Learning Distributions from Multiple Data Providers

Motivated by learning from heterogeneous and overlapping data providers, we study a stylized model of distribution learning from restricted conditional samples. The goal is to learn an unknown distribution $p$ on a finite domain $[n]$. The learner is given a fixed family of queryable sets $\mathscr{S} \subseteq 2^{[n]}$, and each query to $S \in \mathscr{S}$ returns an independent sample from the conditional distribution $p(\cdot \mid S)$. Learnability is governed by the co-occurrence graph associated with $\mathscr{S}$: two domain elements are adjacent if they appear together in some queryable set. Pointwise consistency is achievable when this graph is connected on the target support. PAC learning requires more: it is possible when the co-occurrence graph is complete. The optimal sample complexity of PAC learning ranges from nearly linear to quadratic. Every query family with complete co-occurrence graph admits sample complexity $\widetilde O(n^2/ε^2)$, and this bound is tight in the worst case. On the other hand, if $[n]$ is queryable then ordinary sampling improves the bound to $Θ(n/ε^2)$, and this cannot be improved further even if every set is queryable. More generally, we identify hierarchical comparabilityas a sufficient structural condition on $\mathscr S$ under which the optimal complexity is nearly linear, $\widetilde Θ(n/ε^2)$, with pairwise query families as a canonical example. Finally, the full range of polynomial rates between linear and quadratic is attainable: for every $α\in (1,2)$, there exists a query family with optimal PAC rate $\widetilde Θ(n^α/ε^2)$.

LLM Detection as an Intervention: Downstream Impact under Strategic User Behavior

As LLM adoption becomes more widespread, there is a growing interest in detecting LLM-generated content, for example through LLM detection tools and through heuristics based on language patterns. Detectors operate as an intervention that steers not only the detected attribute itself, but also downstream metrics such as LLM usage and output quality. In this work, we demonstrate how imperfect LLM detectors lead to counterintuitive impacts on these downstream metrics, by distorting how users are incentivized to use LLMs in their workflow. We develop a stylized model which captures how users strategically choose how much to use the LLM and how to post-process content to reduce the detected attribute. Using this model, we show that LLM detection can counterintuitively lead humans to increase their LLM usage. Moreover, even when reducing the detected attribute improves output quality, we find that introducing an LLM detector can lead users to produce lower quality outputs. In contrast, we show that detectors result in a clean "rise-then-fall" pattern for the detected attribute, which we empirically reproduce for word frequencies on arXiv abstracts. Altogether, our work illustrates how LLM detection can distort LLM usage and output quality, uncovering failure modes when LLM detectors operate as an intervention on these downstream metrics.

Language Identification with Succinct Machine-Independent Traces

Motivated by the power of large language models, there has been renewed interest in the Gold-Angluin model of language identification in the limit, with an eye toward variants of the model that might overcome the negative results for its original formulation. Recent papers on this question have proposed looking at computational traces and annotations of training strings as a source of additional power for a learner, reflecting empirical regularities such as the way that commented source code is easier to learn from than arbitrary source code, and text annotated with algorithmically generated chain-of-thought tokens can be easier to learn from than the raw text itself. This recent work has shown positive results for language identification in the presence of such computational traces, but the traces in these positive results come from explicit automata-theoretic machine models that generate the language, where the underlying vocabulary of tokens for the traces is very large. In this paper, we address two fundamental issues left open by this line of work: can we achieve positive results with traces that use only a small alphabet, and can we define traces directly from the language itself, without requiring an underlying machine model that generates it? We establish positive results for both of these questions: for an arbitrary collection of languages, we show how to define computational traces that enable identification in the limit, using an alphabet of tokens that is linear in the size of the alphabet that the languages are defined over, and independent of any other properties of the languages.

