data curation teams

Abstract: When told that robins have sesamoid bones, people are far more prone to extend the property to sparrows than to whales. This behavior, called category-based property induction, is well documented in humans, but whether and how it develops in language models trained only to predict the next token is unclear. To shed light on this, we construct a hierarchical taxonomy of synthetic entities grouped by the properties they share. We then inject facts from this latent taxonomy into the continued pre-training of Pythia and OLMo 2 models from 70M to 7B parameters, and track the development of inductive behaviors. We find that taxonomically aligned induction emerges from next-token prediction alone: a property learned for one entity transfers more strongly within its taxonomic category than across categories. Internally, entities that share a deeper category become more similar in activation space, and clustering recovers the full taxonomic tree from entity representations. Contrary to the usual expectation that scale helps, we find that induction weakens with model size when the data budget is fixed, suggesting that larger models memorize individual facts rather than compress them into shared structure. Induction also strengthens with more pre-training before the taxonomy injection, transfers to novel in-context properties, and can be overridden by a latent taxonomy implied by the prompt. Together, these results show that a model's inductive generalization reflects the latent organization of its training data, with implications for data curation and for predicting how far a learned fact will spread.

Abstract: Image-text alignment is a core problem in computer vision with applications in caption evaluation, hallucination detection, data curation, and the benchmarking of text-to-image (T2I) generators. As T2I models improve, benchmarking has become demanding, requiring metrics capable of finding a series of issues like missing objects, swapped attributes, miscounts, and ignored negations. Recent work addresses this by fine-tuning evaluators on preference data or by prompting a vision-language model, either holistically with the caption or with decomposed verification questions. However, existing approaches fall short: fine-tuned metrics remain bound to one backbone and training distribution; holistic metrics miss fine-grained details; and decomposed metrics rely on a fixed-YES assumption that penalizes faithful images whenever that assumption fails. In contrast, we propose DEPICT, a training-free metric that replaces fixed reference answers with expected agreement between image-based and caption-only answers, weighting questions by how decisively the caption determines them. By replacing fixed references, our agreement rule increases negation accuracy from 19% to 88%. To recover the context lost during decomposition, DEPICT merges this agreement score with a holistic score. We evaluate DEPICT on five benchmarks and eleven backbones from three model families and find that it surpasses all training-free metrics and exceeds fine-tuned evaluators on two out of three human-correlation benchmarks.

Abstract: Recaptioned image-text corpora are now standard for text-to-image (T2I) training, with vision--language model (VLM) captioners replacing sparse alt-text by dense descriptions. A recaptioned corpus is a supervision distribution induced by a documented captioning policy ($π$), captioner ($V_c$), and source corpus ($C$). Length-correlated proxies miss caption-register artifacts and downstream T2I benchmarks entangle the corpus with training choices, so this distribution is hard to audit at corpus scale. We introduce a reusable matched-budget audit framework for recaptioned supervision distributions $D_{π,V_c,C}$: at a fixed text budget of $B = 64$ it reports a five-axis profile spanning prompt-side coverage, image-conditioned faithfulness, and caption-surface health, with claimed controllable basic units (CBU) as the common claim unit. We instantiate the framework on seven paired comparisons over five public source corpora. Across the four cross-corpus pairs, the released surface raises supported CBU per caption by $+3.39$ to $+6.36$ under both Qwen and Gemma Judges, and on CC12M the same framework exposes a long-vs-dense frontier that is consistent across both judges and four budgets. We release the audited multi-source recap corpus ($\approx$ 490M) together with the audit-artifact bundle.

Abstract: People typically assume that LMs stably mature from pattern-matching parrots to generalizable intelligence during pre-training. We build a toy eval suite and show this mental model is wrong: throughout pre-training, LMs frequently and suddenly hop between parrot-like and intelligence-like computations. We call this mode-hopping. Across our suite, LMs suddenly latch onto memorized or in-context patterns instead of in-context learning, use System 1 instead of System 2 thinking, pick up what sounds true instead of what is true, fail at multi-hop persona QA, out-of-context reasoning, and emergent misalignment -- then just as suddenly revert and generalize. Mode-hopping is not explained by standard optimization dynamics: it is locally stable and cannot be fixed by checkpoint averaging. We instead think of it as a capacity allocation problem: in a capacity-bounded model, generalizable circuits must compete with the shallow ones learned early in training, and the data in each pre-training window may decide which circuits win. Our suite provides a new efficient lens on generalization. We demonstrate two concrete applications: (i) select intermediate pre-training checkpoints that strongly generalize reasoning and alignment, better than the final pre- or mid-training checkpoints, and (ii) select pre-training data that controls and stabilizes generalization dynamics.

Abstract: Although bilevel optimization (BLO) has emerged as a powerful framework for addressing many complex and nested machine learning problems in recent years, most existing studies are confined to the lower-level strongly convex (LLSC) or lower-level generally convex (LLGC) settings (i.e., the lower-level objective function is assumed to be, at least, convex). While the LLSC/LLGC assumptions render more tractable algorithmic design and theoretical analysis, they are too rigid to encompass many machine learning problems in practice. The limitations of LLSC/LLGC assumptions in BLO motivate us to investigate solving the BLO problem in the general lower-level nonconvex (LLNC) settings, which remains in its infancy. In the literature on LLNC-BLO, most of the existing works either require additional structures in the lower-level objective function for tractable theoretical analysis, or adopt the first-order stationarity reformulation as a lower-level surrogate problem, which is inherited from the LLSC/LLGC settings but could lose their effectiveness in the LLNC setting. To bridge this gap, we propose to reformulate the nonconvex lower-level problem using a second-order stationarity-based surrogate, the solution of which guarantees a local optimal solution at the lower level. Based on this reformulation, we propose the PROBE (Perturbed gradient algorithm for bilevel problem) and show that it overcomes the limitations of prior works by probing and escaping lower-level saddle points. We prove that PROBE achieves a finite-time convergence rate of $O(T^{-2/5})$, where T denotes iterations. To our knowledge, this work is the first to establish the finite-time convergence for achieving lower-level second-order stationary solutions in general LLNC-BLO. Our experiments on both a large language model-based data curation task and a meta-learning task also show that PROBE outperforms SOTA methods.

Abstract: Large language models trained recursively on their own or other models' outputs undergo model collapse, in which distributional tails and factual accuracy deteriorate while fluency survives. Prior work diagnoses collapse after training; the actionable problem is screening a corpus of unknown provenance before training. We introduce SynthSentry, a corpus-level, model-agnostic contamination signal requiring no access to the generating model, no generation history, and no synthetic labels. The score is a distributional divergence over three statistics: lexical diversity collapse, n-gram tail truncation, and perplexity variance across reference models. We evaluate on corpora contaminated by small open-weight generators and an instruction-tuned open-weight model under a leave-one-generator-out protocol. A domain-stratified study measures false positives on naturally repetitive human text (legal, clinical, source code). The score ranks corpora by severity with little loss when whole generator families are held out. Per-domain calibration holds near its nominal false-positive budget once covariance shrinkage and a bootstrap threshold replace a naive quantile, which runs four times over budget. A downstream fine-tuning check showed no contamination-driven accuracy deficit at our scale, so whether pruning recovers one remains open; the same run shows over-pruning risk once pruning exceeds the true contamination fraction. We frame screening as a data-curation defense rather than a post-hoc diagnosis and release the scoring toolkit. All results are small-scale; scope is English-language, batch-mode corpus screening. Contamination sources are single-generation or hand-authored rather than recursively generated, so results speak to synthetic contamination generally and not to recursion depth.