arXiv subjects

Loris D'Antoni

Publications and source records attributed to Loris D'Antoni.

At least 19 recordsLinked to original sources

arXiv2026

Learning the Error Patterns of Language Models

When generating outputs for domains with specific validity constraints (e.g., a program should compile), LLMs often fail in a small number of focused ways: for example, by using Python function names when generating TypeScript. We observe that these error patterns can be represented using a small number of constraints that can be learned in practice. We propose \emph{prefix filters}, which are per-domain-and-LLM symbolic functions, as objects to capture the error patterns, Palla as an algorithm to learn prefix filters efficiently in practice, and implement Palla. Prefix filters learned by Palla i) help us quantitatively analyze the error patterns of LLMs, and ii) can be used to constrain the outputs of a model via constrained sampling algorithms. For example, Palla boosts compile rates for Qwen2.5-1.5B on TypeScript generation, by over 60%, allowing Qwen2.5-1.5B to achieve similar performance to Llama3.1-8B unconstrained.

cs.LG↗

arXiv2026

Manifold-Guided Attention Steering

Large language models frequently produce errors in reasoning tasks despite possessing the underlying knowledge required for correct reasoning. One possible approach to improve reasoning consistency is through activation steering. However, existing activation steering approaches apply fixed, pre-computed correction vectors, ignoring where the model currently sits along its generation trajectory; the result is indiscriminate perturbation that disrupts already-correct steps as freely as erroneous ones. We propose Manifold-Guided Attention Steering (MAGS), a trajectory-aware inference-time intervention grounded in a geometric observation: the output activations of specific attention heads diverge from a low-dimensional correctness manifold at the point of error, and this deviation compounds through subsequent steps. For each identified attention head, we learn a low-dimensional subspace from contrastive pairs of correct and incorrect traces that capture the directions along which error behavior deviates from correct behavior. During inference, we monitor each head's proximity to this manifold and apply a targeted projection correction when deviation exceeds a learned threshold, steering the attention output back toward the correct subspace before the error propagates. MAGS consistently outperforms both unsteered baselines and static steering approaches across benchmarks spanning mathematical reasoning (MATH-500, GSM8K), code generation (HumanEval, MBPP), and molecular generation (SMILES), suggesting that correctness manifolds are a general feature of LLM attention geometry.

cs.LG↗

arXiv2026

Language-Based Agent Control

This paper introduces language-based agent control (LBAC), a new programming model for agentic applications that brings techniques from programming languages and language-based security to the problem of agent control. In conventional programming, combinations of static typing and runtime enforcement have long been used to guarantee that well-typed programs satisfy user-specified policies, including policies for access control, information flow, data provenance, and more. The key idea behind LBAC is to extend these guarantees to agentic applications by requiring agents to generate programs that are themselves well typed in the context of the surrounding scaffolding code. Unsafe programs are rejected by the type-checker before execution, allowing policies to apply uniformly across the entire application, including both agent-generated behavior and developer-written scaffolding. At the same time, LBAC preserves substantial expressiveness: agents may perform arbitrary side-effect-free computation and recursively invoke subagents, which retain full tool access subject to the same -- or potentially more restrictive -- policies. We demonstrate LBAC with three case studies: I/O sandboxing via filesystem capabilities, data provenance, and information-flow control.

cs.PL↗

arXiv2026

The Format Tax

Asking a large language model to respond in JSON should be a formatting choice, not a capability tax. Yet we find that structured output requirements -- JSON, XML, LaTeX, Markdown -- substantially degrade reasoning and writing performance across open-weight models. The research response has focused on constrained decoding, but sampling bias accounts for only a fraction of the degradation. The dominant cost enters at the prompt: format-requesting instructions alone cause most of the accuracy loss, before any decoder constraint is applied. This diagnosis points to a simple principle: decouple reasoning from formatting. Whether by generating freeform first and reformatting in a second pass, or by enabling extended thinking within a single generation, separating the two concerns substantially recovers lost accuracy. Across six open-weight models, four API models, four formats, and tasks spanning math, science, logic, and writing, decoupling recovers most lost accuracy. Notably, most recent closed-weight models show little to no format tax, suggesting the problem is not inherent to structured generation but a gap that current open-weight models have yet to close. Code is available at https://github.com/ivnle/the-format-tax.

cs.CL↗

arXiv2026

Continuous Diffusion Models Can Obey Formal Syntax

Diffusion language models offer a promising alternative to autoregressive models due to their global, non-causal generation process, but their continuous latent dynamics make discrete constraints -- e.g., the output should be a JSON file that matches a given schema -- difficult to impose. We introduce a training-free guidance method for steering continuous diffusion language models to satisfy formal syntactic constraints expressed using regular expressions. Our approach constructs an analytic score estimating the probability that a latent state decodes to a valid string accepted by a given regular expression, and uses its gradient to guide sampling, without training auxiliary classifiers. The denoising process targets the base model conditioned on syntactic validity. We implement our method in Diffinity on top of the PLAID diffusion model and evaluate it on 180 regular-expression constraints over JSON and natural-language benchmarks. Diffinity achieves 68-96\% constraint satisfaction while incurring only a small perplexity cost relative to unconstrained sampling, outperforming autoregressive constrained decoding in both constraint satisfaction and output quality. Diffinity is open-sourced at github.com/large-loris-models/Diffinity.

