integrate visual understanding, generation, and editing within a single unified framework. However, existing unified multimodal datasets and benchmarks remain dominated by single-turn tasks. Even in recent multi-turn datasets, many turns can be solved independently without tracking information across turns. Consequently, they provide limited supervision and evaluation for multimodal stateful interaction, which requires models to understand, maintain, update, and faithfully realize evolving dialogue states across turns. To address this gap, we introduce YANAMI, the first large-scale multimodal dialogue dataset and benchmark for stateful interaction, comprising YANAMI-400K and YANAMI-Bench. YANAMI-400K contains 400K multi-turn multimodal dialogues, with more than 2M images and 2.6M dialogue turns. The dialogues span four representative scenarios: Multi-Image Fusion, Image Editing, World Modeling, and Visual Reasoning, and require models to continuously track, preserve, and revise evolving multimodal dialogue states across turns. YANAMI-Bench comprises 400 multi-turn dialogues for systematically evaluating multimodal stateful interaction. It jointly evaluates Dialogue State Understanding and Visual Realization, characterizing the gap between models' ability to understand evolving dialogue states and their ability to faithfully realize them in visual outputs. Experiments show that fine-tuning BAGEL on YANAMI-400K improves image editing, generation, and multimodal stateful interaction by 10.3%, 2.0%, and 13.23%, respectively, while preserving its general visual understanding capabilities. Evaluations on YANAMI-Bench further reveal that existing UMMs still struggle to maintain, update, and visually realize evolving multimodal dialogue states.
We aim to build a truly unified motion model that can understand, generate, and edit human motion within a single framework. Existing motion UMMs remain only partially unified: motion understanding is handled by a language model, while generation is often delegated to a separate diffusion module operating over learned motion latents. We introduce Möbius, which brings the entire diffusion process into the same multimodal representation space used for motion understanding. Clean motion, noisy motion, and text jointly participate in shared attention, while role-specific parameters process understanding and generation streams. Importantly, Möbius operates directly on continuous frame-level raw motion, requiring neither discrete tokenization nor a pretrained motion VAE. This simple formulation naturally supports motion captioning, text-to-motion generation, temporal completion, and instruction-guided editing within a single model. On HumanML3D, Möbius achieves state-of-the-art performance among unified motion models. Scaling to 1.7B parameters and 1,700 hours of motion data further demonstrates that our framework is both effective and scalable.
Conformal prediction gives finite-sample coverage when calibration and test points are exchangeable, but exchangeability can fail when deployment changes the data distribution. We study conformal prediction in performative settings, where the deployed predictor affects the data distribution, and seek prediction sets valid at performative equilibrium: coverage must hold under $D(f^\star)$ for a performatively stable predictor $f^\star$. Under total-variation $L$-sensitivity of the distribution map, we prove a phase transition at $L=1$. If $L \ge 1$, no finite-sample procedure can uniformly guarantee both equilibrium-valid coverage and nontrivial prediction sets. If $L<1$, the minimax excess miscoverage over informative procedures is $\widetildeΩ(1/((1-L)\sqrt{n}))$, implying a sample-complexity inflation of order $1/(1-L)^2$ for fixed excess miscoverage. Our main technical tool is a transfer theorem: any procedure returning a predictor within $\varepsilon$ of equilibrium yields equilibrium-valid coverage up to an additive $L\varepsilon$ term after split-conformal calibration. We instantiate the theorem with two retraining algorithms: a model-free method for discretizable classes and a parametric tail-averaged method that matches the $n$- and $L$-dependence of this rate up to logarithmic factors. Experiments, including semi-synthetic credit scoring, demonstrate the predicted threshold behavior and scaling laws.