Globally Consistent Coloring Schemes for Language Identification

We study how little extra information is needed to make adversarial language learning possible. In Gold's model of language identification in the limit, a learner is given an enumeration of the strings from an unknown language chosen from a countable language collection. The learner guesses the identity of the language over the course of the enumeration, and it succeeds if, eventually, all of its guesses are the correct language. Classical results of Gold and Angluin show that many natural collections cannot be learned in this way. Recent work on trace colorings, motivated by the success of thinking-trace strategies in language learning, overcomes this obstruction by annotating every symbol of every string with a color. We ask whether the learner really needs this whole sequence of colors, or whether one color at the end of each string (a terminal coloring) is enough for language identification. We show that just one terminal bit per string is enough for every countable collection of infinite languages. In fact, the colorings can be chosen collection-independently: there is a single assignment of a two-color terminal coloring to every infinite language such that the same preassigned colorings identify every countable subcollection. Thus, in this model, an entire color trace can be compressed to one bit attached to the end of each example. Our global construction uses transfinite recursion, and we prove that this kind of nonconstructivity is unavoidable for any bounded number of colors. As a notion of constructivity, we use the formalism of Borel maps (a regularity condition satisfied by natural explicit constructions); we show that no global terminal coloring with a finite number of colors defined by a Borel map can identify all countable subcollections. By contrast, known trace-coloring constructions are Borel when encoded as terminal colorings, but require infinitely many colors.

On Language Generation in the Limit with Bounded Memory

We study language generation in the limit under bounded memory. In this task, a learner observes examples from an unknown target language one at a time and must eventually output only new valid examples. Prior work assumes access to the entire history, a strong assumption since realistic algorithms retain limited past information. Classical work in learning theory shows memory constraints dramatically alter learnability; we extend this to language generation. First, we study memoryless generators. Under a mild enumeration restriction, every countable collection of infinite languages remains generable without memory. Without this restriction, we exactly characterize when memoryless generation is possible. For finite collections, we characterize the optimal minimax density achievable by memoryless generators -- the best density guaranteed against any collection of a given size. This combinatorial bound relies on Sperner's theorem and symmetric chain decompositions. We further show that a sliding window of the last $W$ examples does not improve this worst-case density, whereas allowing it to store $b$ adaptively chosen past examples improves the achievable density for every $b \geq 1$. Finally, we revisit identification in the limit, where the learner must converge to a single correct hypothesis for the target language. We focus on its incremental variant, where the learner remembers only its previous guess. Here, although exact identification fails on a collection of just three languages, a mild relaxation requiring convergence to an ``approximate'' version of the target is achievable for every finite collection. These results show bounded memory affects these tasks differently: generation remains achievable for every countable collection, while density and identification are confined to finite collections, with guarantees weakening as the collection grows.

Mistake-Bounded Language Generation

We investigate the learning task of language generation in the limit, but shift focus from the traditional time-of-last-mistake metric of a generator's success to a new notion of "mistake-bounded generation." While existing results for language generation in the limit focus on guaranteeing eventual consistency, they are blind to the cumulative error incurred during the learning process. We address this by shifting the goal to minimizing the total number of invalid elements output by a generation algorithm. We establish a formal reduction to the Learning from Correct Demonstrations framework of Joshi et al. (2025), enabling a general recipe for deriving mistake bounds via weighted update rules. For finite classes, we provide an algorithm that simultaneously achieves an optimal last-mistake time of $\mathsf{Cdim}(L)$ and a mistake bound of $\lfloor \log_2 |L| \rfloor$, whereas for the non-uniform setting of countably infinite streams of languages, we prove a fundamental trade-off: achieving logarithmic mistakes $O(\log i)$ necessarily precludes convergence guarantees established in prior work. Finally, we show that our framework can be extended to accommodate noisy adversaries and guarantee mistake bounds that scale with the adversary's suboptimality.

How Many Features Can a Language Model Store Under the Linear Representation Hypothesis?

We introduce a mathematical framework for the linear representation hypothesis (LRH), which asserts that intermediate layers of language models store features linearly. We separate the hypothesis into two claims: linear representation (features are linearly embedded in neuron activations) and linear accessibility (features can be linearly decoded). We then ask: How many neurons $d$ suffice to both linearly represent and linearly access $m$ features? Classical results in compressed sensing imply that for $k$-sparse inputs, $d = O(k\log (m/k))$ suffices if we allow non-linear decoding algorithms (Candes and Tao, 2006; Candes et al., 2006; Donoho, 2006). However, the additional requirement of linear decoding takes the problem out of the classical compressed sensing, into linear compressed sensing. Our main theoretical result establishes nearly-matching upper and lower bounds for linear compressed sensing. We prove that $d = Ω_ε(\frac{k^2}{\log k}\log (m/k))$ is required while $d = O_ε(k^2\log m)$ suffices. The lower bound establishes a quantitative gap between classical and linear compressed setting, illustrating how linear accessibility is a meaningfully stronger hypothesis than linear representation alone. The upper bound confirms that neurons can store an exponential number of features under the LRH, giving theoretical evidence for the "superposition hypothesis" (Elhage et al., 2022). The upper bound proof uses standard random constructions of matrices with approximately orthogonal columns. The lower bound proof uses rank bounds for near-identity matrices (Alon, 2003) together with Turán's theorem (bounding the number of edges in clique-free graphs). We also show how our results do and do not constrain the geometry of feature representations and extend our results to allow decoders with an activation function and bias.

