software developers building ai tools

Abstract: Generative LLMs can label spans without task-specific training, but their output is not tied to positions in the input: a returned span may be absent from the input or occur several times. We present constrained tagging, a training-free method in which the model reproduces the input and wraps each span in labeled inline markup. A logits processor using a byte-level token trie and a finite-state machine admits only tokens that continue the input verbatim or form a valid tag, so the output reproduces the input exactly with labels from the given set, and spans may begin or end at any character boundary, even inside a token. Two-phase decoding lets reasoning models reason before the constraint engages. On six datasets and six open-weight models, the guarantee holds in all but 0.022% of 255,312 constrained generations, each failure a reasoning model exhausting its token budget or never closing its reasoning, and the constraint adds no measurable cost. Unlike context-based prompting, which loses 8.5% of its spans, constrained tagging locates every span it proposes on every dataset; it is as accurate on two datasets, and on the other four its errors lie in which spans the model chooses, not where they are.

Abstract: As claims about recursive self-improvement (RSI) and artificial general intelligence (AGI) proliferate, we ask a simple question: do frontier AI models have a sense of data, i.e., can they reliably select the right data for training? We introduce DataSense-Bench to study this capability through the fundamental problem of data selection and performance forecasting in machine learning. We ask AI agents to select and rank candidate training subsets that can be used to fine-tune a small LLM model. Agents are allowed to inspect the data, write and execute analysis code, and run model forward passes, but can not train the model or access the actual evaluation tasks. We then fine-tune the base model on each selected subset and evaluate its post-training performance under a standardized protocol. We instantiate the benchmark in terminal problem solving and tool use, selecting trajectories from OpenThoughts-Agent and EnvScaler and evaluating on TBLite and BFCL, respectively. We then evaluate the agents along two complementary dimensions: the post-training performance of the top-ranked subset, reflecting the ability to identify high-value training data, and ranking accuracy, reflecting the ability to predict the relative performance of the selected subsets. In our experiments, selection gains over random selection are limited; agents do not reliably rank their selected groups, and ranking ability does not hold consistently across tasks: Astra identifies the best group in all three tool-use runs but in only one of three terminal runs. Analysis of execution traces on both tasks shows that agents often use similar data signals while interpreting their training value differently.

Abstract: We study whether latent sequence compression of prompt prefixes preserves the capabilities that large language models rely on during inference. We introduce SemanticFold, a compression scheme that folds prefix hidden states at learned boundaries, and evaluate it across five model scales: Qwen3-1.7B, Qwen3-8B, SmolLM2-1.7B, Pythia-1.4B, and Pythia-6.9B. We use a fixed-target protocol: a frozen prefix is executed natively or compressed, and both arms teacher-force identical continuation tokens. This design rules out target-selection explanations for likelihood changes. We examine five endpoint families: fixed-target negative log-likelihood, finite-label reasoning accuracy, linear probe accessibility, open-ended generation, and systems-level memory and latency. We find that compression moves these endpoints non-monotonically and that they do not share a single compression threshold. On Qwen3-1.7B at compression ratio R=1.7, compressed-minus-native mean NLL decreases by 0.135 under paired bootstrap with 10000 draws. On SmolLM2 at R=1.2, the mean change is 0.013 higher than native. On both Pythia checkpoints, NLL is effectively unchanged. An NLL decomposition separating sequence shortening from the learned residual transform shows that the favorable Qwen likelihood is attributable primarily to residual adaptation rather than to shortening alone. MLP-only, which applies the transform without shortening, achieves 0.082 lower NLL than Full SemanticFold. Linear probe accuracy and macro AUC change by less than 0.03 in absolute value across conditions, with confidence intervals crossing zero. We conclude that preservation under latent compression has no single scalar certificate: language-model fit, decodability, and reasoning behavior answer different questions and can move in different directions under the same compression operation.

Abstract: A key strength of Proximal Policy Optimization (PPO) is its learned critic, which uses historical trajectories collected during reinforcement learning to estimate expected returns and reduce policy-gradient variance. However, we find that the critic is also a major source of instability in reinforcement learning for large language models (LLMs). We identify two critic failure modes that destabilize PPO. First, filtering truncated rollouts from both actor and critic shifts the policy objective to reward conditioned on completion, allowing truncation to increase even as conditional reward improves. Second, heterogeneous return noise can cause high-variance prompts to dominate critic updates in finite batches. We introduce EasyPPO to address these failures. Actor-only overlong filtering trains the critic on returns from both completed and truncated rollouts. Noise-normalized critic regression weights each prompt's critic loss by the inverse standard deviation of its sampled returns, balancing noise contributions across prompts. Moderately smaller critic mini-batches confine outlier influence to fewer rollouts during gradient clipping. Across continuous-reward coding on FrontierCS, binary-reward mathematical reasoning on AIME24, and multi-turn search on Search-R1, EasyPPO remains stable throughout the full training horizon and consistently outperforms vanilla PPO, VAPO, and HL-Gauss PPO. Its best validation scores show relative gains of 14.89%, 2.28%, and 9.47% over PPO, respectively.