Quantification of Tail-trimming Effects for Interval CVaR Based Nonlinear Regression
Yulei You, Junyi Liu
AI 速读 基于摘要
该文从理论上量化了In-CVaR风险度量对损失右尾的截断在非线性回归抗异常值中的必要性与有效性。
math.OC · cs.LG · stat.ML
查看原文摘要
The tail trimming strategy is widely adopted for robustness against outlier contamination in statistical learning and operations research. However, it generally lacks the thorough understanding that to what extent the tail-trimming technique is effective for nonlinear regression. In this paper, we address the robust nonlinear regression based on the so-called interval conditional value-at-risk (In-CVaR), which is introduced to guard against outliers by trimming extreme losses. We first analyze the trimming effect of the In-CVaR based estimator under distributional perturbation. Our first finding is that the In-CVaR based estimator is qualitatively robust under distributional perturbation in terms of the Prokhorov metric if and only if the largest portion of losses is trimmed. Furthermore, under sample contamination, we derive a probably approximately correct (PAC) bound for the In-CVaR based estimator, showing that the preliminary covariate filtering that is customary in the literature is not needed for In-CVaR based regression. The nonconvexity of the In-CVaR objective, together with unboundedness of covariates, responses and parameter space create the technical challenges for our developed theory, and thus necessitate the new treatment of the qualitative robustness and PAC analysis. Overall, this study rigorously analyzes robustness properties of In-CVaR based nonlinear regression models, which illustrates to what extent the right tail-trimming of the loss by In-CVaR risk measure is essential and effective for robust nonlinear regression. These theoretical findings are further supported by numerical experiments.
As agents become more capable, they are moving beyond one-off tasks toward sustained assistance across successive user goals. Such assistance requires agents to discover how unfamiliar environments work, reuse what they learn across tasks, and keep that experience current as environments evolve. Existing benchmarks increasingly study experience reuse and adaptation, but provide limited insight into which parts of accumulated experience agents revise, retain, or lose. To address this gap, we introduce ReScopeBench, a benchmark for the experience lifecycle across 18 interactive environments. It distinguishes structural experience about how an environment is organized and accessed from rule experience about reusable principles and constraints that guide decisions across tasks. To study how experience is acquired, reused, and revised, we propose an experience-first construction approach. It builds task streams around shared experience and introduces silent environmental changes. We evaluate task performance and track changes in saved experience throughout each stream. Across 20 models, we find that retained experience improves task performance, but agents struggle to keep it current as environments change. Even when the required experience is correctly recorded, agents do not always complete tasks successfully. These findings identify selective maintenance and effective use as complementary challenges for long-lived agents.
We attempt to alleviate failure modes of classical physics-informed neural networks (PINNs) in scenarios of topological pathologies corresponding to punctures, holes, poles, and voids by introducing de Rham cohomological augmentations, or Hodge-de Rham PINNs. Vanilla PINNs often encounter pitfalls in hole-based settings due to discontinuous interfaces, nonzero closed path integration, path dependence, correspondence to exact forms via the exterior derivative, etc., therefore we absolve traditional failures via parameterizing a 1-form by the Hodge-Morrey-Friedrichs (HMF) decomposition for manifolds with boundary, drawing a connection to de Rham cohomology such as through the HMF isomorphism theorem. 1-forms can act as representations of scalar-valued PDE solutions through integral operator mappings. We propose performing a PINN-based loss on the parameterized 1-form through an associated PDE, which subsequently solves the potential-based PINN problem and moreover remedies the failures of vanilla PINNs via the Sobolev-regular operator under a Fréchet differentiation interconnection. The operator relationship often has a nontrivial closed-form and often requires integration or a case-by-case analysis, which we address thoroughly in our experiments. When an integrability condition is not met on our losses, we can equip exhaustion functions to train the networks. We demonstrate our methods on the Aharonov-Bohm effect, which, to our understanding, is largely nonexistent in PINN literature. We develop theoretical guarantees of our methods and demonstrate our pathological safeguards on a broad selection of topologically-adverse problems, including scenarios of inexact forms, and those of nontrivial harmonics and coexact terms.
