Many agent tasks recur with new data or small changes, yet agents regenerate their reasoning and actions on each run, producing different results. Distilling a prior run into a skill helps the agent avoid earlier mistakes, but it still re-runs the full process each time, so runs can vary and incur that cost. We compile a successful agent run into a reusable neuro-symbolic program (NSP) with a set of guidelines we call FlashScript. Actions and data transformations become code, instance-specific values become parameters, and decisions that need model judgment become LLM calls. We identify three layers where these calls belong and three techniques that replace most of them with code. The compilation does not remove nondeterminism; it localizes it, so the deterministic program and its neural calls can be reviewed and tested separately. Across 120 WorkBench analytics tasks with Claude Haiku 4.5, the NSPs were correct and made identical tool calls on every run, using about 70 times fewer input tokens and a third of the runtime of a skill-guided agent whose runs were inconsistent. In a composite version of the tasks, the skill-guided agent's accuracy fell to 23%, and its output inconsistency increased, while the NSPs remained correct and identical across all runs. The NSPs successfully generalized to all 108 unseen instances of the same templates. Our findings suggest that NSPs from agent trajectories can be considered cumulative software assets for end-users to test, maintain, and reuse in their recurring workflows.
Many supervised dimensionality reduction methods promote intra-class compactness and inter-class separation but do not explicitly constrain class geometry, leading to variability in the relative arrangement of classes across independent runs. Such variability limits the reproducibility and interpretability of low-dimensional embeddings. We propose the Isometric Centroid-Encoder (ICE), a supervised dimensionality reduction framework in which class centroids are learned directly in the latent space while preserving a prescribed pairwise class-centroid geometry. Unlike existing centroid- or prototype-based methods that primarily encourage intra-class compactness, ICE explicitly regularizes the relative arrangement of latent class centroids. Unlike multidimensional scaling and many metric-learning approaches that preserve sample-level relationships, ICE operates on class centroids, preserving inter-class geometry without requiring large numbers of sample-pair or triplet constraints. We further show that the intrinsic dimensionality required to represent the centroid-distance configuration of M classes is at most M-1, explaining the observed saturation of distance-preservation measures as the latent dimensionality increases. To quantify embedding reproducibility, we introduce the Centroid Neighborhood Stability (CNS) metric, which measures the consistency of class-level neighborhood topology across independent runs and complements existing embedding quality measures. Experiments on image and biological datasets demonstrate that ICE produces stable embeddings with improved preservation of class-centroid geometry while remaining computationally efficient. Additional analyses of optimization dynamics, sensitivity, and preprocessing further provide practical insights into latent dimensionality and explicit geometry modeling for supervised dimensionality reduction.
Repeated attempts at interactive tasks uncover facts, useful operations, and failures that can inform subsequent execution. We introduce ReGE, a framework for turning this recorded experience into guidance for a fresh attempt. Its main implementation, ReGE-Syn, uses a single coach call to jointly synthesize bounded execution records and available artifacts into self-contained guidance, connecting task-specific findings with proposed actions and relevant pitfalls. A student starts from the task's initial environment state and adapts this guidance to current observations. Source generation, synthesis, and student execution use the same base model, without parameter updates or a stronger teacher. Across four benchmarks, ReGE improves accuracy by an average of 9.9 percentage points for GPT and 8.8 percentage points for Claude compared with unguided execution. On matched BrowseComp-Plus subsets, ReGE-Syn reduces guidance-construction tokens by 98.2\% on GPT and 95.5\% on Claude relative to adapted AggAgent, with observed success. Combined construction and execution tokens decrease by 15.7\% and 35.4\%, respectively, counting input plus output and excluding source generation. BrowseComp-Plus ablations further associate access to process information with fewer student steps and show that joint synthesis produces substantially shorter guidance than Direct history or Independent summaries, with observed success. These results demonstrate how agents can use their own recorded experience to improve subsequent task completion, and suggest a low-cost way to consolidate the execution records of parallel agent systems such as agent swarms.
