Frontier web agents using closed models have high inference and computational costs. Smaller open models can be an alternative, but to be performant, they require training supervision from a human or teacher model. Token-level teacher distillation is not an option for closed teachers as it needs access to the model's logits. The standard approach is to imitate teacher demonstrations offline, but it trains the student on states it has never visited, and in a style it does not write in. We introduce WebKD, where the teacher is the editor of the student's failures. Given a failed episode, the teacher finds the first point of mistake, and rewrites the student's reasoning traces, harness states, and actions from that point onward. The editing reveals that the teacher's corrections can be more surprising to the student than the student's own text. WebKD instead aims to edit minimally. In the data space, the teacher writes each correction in the student's style. In the objective space, the reasoning is weighted by its contribution to the action. A role-asymmetric KL anchor allows the corrected action to move freely while keeping the reasoning near the student's prior. On the TimeWarp benchmark, WebKD students gain about 5.5-10.7 pp over behavior cloning, outcome-reward RL, and teacher editing baselines. The performance transfers to unseen versions and to the live web in the OnlineMind2Web benchmark. WebKD opens a new way of distilling black-box models to improve the capabilities of open-weight web agents.

Multimodal Large Language Models (MLLMs) are increasingly deployed in safety-critical domains, yet their susceptibility to backdoor attacks remains a formidable challenge. While existing defenses have evolved from visual-centric patch cues to unsupervised internal attention statistics, they are fundamentally constrained by a reliance on static snapshots and attack-specific heuristics. Given the intricate interplay between dual modalities, the vast design space for multimodal triggers allows diverse threats, such as textual injections, sub-perceptual noises, and coordinated cross-modal triggers, to easily circumvent such rigid priors, particularly when attack types are unknown in real-world applications. In this paper, we rethink backdoor detection by shifting the diagnostic paradigm from static observations to dynamic output evolution. We observe that a successful attack inevitably hijacks the model's decoding trajectory, skewing internal modal contributions during generation while maintaining an abnormal output invariance despite input perturbations. To capture this behavior, we propose EVO-Detect, a unified detection framework that monitors the dynamic evolution of internal signals during early decoding steps through two complementary channels: Attention Flow and Logits Flow. The attention branch measures how image-text attention allocation evolves during early generation, whereas the logits branch quantifies the sensitivity of the output distribution to modal perturbations through interventional ablations. We then convert these trajectories into compact sample-level metrics and normalize them against negative-control signatures so that natural perturbations are less likely to be mistaken for backdoor evidence. Extensive evaluations show that our approach substantially outperforms state-of-the-art baselines such as BYE and TCAP across diverse attack families.

Distilling capable language-model agents into smaller models requires aligning the student's behavior with the teacher's. This raises a fundamental question: at which states and at what granularity should this alignment be performed? Offline imitation provides dense but token-specific supervision and can suffer from state-distribution mismatch. Token-level on-policy distillation (OPD) addresses this mismatch but typically requires teacher token probabilities and compatible tokenization. We argue that semantic actions provide a natural unit of supervision for agent distillation. We introduce ActionOPD, an action-level on-policy distillation method motivated by maximizing expected teacher-action likelihood at student-encountered states. At each such state, ActionOPD samples actions from both the student and a black-box teacher, rewarding student actions according to their semantic agreement with the sampled teacher actions. Through policy optimization, ActionOPD aligns the student's action-level behavior with the teacher's decisions without requiring reproduction of teacher-written reasoning or access to teacher token probabilities. Training Qwen3-4B and Qwen3-8B with two teachers on diverse, curated tool-use tasks yields average gains of 3.3-6.9 percentage points across four out-of-distribution benchmarks. ActionOPD scales favorably with student size and teacher capability, without requiring the teacher and student to share a model family or tokenizer. On $\tau^2$-bench, ActionOPD achieves competitive performance against offline distillation and token-level OPD. We further explore Self-ActionOPD with a same-model teacher and privileged context. These results establish ActionOPD as a promising approach to agent distillation.