Artificial general intelligence (AGI) requires a method that can understand, generate, reason, and act across diverse domains and modalities. Rather than relying on a single model to master every capability, a promising path toward AGI now is to orchestrate a heterogeneous ecosystem of specialized models, tools, and policies (candidates) into a unified AI system. At the heart of this vision lies a fundamental question: how can an AI system dynamically select and coordinate the most capable candidate for each task while balancing performance and computational cost? In this work, we present Route Anything, a unified framework for studying model and policy routing across general-purpose AI systems. We develop a unified theoretical foundation that characterizes optimal routing under quality--cost trade-offs and establishes how the information available to a router determines its attainable performance. To systematically investigate routing across diverse capabilities, we first construct an omni-modal trajectory matrix spanning 25 benchmarks across language understanding, image, video, and speech generation, coding and workplace agents, and robotic manipulation, comprising 374,401 scored runs from 452 candidate configurations. Building on this framework, we train and test different routers to investigate the potential of omni-modal model routing, the capabilities of learned routers, and the conditions under which routing improves efficiency. Our results reveal substantial opportunities for capability orchestration: per-input oracle routing outperforms the best fixed candidate by 5.4 points, and learned routers outperform it by 0.3 points across 16 benchmarks, by 2.1 points on video generation, and by 1.1 points on robot manipulation while using 24% less policy time. More importantly, routing raises efficiency by 8.82x on average compared with the reference model.
Evaluating whether a large language model (LLM) genuinely knows a concept remains a fundamental problem, since behavioral metrics can be fooled by suppression: a model that refuses to answer may still harbor recoverable knowledge. The framework measures the geometric coherence of \textit{anon-vectors}, the differences in hidden states between a concept-specific query and its anonymized paraphrase, and shows that genuine knowledge forms a tight, low-entropy cluster while ignorance forces geometric fragmentation. Across ten instruction-tuned LLMs from three model families, all fragmentation metrics reliably separate known from unknown concepts, and post-unlearning, fragmentation geometry correlates with adversarial recovery across four attack types and five unlearning algorithms, though the direction of that relationship is shown to depend on attack methodology. Furthermore, a layer-resolved analysis reveals that different adversarial mechanisms exploit this geometry at varying depths, with high attack-specific correlations that are invisible to behavioral evaluation. Together, these results constitute a single-phase, attack-free proxy for unlearning robustness that captures aspects of parametric knowledge inaccessible to output-level metrics. A reference implementation is available at https://github.com/AdarPeleg/geometry-of-ignorance
Antibody affinity maturation is a natural optimization process in which B cells acquire mutations through somatic hypermutation (SHM), undergo selection for improved antigen binding, and iteratively evolve toward higher affinity. Computationally modeling this process requires capturing both biologically plausible mutations and antigen-specific selection. However, existing mutation models do not directly optimize binding, while affinity optimization often neglects realistic mutation trajectories. We introduce iGC (in silico Germinal Centers), a two-stage reinforcement learning (RL) framework for antigen-specific antibody optimization. Stage 1 learns an SHM mutation prior from parent-child pairs in B-cell lineage trees and adapts it to a target affinity-maturation dataset. Stage 2 initializes an RL policy from this prior and optimizes antigen binding with Kullback-Leibler (KL) regularization toward the frozen prior. To prevent reward exploitation, including repetitive high-reward substitutions and structurally tolerated framework mutations, we introduce a Reward-Tax Meta-Gradient objective. RewardNet smooths the affinity landscape, while TaxNet penalizes reward overexploitation, jointly defining a learned meta-reward for policy optimization. We evaluate iGC on mutation prediction, SHM consistency, affinity optimization, and mutation diversity. The mutation prior more than doubles the strongest baseline in mutation-site prediction. Following RL optimization, iGC maintains SHM sequence structure. It also achieves a strong exploration-exploitation balance, yielding competitive reward-diversity products across candidate-selection cutoffs. These results demonstrate that iGC integrates biologically plausible mutation generation with antigen-specific affinity selection, providing an in silico framework for iterative antibody affinity maturation.