Numina Fuse is an agent-native application for formalizing mathematics in Lean 4. Since Lean can compile while not representing its intended informal meaning, and agents now write Lean faster than people can read it, we built Fuse around human supervision of agent work. A user works with an AI orchestrator on a Lean project and supervises it through views of the blueprint, the dependency graph, and the Lean code, while the backend keeps the project's derived state and runs agents for hours without the user present. We describe the development of Fuse throughout 2026, as it moved from a prescribed pipeline of agents paced by human review to an orchestrator that dispatches specialists, provers, and parallel children on its own. We designed for this shift, keeping project state, tooling, and interface design separate from orchestration so that agents could take over more of the work as models improved. We report Fuse's use in a formalization of the three-dimensional Kakeya conjecture and a study of prover models. We close with lessons for the next generation of autoformalization systems. Fuse Desktop, an open-source local version, is available at https://github.com/project-numina/fuse-desktop.
From Apple's repeatedly delayed Siri AI to OpenAI's messy "super app" launch, here's a look at the AI projects that shut down or missed expectations.
Software engineering (SE) expertise can shape how coding large language models (LLMs) are trained and which engineering capabilities they acquire. We examine this role through an industrial case study of capability-focused training for a 92B-parameter coding LLM serving more than 100,000 developers. We demonstrate how established SE knowledge can guide the selection of engineering capabilities, the construction of verifiable training data, and the evaluation of model improvements. To target repository-level refactoring, we developed RefactorForge, a pipeline that combines refactoring knowledge, established SE tools, and verification techniques to construct training data. Across 306 Python repositories, RefactorForge generated 5,704 executable tasks and retained 4,572 verified teacher trajectories totaling 72.51 million tokens. Incremental supervised fine-tuning on these trajectories, accounting for 0.46% of the preceding coding mixture's tokens, improved RefactorBench performance by 18.00 percentage points and SWE-Bench ProMax performance by 7.06 points. Under OpenCode, these improvements extended to SWE issue resolution, with gains of 4.97 points on SWE-Bench Verified and 1.23 points on SWE-Bench Pro. Improvements also persisted across other coding-agent harnesses, while greenfield repository generation showed mixed results. We report lessons from constructing capability-focused training data and evaluating its effects across engineering workloads. Building on this industrial experience, we propose a research agenda for expanding SE's role in transforming coding LLMs into SE LLMs: incorporating SE knowledge into model training, applying engineering methods to model development and maintenance, and systematically defining the engineering capabilities these models should acquire.
Query-time filters that screen each retrieved document before it reaches the generator are the dominant published defense against retrieval-augmented generation poisoning, and they report near-perfect detection on domain benchmarks. We show that this performance is predicted by how the poison set was built rather than by what the filter reads: removing a single construction cue destroys the deployed decision before it destroys the ranking. Two label-free statistics of a poison set and its clean corpus profile the cue: containment, how often the target question appears verbatim in the poison, and length overlap. PoisonedRAG's released code prepends the target question to every inserted passage. Of the 21 published defense evaluations we survey, 8 verifiably run poison built this way and 5 substitute a length cue; separately, 12 report no false-positive rate at the operating point they deploy. Deleting only that prefix from the attack authors' own passages (0.93 still retrieved for their target) takes RAGuard, a published content filter we re-implement, from AUROC 0.953 to 0.758 and its miss rate from 0.191 to 0.763. Across four corpora, four attack families and ten per-document defenses on one shared candidate set, the profile predicts where detection saturates. Padding medical poison to clean length with word-shuffled filler that carries no content takes TrustRAG's set-level stage from a miss rate of 0.000 to 0.827, and one learned scorer of ours from 0.000 to 0.986 (0.321 at a matched 5% false-positive rate) while its AUROC holds at 0.953. The ranking survives; the decision does not. A defense-aware attacker defeats our own scorer, so we claim no defense. A detection result is evidence only when it is reported with its poison set's construction profile and the false-positive rate at the threshold actually deployed; we release the script that computes the profile.