Energy-based models can represent distributions over discrete structures such as molecular graphs, but an energy that is accurate at equilibrium need not yield accurate samples when the sampler stops after a fixed number of steps. We introduce finite-horizon energy learning, which adapts the acceptance energy of a Metropolis--Hastings sampler to the distribution it returns after a prescribed number of steps, with initialization, proposals, and annealing schedule held fixed. For graph generation, a frozen pretrained model supplies initialization and proposals, and a learned energy governs acceptance. A training-only discriminator scores completed samples, and coupled continuations of the accepted and rejected branches of a proposal give an acceptance-gradient estimator that is unbiased for a fixed discriminator under stated conditions, without equilibrated negative samples or differentiation through discrete states. Generation uses the learned energy in the original single-trajectory sampler, without discriminator evaluations or reranking. On twelve pairwise binary targets with exactly evaluated terminal laws, learned energies have lower median terminal total variation than the equilibrium-correct energy in every tested short-horizon setting, and than its numerically tuned scalar multiple in most settings. On MOSES, a learned two-coefficient combination of pretrained Graph Energy Matching (GEM) potentials lowers FCD by 3.7--10.5\% relative to scale-calibrated GEM at matched sampling steps, with validity gains at the shortest budget and a sub-point validity trade-off at higher budgets.

Sampling multiple chain-of-thought rationales from a language model can uncover correct solutions missed by a single attempt. However, these rationales are typically lengthy, making repeated autoregressive generation costly. Continuous latent reasoning offers a compact alternative, but point-valued methods condition repeated answer sampling on a single rationale, which can limit exploration of diverse paths and answer coverage. We thus propose rationale distribution distillation (RDD), training a conditional flow over fixed-length latent sequences by distilling from multiple textual reasoning paths. Furthermore, we expand the distillation targets through within-query mixup, interpolating between complete latent rationales for the same query. The flow nonautoregressively generates the entire latent sequence, and a frozen language model autoregressively decodes answers from sampled rationale contexts. Empirically, the resulting distributional reasoner achieves higher answer coverage than the evaluated point-valued latent reasoning baselines, particularly on advanced mathematical tasks.

Spatio-temporal forecasting is fundamental to many real-world systems, yet deployment environments are inherently varying: temporal distributions may evolve, spatial domains may change, and observation availability may shift from training conditions. Existing models are often tightly coupled to their training environments through node-specific representations, predefined spatial support sets, and reliance on consistently available observations, limiting generalization to changing environments. To this end, we propose VaryST, a compact spatio-temporal forecasting model trained from scratch that achieves strong generalization by dynamically constructing predictive context from available long-term observations. To accommodate temporal changes, VaryST derives node representations from long-term histories, capturing current behavior in the context of longer-term patterns. To adapt to spatial changes, it dynamically selects functionally relevant support nodes beyond predefined neighborhoods and integrates their information with recent and long-term target states and global context. To handle varying observation availability, VaryST selectively masks historical patches on which the model relies most during training, encouraging it to exploit alternative temporal and spatial dependencies for accurate forecasting when key observations become unavailable. Extensive experiments on six real-world datasets against seventeen baselines show that VaryST achieves strong in-domain forecasting accuracy, substantially improves zero- and few-shot generalization under spatial and temporal changes, and remains robust to varying observation availability.

Large language models (LLMs) hold broad knowledge in one set of weights, but lack interpretability and trustworthy reasoning, especially in high-stakes use cases. A neuro-symbolic model, with its interpretable architecture and System 2 reasoning capacity, is a promising alternative if built at the scale of LLMs; however, its symbolic backbone is usually constructed with human effort, so existing neuro-symbolic models are bottlenecked by the scope of the symbol component and usually domain-specific. A data-driven method could instead build a larger symbolic system by aligning and merging existing ones. We target its first, foundational step: matching units across these heterogeneous systems. Our method, SAE-IDF, encodes a symbol in its native context into a sparse code with a sparse autoencoder (SAE), re-weights each dimension by its rarity and matches across systems. On 12 tasks across 6 kinds of symbolic system, SAE-IDF outperforms the same LLM's dense baseline on all tasks with Gemma-2-9B and all but one with Llama-3.1-8B, by up to 45 points, consistently across SAE layers and configurations. SAE-IDF is also interpretable: every dimension of the sparse code has a human-readable meaning, so each match decomposes into named concepts (features). Our work paves the way to a universal symbolic backbone.