Local LLM agents often work on shared hosts. Unlike multi-tenant cloud clusters, these hosts often have tightly bounded memory, CPU, and I/O resources. Concurrent agents frequently suffer from resource contention, leading to failures such as OOM and reduced resource efficiency. Considering that agents' resource demands are bursty and shell command delay is imperceptible to agents, there is a huge scheduling opportunity to arrange when and in what order commands are actually executed. Thus, we present ToolSched, a Shell MCP Server for shared local hosts that serves multiple agents and schedules their operations. ToolSched embeds a resource model into the MCP interface: agents estimate resource demands along with command requests. These hints, together with real-time host resource conditions, determine when commands are actually executed. Internally, ToolSched applies a scheduling algorithm to stagger commands, reducing contention and improving resource utilization. Prediction correction and checkpoint-based recovery mitigate estimation bias and resource errors. Our evaluation on Terminal Bench shows that, compared with unscheduled execution, ToolSched improves resource utilization by 148.4% and end-to-end command throughput by 29.4%, demonstrating its effectiveness and practicality.
Modern artificial intelligence models are deployed on inference engines to optimize runtime efficiency and resource allocation, particularly for transformer-based large language models. The vLLM project is a major open-source library to support model serving and inference. However, the current implementation of vLLM limits programmability of the internal states of deployed models. This prevents the use of popular test-time model alignment and enhancement methods. For example, it prevents the detection of adversarial prompts based on attention patterns or the adjustment of model responses based on activation steering. To bridge this critical gap, we present vLLM Hook, an opensource plug-in to enable the programming of internal states for vLLM models. Based on a configuration file specifying which internal states to capture, vLLM Hook provides seamless integration to vLLM and supports two essential features: passive programming and active programming. For passive programming, vLLM Hook probes the selected internal states for subsequent analysis, while keeping the model generation intact. For active programming, vLLM Hook enables efficient intervention of model generation by altering the selected internal states. In version 1 (named vLLM Hook MIA), the plugin runs on vLLM's V2 model runner and preserves the engine's CUDA graphs, so that capturing and steering internal states incurs only a small decoding overhead. It supports both offline use and serving through vllm serve. In addition to presenting the core functions of vLLM Hook, we demonstrate use cases including prompt injection detection, attention-based re-ranking for retrieval-augmented generation, activation steering, hidden-state probing, hallucination detection, schema linking for text-to-SQL, and tool-call risk assessment. Finally, we welcome the community's contribution to improve vLLM Hook via https://github.com/IBM/vLLM-Hook.
We introduce Gaussian partitioning, an algorithmic primitive for distributions smooth with respect to a Gaussian. It partitions space into polyhedral regions so that almost all mass lies in thin strips restricting a few orthogonal directions. Within each strip, the conditional distribution approximately matches low-degree Gaussian moments in the unrestricted directions. Our construction builds on logconcave partitioning of Diakonikolas, Kane, and Ma (FOCS 2026). Under input marginals smooth with respect to the standard Gaussian, we obtain agnostic learners with error $(1+c)\mathrm{opt}+ε$ for broad classes of Boolean multi-index models, including functions of halfspaces, low-dimensional convex concepts, and low-dimensional polynomial threshold functions. For functions of $k$ halfspaces, both agnostic and testable agnostic learners achieve this error with dimension dependence $d^{\mathrm{poly}(k/c)}$, improving the previous exponential dependence on $k/c$ in the agnostic exponent. For learning from positive examples with imperfect unlabeled data (PUI), we give algorithms and Statistical Query (SQ) lower bounds that qualitatively characterize complexity through Gaussian moment matching. For fixed smoothness and positive-mass parameters, we learn intersections of $k$ halfspaces with dimension dependence $d^{O(\log(k/ε))}$ and matching SQ lower bounds, improving the previous $d^{\mathrm{poly}(1/ε)}$ upper bound even for one halfspace. A general reduction from agnostic reliable learning to PUI yields negative reliable learners for these intersections with the same dimension dependence, for fixed smoothness. Finally, for PQ learning and testable learning with distribution shift, we prove qualitatively optimal SQ lower bounds of $d^{Ω(1/ε^2)}$, even for intersections of two halfspaces with Gaussian training data and arbitrary test distributions.