Early stopping is commonly used to prevent fitting label noise from degrading generalization. Does this mean that, when a positive fraction of the training labels is intrinsically unpredictable, the best test error must occur before interpolation? Experiments on a CIFAR10 classification task with randomly labeled examples motivate the opposite possibility: although fixed-feature and shallow models exhibit late-time deterioration of the test error, the need for early stopping essentially disappears in larger, deeper convolutional networks even after the random labels have all been classified correctly. We determine when late-time deterioration disappears entirely in a solvable setting. For kernel gradient flow on a synthetic noisy classification task, we derive sharp high-dimensional learning curves and obtain a phase diagram controlled by the balance between the linear kernel component that learns the shared relation between inputs and labels and the higher-order components that enable fitting the label noise. In one regime, generalization is optimized at finite time and early stopping is necessary. In another, generalization improves until interpolation, but the endpoint is suboptimal across kernels. At the boundary, the interpolating endpoint attains the lowest generalization error across the considered kernel family and all training times. Thus, even with extensive irreducible label noise, interpolation can outperform early stopping. The neural-network experiments leave open how learned representations acquire an analogous favorable geometry.
Can LLM agents carry a hard real-world problem from raw data to a system that holds its own against people? We study this in Hearthstone Battlegrounds, an eight-player auto-battler with stochastic shops and battles, hidden information and nearly a thousand interacting cards. We present AutoTavern, a self-improving research system: specialised LLM research agents share a typed research memory, every claim passes verification gates, and live games close the loop, so the agents improve the simulator, opponents and policy and, from the same feedback, their own research process; a human sets goals and rules on ambiguous mechanics. In 20 days AutoTavern built TavernGym, the first faithful, full-content reproduction of the live game, and delivered 23 versions of TavernAgent, trained with PPO against opponents calibrated to real games. TavernAgent plays complete games on the live ladder by itself: its rating rose from 6,000 to about 15,000 and, on the public leaderboard's measure, its last 20 records average a placement of 1.60 with 80% first places, better than the server's ten highest-rated players on every statistic (2.34, 59%). TavernBench, 960 tasks in 40 scenarios, shows that the system the agents built is the strongest player: TavernAgent answers 91% of the play tasks against at most 63% for frontier LLMs, and decides in 0.03 s where they need up to 26 s; within the second a live turn allows, it keeps 90% and no LLM more than 26%. LLM agents build far better than they play. We will open-source TavernBench and TavernGym at https://github.com/MCG-NJU/AutoTavern.
On Thursday, Pennsylvania Governor Josh Shapiro signed the Artificial Intelligence Deceptive Advertising Act into law, requiring disclosures on AI-generated ads. The law requires any ad…
Bayesian mixtures, and Solomonoff induction in particular, provide an idealized model of prediction learned by minimizing population log loss, and they underlie universal agents such as AIXI. An agent that plans with an approximation of such a predictor decides where that predictor is evaluated, and a planner can be drawn toward branches on which its model is optimistic. We study which approximation guarantees survive this feedback. For normalized finite-horizon causal mixtures, the transfer coefficient factors into a policy likelihood ratio and task evidence available before each prediction, so task rarity and action-branch rarity can compound. Within one fixed universal mixture, infinitely many computable environments admit excess prediction error of order κη\,2^{K(μ)}, where η is the training KL budget and κ measures policy coverage. This order is sharp, and for each target a computable predictor attains it at a suitable finite horizon, so computability alone does not remove the penalty. For bounded returns, value error is controlled by the probability of entering a poorly covered branch plus a square-root term in η. Both terms are necessary, and finite maximum coverage ratios are not required. Motivated by the mentor override of Golden Handcuffs, we derive margins for preserving deferral decisions and conditional KL budgets with an unavoidable inverse-history-probability penalty, and a counterexample separates agreement of deferral decisions from task-wise event avoidance. The guarantees apply to any predictor architecture that meets explicit population approximation and coverage conditions, which gives concrete targets for implementing idealized universal agents.
Block said that the voice model can run locally on a MacBook M4 Max