We propose DAGFormer, a Transformer language model whose cross-layer connections are planned by a jointly trained topology predictor. Beyond the residual connections in a standard Transformer, DAGFormer adds direct connections from earlier layers to later ones, and the predictor controls the strength of every connection based on input tokens. Unlike existing cross-layer routing architectures, which set each layer's connections separately, DAGFormer plans the routing of the whole model at once. On matched OLMo-2 pairs from 75M to 1B parameters, DAGFormer lowers WikiText-2 perplexity by 17% to 25% over the vanilla Transformer, and a 1B DAGFormer trained on 5B tokens outperforms the vanilla model trained on 10B. Against four routing baselines with the same data and budget, it reaches the lowest perplexity at 75M, 300M and 1B. The gain carries over to 13 commonsense, science, reading and math evaluation tasks (2.0 to 2.5 points higher mean accuracy at 300M and 1B) and to instruction finetuning (63% to 70% AlpacaEval win rate against the vanilla Transformer, leading every routing baseline), and under structured pruning DAGFormer degrades the least in five of six settings.

Repeated event forecasting requires extracting predictive signals from noisy evidence and evaluating their value across successive decisions. We introduce JockeyBench, a benchmark and evaluation protocol that studies this problem through historical Japanese horse racing. The protocol combines time-gated evidence, priced tools, paired frozen-program controls, and two workflows: executing an LLM-written predictor or acquiring evidence for a fresh per-race LLM forecast. Evaluation separates forecast quality, benchmark-specific net utility, and retrospective betting returns and risk under fixed decision rules. Across 28 model--scaffold configurations on 324 races, paired tests identify seven revision gains and two declines after multiple-testing correction. Yet every configuration has negative mean profit under capped-Kelly settlement. In a four-model comparison, the reusable-program workflow has higher mean betting profit and lower research fees, although both workflows lose on average. Further diagnostics reveal that mean probability bias reverses on selected bets and that profitable individual runs can depend on concentrated returns. These findings motivate evaluating predictive competence, adaptation, resource use, and decision value separately when assessing agents for repeated event forecasting.

Multi-session agents can switch between LLMs to balance performance and inference cost. Such switches require transferring a history for continuity, yet their effects on the recipient remain underexplored. Through controlled experiments with multi-session tasks, we uncover behavior transfer: recipients tend to adopt reasoning and action patterns from the models that generated their histories. This effect is most evident with raw-history transfer but is substantially reduced by summarization. Based on these findings, we propose Direction- and Authorship-aware Summarization of History (DASH) that preserves stronger-model behavior while minimizing weak-to-strong transfer via summarization. Experiments with repeated swaps show that DASH outperforms raw-history transfer and naive summarization baselines. These results suggest that effective multi-model orchestration requires considering not only which model to use, but also which behavioral patterns to carry forward.

Instruction-Conditioned Motion Control (ICMC) separates parameter specification, instruction generation, and code synthesis for language-driven robot control, making execution outcomes traceable to instructions and programs. We compare one-shot generation with ICMC under o1 and o4-mini on cube pick-and-place and letter-reshaping tasks for a UR5e arm in MuJoCo. We then train an open-weight instruction policy by GRPO through a frozen code generator on cube placement onto two- and five-cube letter layouts, averaging the rewards of five programs per instruction. The logged positive-reward rate rose from 31.6% to 35.3% (253/800 to 282/800) between the first and last five training steps. A log audit identifies positive binary labels associated with geometric-failure messages: at least 19 in that run, 18 in the dense-reward run, and 5 in the run with one program per instruction. In one case, all five rewards were 1.0 while the log reports a 0.9515 m miss against a 0.05 m tolerance; the verdict came from the program's own output stream. We distinguish logged reward rates from independently verified task success and report descriptive results for each configuration's single training trajectory. We introduce an independent evaluation protocol that checks task completion against fixed specifications, separating robot control from success assessment to support reproducible evaluation.

Large language model agents increasingly undertake long-horizon tasks in complex environments characterized by fragmented evidence, partially completed states, and dependencies across resources and artifacts. Existing agents may satisfy individual requirements while failing to maintain these relationships, producing locally correct but incomplete outcomes. These failures arise because task completion depends on relationships among environmental evidence, consequential decisions, and required effects. To address this completion gap, we propose Interlock, a framework for task synthesis that represents each task as a typed completion graph. Grounding edges connect public evidence to decisions, coupling hyperedges impose joint constraints across components, and realization edges connect decisions to all required downstream effects. Interlock compiles this graph into a resettable workspace, a public request, and an evaluation rubric. A source-coverage review checks whether the original requirements survive generation, while a public-input review checks whether the rubric is supported by the materials available to the solver. An agent-as-judge inspects execution artifacts and runs verification scripts, retaining trajectories judged to satisfy all mandatory requirements. To evaluate training utility, we fine-tune two Qwen3.6 backbones on their own accepted trajectories. Across the two backbones, training improves AutomationBench task pass rates by 4.00-4.50 percentage points, JobBench weighted scores by 3.20-6.64 points, and Harvey LAB criterion pass rates by 3.89-4.16 percentage points. These results demonstrate the training utility of the synthesized experience across external task environments.