LLM-based agents, which increasingly operate persistently across sessions and accumulate memory through interaction, already exhibit the familiar symptoms of unchecked plasticity: semantic drift, degraded retrieval precision, propagation of stored errors, and saturation of memory capacity. Drawing on the selectionist "ringing, resonance and renormalization" (RRR) account of sleep, we propose that each of its three phases maps onto an operation feasible in digital memory architectures. Engineers building stateful agents, meanwhile, are independently rediscovering the problems sleep evolved to solve: what to forget, and when to do the forgetting. From this we derive the falsifiable prediction that selectionist consolidation will outperform instructionist alternatives as memory stores grow large. The core claim concerns the nature of continuous learning itself: plasticity has a price, and for any system that learns continuously under finite capacity while coupled to its environment, that price is sleep.
Learning on boundary representations (B-reps) requires robustness to the pose of a part, since CAD models are not stored in a canonical orientation. Existing B-rep networks either operate on world-frame coordinates and normals, and thus degrade under rotation, or encode each face in a local reference frame. Local frames yield invariance, but reduce inter-face messages to scalars and discard the relative orientation of adjacent faces. We propose SO3brep, an SO(3)-equivariant graph neural network on the B-rep face-adjacency graph. Node features are computed in local frames and mapped into a shared frame by Wigner D-matrices. Clebsch-Gordan tensor products couple these features with the directions between adjacent faces, so that relative orientation is preserved throughout message passing and the output remains SE(3)-invariant. Across three face-segmentation and two classification benchmarks with randomly rotated test parts, SO3brep achieves the highest accuracy among the compared B-rep networks and equivariant point cloud networks, with the fewest parameters. Its advantage is largest with few labeled parts, a setting typical of industrial CAD. Code is available at https://github.com/UGent-CVAMO/so3brep.
Given the full truth table of a Boolean function whose input is split between two players, we study the complexity of determining the minimum worst-case number of bits they must communicate to compute it. This is a basic meta-complexity question, posed by Yao in 1979. We prove NP-hardness of approximation within every fixed additive error for deterministic communication complexity, under deterministic polynomial-time reductions. The result holds both in the standard model and in Yao's original model, where the players alternate after every bit and only the receiver of the last bit must determine the answer. In particular, it resolves the exact computation question under Yao's original conventions. Earlier versions of this same paper established exact NP-hardness for the standard model, independently and concurrently with Hirahara, Ilango, and Loff. In both models, a polynomial-time estimator with any fixed additive error would yield a polynomial-time $3$-SAT algorithm, deterministic or bounded-error randomized according to the estimator. For the standard model, we also give quantitative implications for errors that grow with the number $T$ of table entries. A bounded-error randomized polynomial-time estimator with additive error $c\log T/\log\log T$, for any fixed $0\le c<1/8$, yields a bounded-error randomized $3$-SAT algorithm on $N$ variables and $O(N)$ clauses running in time $2^{O(N^α)}$ for every fixed $8c<α<1$. A deterministic polynomial-time estimator with additive error $c_*\log\log T$, for an effective absolute constant $c_*>0$, yields a deterministic algorithm running in time $2^{O((\log N)^2)}$ on such formulas.
Multimodal web agents have shown promising capabilities for interacting with websites through graphical user interfaces (GUIs), yet realistic web tasks often require long-horizon, multi-step planning that remains challenging for monolithic policies. Such agents must jointly infer high-level plans, navigate through intermediate states, track progress over extended interaction histories, and execute visually grounded actions. Although prior work has explored modular multi-agent systems with a separate planner, the planner is typically not updated jointly with the actor. We investigate whether explicitly separating and jointly training planning and execution provides a stronger inductive bias for long-horizon web navigation. We introduce JHiMA-WebRL, a joint hierarchical multi-agent reinforcement learning framework for web agents that couples a task decomposer with an actor. To train the decomposer, we use a strong teacher model to construct structured subtask supervision from multimodal GUI trajectories, and use these annotations to fine-tune both the decomposer and the subtask-conditioned actor. We then extend Group Relative Policy Optimization (GRPO) to jointly optimize the decomposer and actor in the interactive WebArena environment through hierarchical rollout grouping and role-specific advantages. On WebArena-Lite-v2, JHiMA-WebRL-14B, consisting of a 7B actor and a 7B decomposer, achieves an average success rate of 37.2%, compared with 24% for a ScaleCUA-32B agent that we fine-tune with RL on the same training tasks and evaluate under the same protocol. Ablations show that jointly updating the decomposer during RL outperforms updating the actor alone. These results suggest that jointly learned role factorization with smaller models can be a competitive alternative to scaling monolithic agents for long-horizon web tasks.