Abstract: Transformers and state-space models (SSMs) are two prominent sequential learning architectures, yet their comparison remains largely empirical and existing theoretical analyses are typically task-specific or architecturally restricted. In this paper, we develop belief geometry, a unified analytical framework for comparing the representational capabilities of broad classes of attention and SSMs. Starting from a generalized formulation of in-context linear regression and using cumulative Bayes regret as our measure, we abstract three capabilities required by many sequential learning problems in our belief geometry: evidence assembly, belief maintenance, and addressing. We then study three cases of our formulation that isolate these capabilities and yield sharp architectural lessons: For belief maintenance, SSMs attain the optimal regret over stationary aggregation kernels; for positional assembly, SSMs have a memory advantage; and for content addressing, softmax attention has an exponential width advantage over sigmoid-selective SSMs. Experiments with LLaMA-type Transformers and Mamba-2 show that these architectural insights extend beyond our analytically tractable classes and linear-regression testbed.

Abstract: Transformers achieve remarkable performance by jointly learning broad families of tasks during pretraining and adapting to unseen tasks from only a short prompt. Yet a rigorous mathematical and statistical understanding of this phenomenon remains limited. This paper aims to study how Transformers exploit shared cross-task structure and how this structure affects the sample complexity of in-context learning (ICL). Specifically, we characterize task-space complexity through covering numbers under a prescribed metric, thereby quantifying the low-dimensional cross-task structure without requiring an explicit parametric representation. The resulting cover provides a set of anchor functions, which we use to introduce a task-identification-and-evaluation procedure: context observations localize an unseen task among the anchor functions, and the response at a query is predicted by aggregating the corresponding anchor function query evaluations. For approximation, we explicitly construct a Transformer with Softmax attention to approximate this procedure. For generalization, we derive an error bound that separates the effects of the number of pretraining tasks and the prompt length. The scaling with respect to the number of pretraining tasks is governed by the intrinsic dimensions of the task space and input domain; once sufficiently many tasks are available, the dependence on the prompt context length becomes dimension-free. To the best of our knowledge, this is the first work to quantify cross-task complexity for general nonlinear task families and explicitly construct a Transformer that exploits their low-dimensional structure to perform ICL. Our theory provides a quantitative explanation of how joint pretraining across related tasks improves in-context generalization.

Abstract: Automated red teaming often replays a fixed set of prompts, which measures known risks but cannot learn from failures found during testing. We present CART (Closed-Loop Adaptive Red Teaming), a framework that uses each result to guide what it tests next. CART begins with broad risk coverage, follows weaknesses that emerge, keeps new probes diverse, and records the evidence and source of every finding. It separates the Challenger that creates tests, the Target being tested, which may be a text-only model or a bounded tool-using agent, and the Judge that evaluates the results, allowing these roles to be studied independently. Across three evaluation families (Frontier, JAH, and Agentic), CART discovers more failures and higher average risk than static seed replay for every Target with an available baseline. The gains extend to tool-mediated agent tests, suggesting that contextual adaptation can reveal weaknesses that direct prompt replay does not exercise. These results describe what the test policies discover, not how often failures occur in real deployments. We also find that Challenger-Judge choices affect the evidence uncovered, highlighting the need for role separation and independent review. Overall, CART turns red teaming from a one-time checklist into a continuous, adaptive, and auditable search for model and agent weaknesses.

Abstract: We present the first method for native, continuous gradient descent for machine learning models with $p$-adic parameters. Existing native optimizers are discrete, mostly combinatorial searches, as the $p$-adic numbers $\mathbb{Q}_p$ are totally disconnected, with standard losses that are flat away from their minima. To enable continuous optimization, we propose working with $\mathbb{Q}_p$ via its Berkovich affine line: a canonical, path-connected expansion of $\mathbb{Q}_p$ that preserves its isometries and uniquely extends its analytic maps. This hull is a metric tree with interpretable points and local derivatives, which we show enables effective optimizers and backpropagation. We formulate gradient descent and show that its approximations efficiently learn linear models with coefficients in $\mathbb{Q}_p$ to do modular arithmetic, an XOR-like task not expressible by linear models in $\mathbb{R}$. We also demonstrate momentum and Adam variants, linear regression, and classification on binary-encoded hierarchies (Quillian semantic networks), addressing open problems posed by Martins (2025). Library at https://github.com/google-deepmind/padic-ml

Abstract: Compound LLM systems often solve a coordination problem by adding a higher-level LLM. The resulting meta-agent reads workers' outputs, writes the final answer, allocates later calls, and decides when to stop. It is expressive, but it also concentrates three control decisions in an opaque, order-sensitive model call. We ask whether the manager needs to be generative at all. UnitBoost replaces that model with a defined meta-level operator: a task-given unit map turns worker outputs into slot-value proposals, a constrained argmax assembles the output, and the slots left unfilled or unsupported become an explicit residual for the next round. The operator is order-free, records unit provenance, and gives a simple guarantee: without coupling constraints, unit-wise maximization under the same admission score dominates selection of any complete candidate. On three held-out benchmarks, it exceeds the best single candidate chosen with gold labels by 0.060-0.195 absolute task-score points and input-matched generative managers by 0.048-0.076. Replacing only the management step improves six compound-system configurations by 0.013-0.182. Residual-directed rounds raise FanOutQA cell F1 from 0.4778 to 0.5524; matched controls show that the true residual outperforms random targets and ordinary rereading, while a label-free supply signal flags exhaustion after one unproductive round. The same analysis measures three conditions in which no such gain is available (one indivisible unit, unavailable unit identity, and an endpoint that charges for every emitted unit) and quantifies cross-unit coupling as a repair cost. The manager gives up semantic freedom and gains order invariance, unit provenance, and testable failure conditions.