We study proportional cake cutting under an equal-length constraint, where each of $n$ agents must receive a bundle of length exactly $1/n$ that is worth at least $1/n$ to that agent. Our main result determines the exact worst-case number of cuts needed to guarantee such an allocation, namely $2n-2$ for every $n$. This improves on the previously known quadratic upper bound. The proof combines the constrained necklace-splitting theorem with a recursive assignment argument based on Hall's theorem. The same bound extends to the stronger requirement of envy-freeness when $n$ is a prime power, and an $O(n^{1.525})$ upper bound holds for arbitrary $n$. We then ask whether such allocations can be computed. We first show that no deterministic protocol in the Robertson--Webb model terminates after finitely many queries on every instance and always returns an exact equal-length proportional allocation, even for two agents with identical valuations. For every $\varepsilon>0$, however, $O(n^2/\varepsilon)$ queries suffice to produce an allocation in which every agent receives a bundle of length exactly $1/n$ and value at least $1/n-\varepsilon$, again using at most $2n-2$ cuts. Finally, we augment the Robertson--Webb model with two-knife queries. We show that recursive balanced halving attains the optimal bound of $2n-2$ cuts when $n$ is a power of two. For arbitrary $n$, we give a procedure using $O(n\log n)$ cuts.

In this work, we reframe agentic trajectories as paths through a shared embedding space of actions and environmental observations. We show that successful and failed rollouts of the same task separate in this space, with simple per-task discriminators predicting held-out rollout outcomes at up to 0.95 AUROC across three agent benchmarks and three policy models. To exploit this signal across tasks, we introduce Precog-RM, a transformer-based model that meta-learns a representation in which outcome prediction can be performed in a few-shot setting. We show that Precog-RM generalizes to unseen tasks and models, and gives confident predictions of trajectory success long before termination. Deployed as an early stopping controller for Group Relative Policy Optimization (GRPO) on Tau2-Bench, Precog-RM cuts training FLOPs by up to 47.3\% while maintaining performance comparable to baseline GRPO.

As large language models become widely used in scientific review and manuscript revision, AI scores are increasingly serving as feedback for assessing paper quality and guiding edits. Yet whether a single AI score reliably reflects the reviewer's judgment, and whether revision gains based on that score are endorsed by other reviewers, remain understudied. Using 200 papers from ICLR 2025 and 2026 and approximately 14,000 AI reviews, we study score reliability, score-distribution estimation, and cross-model agreement on revision gains. Repeated reviews fluctuate even under identical conditions. Finer-grained scoring protocols reduce score variance but also systematically shift mean scores, while the gains assigned to manuscript revisions depend strongly on the reviewer.

Simulation and experimental measurements provide complementary data for learning spatiotemporal physical systems, but standard simulation-to-experiment fine-tuning optimizes only the experimental objective after transfer and can degrade simulation performance. We formulate simulation--experiment prediction as a multi-objective learning problem with domain-specific simulation and experimental risks. On four fluid systems from RealPDEBench and two model capacities, we compare Simulation only, Experiment only, Sim$\rightarrow$Exp, and Joint training, evaluating every final model on both held-out domains. Sim$\rightarrow$Exp tends to specialize more strongly to experimental data at the cost of simulation-domain forgetting. Joint training consistently achieves the best balanced performance over a broad range of simulation--experiment evaluation weightings, while substantially improving simulation retention over Sim$\rightarrow$Exp. Joint also better preserves simulation-only fields absent from experimental measurements. Project page: https://mahindrautela.github.io/morph.

As machine learning increasingly relies on public, untrusted data sources, data poisoning attacks, which inject malicious examples into training data to induce misclassification of a chosen target, pose a growing threat. Existing defenses either assume zero ground-truth information about which examples are poisoned, or they assume access to a large set of examples verified to be clean. Satisfying the latter assumption incurs significant cost since reliable verification can be very resource- or labor-intensive. This cost is particularly high for clean-label attacks, where poisoned examples are visually indistinguishable from clean data. Since requiring a large set of verified examples is impractical, we propose relying on a small set of verified examples including both clean and poisoned ones, i.e., each example verified either to be clean or poisoned through inspection by a forensic expert. The challenge is then to detect poisons based on a set of verified examples that is so small that most classification models would overfit. To address this challenge, we propose Similarity-based Approach for Ground-truth-driven Exclusion (SAGE), which trains a generic feature extractor on a separate dataset and then flags poisoned training examples using a non-parametric, similarity-weighted prediction based on the verified set. On standard benchmarks against seven clean-label attack methods, we demonstrate that having access to even a handful of verified poisoned examples provides a substantial advantage. We also find that the distribution of verified clean examples across classes matters more than the number of verified examples.