De novo peptide sequencing has moved from dynamic programming over hand-designed or generative scores to large transformers that learn the task end to end. We present NovoCRF, a conditional random field (CRF) whose states are prefix masses, so that its paths are exactly the peptides within +/-0.06 Da of the precursor mass. A small network scores each edge from physics-derived ion-match features and is trained with its scores globally normalized over all mass-feasible peptides. To our knowledge, NovoCRF is the first de novo method trained with exact normalization over all mass-feasible peptides. Our software has three modules, each with exactly one job: a precursor mass finder recovers the accurate precursor mass from MS1, the CRF finds the peptides that fit the precursor mass by exact k-best decoding, and two BiGRU rerankers add sequence context and internal-fragment evidence. The model has 310k parameters and trains in about 17 GPU-hours. On seven held-out data-dependent acquisition datasets published in 2026 (126,262 labeled spectra), NovoCRF's mean peptide recall is 42.4%, against 39.1% for Casanovo v5 with five beams, and its amino-acid recall 64.0% against 53.1%, at 1,746 spectra/s on 18 CPU threads, 55x Casanovo's throughput on a GPU. Split by fragment evidence, NovoCRF leads on labels with up to three peptide bonds without a b or y ion, and Casanovo only beyond. NovoCRF can also search modifications it was never trained on at their exact mass: it finds 32.1% of the hydroxyproline peptides of a lion bone sample.
Controlling the motion of multiple objects in image-to-video (I2V) generation requires preserving object identities while enforcing adherence to distinct target trajectories. This becomes particularly challenging as the number of objects increases and their paths intersect or occlude one another. Existing approaches entangle multiple trajectories within a shared, dense conditioning signal, making object-level correspondence difficult to preserve in crowded scenes. We depart from this paradigm and enforce a strict, per object spatial constraint that isolates instances independently. Our method, TrajLoc, achieves this directly within the attention layers by substituting the cross-attention weights of each object token with a Gaussian heatmap centered on its target location at every frame. The same per object token interface carries trajectory and depth through a learned embedding and preserves identity by encoding first frame appearance in place of an object token. Evaluations across six datasets, featuring up to 20 simultaneously controlled objects and out of distribution real world scenes, demonstrate that our method consistently improves both visual fidelity and trajectory adherence. Applied to two architecturally distinct backbones (CogVideoX 5B and WaN 2.1 14B), our approach achieves average gains of +4.3 dB PSNR and a 51% reduction in trajectory end point error compared to the strongest baselines. Project page: https://sela-omer.github.io/traj-loc/
The distortion framework in social choice theory allows quantifying the efficiency of (randomized) selection of an alternative based on the preferences of a set of agents. We make two fundamental contributions to this framework. First, we develop a linear-programming-based algorithm for computing the optimal randomized decision on a given instance, which is simpler and faster than the state-of-the-art solutions. For practitioners who may prefer to deploy a classical decision-making rule over the aforementioned optimal rule, we develop an algorithm based on non-convex quadratic programming for computing the exact distortion of any (and the best) randomized positional scoring rule. For a small number of alternatives, we find that the exact distortion bounds are significantly better than the asymptotic upper bounds established in prior literature and lead to different recommendations on which rules to use. These results rely on a novel characterization of the instances yielding the worst distortion, which may be of independent interest.