Language Generation and Identification From Partial Enumeration: Tight Density Bounds and Topological Characterizations

The success of large language models (LLMs) has motivated formal theories of language generation and learning. We study the framework of \emph{language generation in the limit}, where an adversary enumerates strings from an unknown language $K$ drawn from a countable class, and an algorithm must generate unseen strings from $K$. Prior work showed that generation is always possible, and that some algorithms achieve positive lower density, revealing a \emph{validity--breadth} trade-off between correctness and coverage. We resolve a main open question in this line, proving a tight bound of $1/2$ on the best achievable lower density. We then strengthen the model to allow \emph{partial enumeration}, where the adversary reveals only an infinite subset $C \subseteq K$. We show that generation in the limit remains achievable, and if $C$ has lower density $α$ in $K$, the algorithm's output achieves density at least $α/2$, matching the upper bound. This generalizes the $1/2$ bound to the partial-information setting, where the generator must recover within a factor $1/2$ of the revealed subset's density. We further revisit the classical Gold--Angluin model of \emph{language identification} under partial enumeration. We characterize when identification in the limit is possible -- when hypotheses $M_t$ eventually satisfy $C \subseteq M \subseteq K$ -- and in the process give a new topological formulation of Angluin's characterization, showing that her condition is precisely equivalent to an appropriate topological space having the $T_D$ separation property.

Density Measures for Language Generation

The recent successes of large language models (LLMs) have led to a surge of theoretical research into language generation. A recent line of work proposes an abstract view, called language generation in the limit, where generation is seen as a game between an adversary and an algorithm: the adversary generates strings from an unknown language $K$, chosen from a countable collection of candidate languages, and after seeing a finite set of these strings, the algorithm must generate new strings from $K$ that it has not seen before. This formalism highlights a key tension: the trade-off between validity (the algorithm should only produce strings from the language) and breadth (it should be able to produce many strings from the language). This trade-off is central in applied language generation as well, where it appears as a balance between hallucination (generating invalid utterances) and mode collapse (generating only a restricted set of outputs). Despite its importance, this trade-off has been challenging to study quantitatively. We develop ways to quantify this trade-off by formalizing breadth using measures of density. Existing algorithms for language generation in the limit produce output sets that can have zero density in the true language, and this important failure of breadth might seem unavoidable. We show, however, that such a failure is not necessary: we provide an algorithm for language generation in the limit whose outputs have strictly positive density in $K$. We also study the internal representations built by these algorithms, specifically the sequence of hypothesized candidate languages they consider, and show that achieving the strongest form of breadth may require oscillating indefinitely between high- and low-density representations. Our analysis introduces a novel topology on language families, with notions of convergence and limit points playing a key role.

The Backfiring Effect of Weak AI Safety Regulation

Recent policy proposals aim to improve the safety of general-purpose AI, but there is little understanding of the efficacy of different regulatory approaches to AI safety. We present a strategic model that explores the interactions between the regulator, the general-purpose AI technology creators, and domain specialists--those who adapt the AI for specific applications. Our analysis examines how different regulatory measures, targeting different parts of the development chain, affect the outcome of the development process. In particular, we assume AI technology is described by two key attributes: safety and performance. The regulator first sets a minimum safety standard that applies to one or both players, with strict penalties for non-compliance. The general-purpose creator then develops the technology, establishing its initial safety and performance levels. Next, domain specialists refine the AI for their specific use cases, and the resulting revenue is distributed between the specialist and generalist through an ex-ante bargaining process. Our analysis of this game reveals two key insights: First, weak safety regulation imposed only on the domain specialists can backfire. While it might seem logical to regulate use cases (as opposed to the general-purpose technology), our analysis shows that weak regulations targeting domain specialists alone can unintentionally reduce safety. This effect persists across a wide range of settings. Second, in sharp contrast to the previous finding, we observe that stronger, well-placed regulation can in fact benefit all players subjected to it. When regulators impose appropriate safety standards on both AI creators and domain specialists, the regulation functions as a commitment mechanism, leading to safety and performance gains, surpassing what is achieved under no regulation or regulating one player only.