Step-size selection remains a central challenge in large-scale neural network optimization; conservative steps slow convergence, while aggressive steps can destabilize it. We combine \textbf{Z}ero-and-\textbf{F}irst-\textbf{O}rder optimization~(ZFO) and propose a lightweight framework that decouples direction selection from step-size. ZFO uses a trusted first-order optimizer to determine the direction and performs zeroth-order evaluations only along this one-dimensional subspace to choose how far to move. Using the current {gradient information} and two additional objective function evaluations, ZFO instances construct a local model of the objective function along the proposed direction and select a curvature-aware step within a bounded search interval. This yields an adaptive step-selection mechanism that costs less than a full line search. We provide theoretical guarantees to show that shared-sample evaluations produce reliable finite-difference curvature estimates, that the induced local model selects a near-optimal step along the search interval, and that ZFO converges to a neighborhood of a stationary point. Across the evaluated settings, language models and datasets, ZFO frequently improves optimization and final performance relative to fixed-step first-order baselines, with the magnitude and preferred local model depending on the objective. Our code is publicly available at: https://github.com/nizswan/Zeroth-First-Order-Framework.

Given a set of $n$ points $X$ in a metric space and an integer $k$, max-min diversification aims to select $k$ points of $X$ maximizing their minimum pairwise distance. This objective function is however highly vulnerable to noisy points. In[Amagata, AAAI23], a robust formulation is proposed which addresses this vulnerability by excluding solutions containing any of $z$ outliers, defined as the $z$ points in $X$ with the largest nearest-neighbor distances. That paper also presents a coreset-based streaming algorithm for the new formulation, based on a suitable inlier-outlier separation assumption. However, we identify three shortcomings in the algorithm by [Amagata, AAAI23]: its coreset construction requires an offline computation over $X$, which needs memory linear in $n$, in stark contrast with the typical goals of stream processing; the one-pass procedure used to extract the solution from the coreset may return fewer than $k$ points (hence, an unfeasible solution) because it permanently discards points too far from the current solution; and its outlier-exclusion guarantee is only probabilistic and weakens as the coreset size shrinks. In contrast, we present a deterministic coreset-based algorithm that, under a natural inlier-outlier separation assumption (similar to the one used in [Amagata, AAAI23]), returns exactly $k$ inliers which are a $(2+\varepsilon)$-approximate solution, for any $\varepsilon>0$, thus only $\varepsilon$ above the best polynomial-time sequential approximation, even without outliers. Its one-pass streaming implementation adapts obliviously to the dataset's doubling dimension $D$ and, for wide ranges of $k$, $z$, $\varepsilon$, and $D$, it uses memory independent of $n$. For sufficiently long streams, its amortized update time is proportional to the coreset size, thus also independent of $n$.

Inference-time scaling adapts pretrained diffusion models to new sampling tasks without additional training. Existing methods rely primarily on Monte Carlo sampling with more particles, yet are premised on the pretrained model being exact. In practice, data and training limitations make the model imperfect, and these methods inherit its error. More particles reduce Monte Carlo error but cannot remove the mismatch between the endpoint and the desired target or the error in tracking the prescribed probability path. We introduce the Energy-based Feynman-Kac Corrector (EBFKC), a framework for energy-based diffusion models that corrects these errors on the fly given a reference energy. We first derive Feynman-Kac dynamics that track a prescribed path exactly in the continuous-time population limit even when the model is imperfect, and approximate these dynamics using sequential Monte Carlo with variance-controlling guidance. To remove the endpoint mismatch, we use the pretrained energy as a surrogate along the diffusion path and progressively incorporate the discrepancy between the learned and target terminal energies. Experiments on Gaussian mixture models, particle systems, alanine dipeptide, and alanine tetrapeptide show that our method closely matches target distributions and molecular free-energy profiles under annealing and reward tilting, whereas standard inference-time scaling baselines retain substantial sampling errors.