Byte level language models avoid tokenizer dependence and provide universal coverage, but processing every byte at full resolution is expensive. We introduce DeepMem, a causal multiscale architecture that separates semantic compression from exact byte preservation. A learned invertible subband transform constructs a hierarchy of shorter sequences, while a lossless carrier and explicit detail sidebands preserve byte identity. Fast weight memory blocks operate along the encoder, with iterative shared state computation at the compressed bottleneck, while an attention based super resolution decoder restores byte resolution representations. DeepMem achieves stronger language modeling performance than a matched byte level Transformer on FineWeb-Edu and PG19, and obtains the best macro average performance across the evaluated FLORES languages. It also improves PIQA accuracy while remaining competitive on code modeling, multiple choice reasoning, and byte level corruption robustness. Bottleneck refinement requires approximately $7{,}600\times$ fewer counted FLOPs than equivalent full resolution reasoning, while maintaining competitive evaluation throughput.
Predicting social network churn before it becomes visible in aggregate metrics remains an open challenge, particularly in enterprise settings where heterogeneous usage patterns, sparse data, and deployment constraints preclude standard supervised approaches. We formalize this problem as finite-step engagement projection over contributor-role Markov chains, with a stationary-distribution interpretation for ergodic chains in the long-run limit. For each engagement feature type within a network (e.g., Q&A, communities, articles, etc.), users are classified monthly into four behavioral states and a transition matrix is estimated from six months of history. Six-step propagation yields a label-free projection of the contributor mix; for ergodic chains, the Perron-Frobenius theorem establishes convergence under repeated propagation. Comparing projected future engagement to current engagement produces the Health Ratio, a single score requiring no trained model, no labels, and no infrastructure beyond SQL. Validated across seven rolling prediction windows on 11,500+ tenant-level networks, the Health Ratio achieves mean AUC 0.917 with a 156 times churn-rate gradient across risk tiers, outperforming all single-signal baselines and the growth-adjusted variant previously deployed in production. An ablation study shows that six-step propagation achieves similar churn discrimination to the eigenvector-based score. Deployed on Microsoft Engage since October 2025, the framework powers customer health monitoring and has enabled proactive intervention on at-risk networks before decline becomes visible in traditional metrics.
Task instructions from users for agents are inherently underspecified, especially regarding safety boundaries--creating behavioral risks. To study these risks, we introduce BARC-Bench, a collection of 110 tasks across nine simulated workspace environments where benign requests leave openings for unsafe actions. In this benchmark, we evaluate agent behavior against safety boundaries established by workspace records, implicit environmental context, and unstated social norms. For example, an agent asked to schedule off-boarding discussions must arrange the meeting without prematurely revealing an employee's resignation through a shared calendar invitation. Evaluating four frontier model configurations across five attempts per task reveals that models are not consistently safe: the share of tasks completed safely on all five attempts ranges from 9.1% to 72.7%, with even the strongest configuration producing a safety violation on 27.3% of tasks. These results highlight an asymmetric operational risk: while models frequently complete complex workflows, their inability to consistently respect unstated boundaries presents a substantial barrier to safe autonomous delegation.
Large language models understand passengers' driving commands, but as driving agents they reason in generated text for seconds and must infer each action's consequences implicitly from the scene. We present Jev-AD, which pairs Jev, a typed decision model that answers predefined questions with probabilities, not generated text, with a language world model that predicts for every candidate action how nearby vehicles move over the next two seconds. Code turns these predictions into foreseen consequences, which Jev compares instead of inferring them from coordinates. In simulated highway merging (500 car-episodes), this foresight lowers the same decision model's collision rate from 26.0% to 2.4% at 0.25 seconds per decision, and Jev-AD achieves the passenger's intent in 93% of free-form command trials. In 500 replayed real-world NGSIM episodes, foresight lowers the collision rate from 48.2% to 2.8%, and the world model, learned from this traffic, foresees 93% of dangerous actions. Project page: https://ys-qu.github.io/jev-ad-website.
We study the degree needed for sum-of-squares (SoS) proofs of several basic principles of real linear algebra. For any $n\times n$ matrix $A$ of rank greater than $r$, the Weak Rank Principle states that the equations $XY^\top=A$ with $n\times r$ matrix variables $X, Y$ have no solution. We show that this principle requires degree exactly $2r+2$ to prove. On the other hand, for $A=I_n$, setting $X = Y$ gives the Tight Frame Principle, which has a degree-$4$ proof. For $n\times n$ matrices, proving that a right inverse is also a left inverse requires degree exactly $2n+2$, while proving that orthonormal rows imply orthonormal columns again has a degree-4 proof. Using our lower bounds, we also derive related SoS degree lower bounds for polynomial systems encoding sign rank, matrix rigidity, small-width homogeneous non-commutative algebraic branching programs, and tensor CP rank. Lastly, we show that quantitative bounds on the Frobenius norms of the factors in the Weak Rank Principle can lower the degree required to refute $XY^\top=I_n$, but sufficiently large bounds still preserve the exact degree $2r+2$ from the unbounded case, giving a quantitative version of the Archimedean property arising in classical results on SoS proofs.
Vision-language-action (VLA) models increasingly pair a pre-trained vision-language model (VLM) with a flow-matching expert that conditions its action predictions on the VLM's key-value representations. Those representations must carry information useful for control, yet the VLM was never trained for the action expert's flow-matching objective. Two signals can address this: gradients backpropagated from the expert's loss (coupling), or VLM supervision to predict discrete tokens encoding the same target actions. Prominent training recipes forgo the first, blocking expert-to-VLM gradients (insulation) to protect the VLM's existing capabilities, while also supervising it on vision-language data and action tokens. Because vision-language co-training also aims to preserve those capabilities, the additional protection provided by insulation is unclear. At the same time, insulation removes a direct signal about which features continuous action prediction requires. On the other hand, action-token supervision offers an alternative way to learn action-relevant representations, which may explain why insulation succeeds in practice. We disentangle these factors in autonomous driving and manipulation by independently varying action-token supervision and whether the expert's loss updates the VLM. In open-loop driving, insulation is competitive with coupling when action tokens are supervised; without them, coupling reduces trajectory-prediction error by roughly 25%. When both components are trained jointly from the outset, coupling increases closed-loop driving route progress by approximately 17% with action-token supervision and 40% without it, and similarly improves robot manipulation performance under both settings. These findings show that the value of gradient insulation depends on the VLM's action supervision and the evaluation setting, challenging insulation as a default VLA training recipe.
A computational path records the steps used to establish an equality. Topologies on computational equality witnesses can depend on the chosen generators and execution schedule, even when their algebraic and homotopical meanings agree. We study continuous presentations of finite computational paths whose declared rewrites are realized by endpoint-fixed homotopies. We compare a topology retaining the full signed trace with an observable topology recording endpoints, length, and exact path realization. Continuous finite-word interpretations that are inverse up to scoped rewriting preserve trace-sensitive quotient arrows and their final composable-pair spaces. Observable quotients behave differently. Over a point with a nonempty primitive alphabet, we classify their topology: it records length parity when all named relations preserve parity, and is otherwise indiscrete. Adjoining a conservative abbreviation can therefore turn a discrete two-element quotient into an indiscrete space, even though ordinary multiplication remains continuous. A circle presentation also gives nonhomeomorphic quotients under binary and equal-slot timing. Geometric completeness alone does not identify the trace-sensitive and observable topologies. Multiplication is continuous on the quotient of composable representatives. We characterize when this topology agrees with the ordinary pullback topology and give an open-projection criterion. Universal presentations recover classical fundamental-groupoid topology. Normal forms, primitive interchange, and a Cayley covering prove completeness for circle, torus, product, and rose presentations. A literal-word correspondence relates the semantics to the reflexivity, symmetry, and transitivity fragment of equality derivations. Selected results are formalized in Lean, with the realization models and proof coverage stated explicitly.