cs.LG↗

arXiv2025

Nice to Meet You: Synthesizing Practical MLIR Abstract Transformers

Static analyses play a fundamental role during compilation: they discover facts that are true in all executions of the code being compiled, and then these facts are used to justify optimizations and diagnostics. Each static analysis is based on a collection of abstract transformers that provide abstract semantics for the concrete instructions that make up a program. It can be challenging to implement abstract transformers that are sound, precise, and efficient, and in fact both LLVM and GCC have suffered from miscompilations caused by unsound abstract transformers. Moreover, even after more than 20 years of development, LLVM lacks abstract transformers for hundreds of instructions in its intermediate representation (IR). We developed NiceToMeetYou, a program synthesis framework for abstract transformers that are aimed at the kinds of non-relational integer abstract domains that are heavily used by today's production compilers. It exploits a simple but novel technique for breaking the synthesis problem into parts: each of our transformers is the meet of a collection of simpler, sound transformers that are synthesized such that each new piece fills a gap in the precision of the final transformer. Our design point is bulk automation: no sketches are required. Transformers are verified by lowering to a previously created SMT dialect of MLIR. Each of our synthesized transformers is provably sound and some (17 percent) are more precise than those provided by LLVM.

cs.PL↗

arXiv2025

Bootstrapping Fuzzers for Compilers of Low-Resource Language Dialects Using Language Models

Modern extensible compiler frameworks-such as MLIR-enable rapid creation of domain-specific language dialects. This flexibility, however, makes correctness harder to ensure as the same extensibility that accelerates development also complicates maintaining the testing infrastructure. Extensible languages require automated test generation that is both dialect-agnostic (works across dialects without manual adaptation) and dialect-effective (targets dialect-specific features to find bugs). Existing approaches typically sacrifice one of these goals by either requiring manually constructed seed corpora for each dialect, or by failing to be effective. We present a dialect-agnostic and dialect-effective grammar-based and coverage-guided fuzzing approach for extensible compilers that combines two key insights from existing work: (i) the grammars of dialects, which already encode the structural and type constraints, can often be extracted automatically from the dialect specification; and (ii) these grammars can be used in combination with pre-trained large language models to automatically generate representative and diverse seed inputs from the full dialect space without requiring any manual input or training data. These seeds can then be used to bootstrap coverage-guided fuzzers. We built this approach into a tool, Germinator. When evaluated on six MLIR projects spanning 91 dialects, Germinator generated seeds improve line coverage by 10-120% over grammar-based baselines. We compare against grammar-based baselines because they are the only class of existing automatic seed generators that can be applied uniformly across MLIR's heterogeneous dialect ecosystem. Germinator discovers 88 previously unknown bugs (40 confirmed), including 23 in dialects with no prior automated test generators, demonstrating effective and controllable testing of low-resource dialects at scale.

cs.SE↗

arXiv2025

ChopChop: a Programmable Framework for Semantically Constraining the Output of Language Models

Language models (LMs) can generate code but cannot guarantee its correctness$\unicode{x2014}$often producing outputs that violate type safety, program invariants, or other semantic properties. Constrained decoding offers a solution by restricting generation to only produce programs that satisfy user-defined properties. However, existing methods are either limited to syntactic constraints or rely on brittle, ad hoc encodings of semantic properties over token sequences rather than program structure. We present ChopChop, the first programmable framework for constraining the output of LMs with respect to semantic properties. ChopChop introduces a principled way to construct constrained decoders based on analyzing the space of programs a prefix represents. It formulates this analysis as a realizability problem which is solved via coinduction, connecting token-level generation with structural reasoning over programs. We demonstrate ChopChop's generality by using it to enforce (1) equivalence to a reference program and (2) type safety. Across a range of models and tasks, ChopChop improves success rates while maintaining practical decoding latency.

cs.PL↗

arXiv2025

Constrained Adaptive Rejection Sampling

Language Models (LMs) are increasingly used in applications where generated outputs must satisfy strict semantic or syntactic constraints. Existing approaches to constrained generation fall along a spectrum: greedy constrained decoding methods enforce validity during decoding but distort the LM's distribution, while rejection sampling (RS) preserves fidelity but wastes computation by discarding invalid outputs. Both extremes are problematic in domains such as program fuzzing, where both validity and diversity of samples are essential. We present Constrained Adaptive Rejection Sampling (CARS), an approach that strictly improves the sample-efficiency of RS without distributional distortion. CARS begins with unconstrained LM sampling and adaptively rules out constraint-violating continuations by recording them in a trie and subtracting their probability mass from future draws. This adaptive pruning ensures that prefixes proven invalid are never revisited, acceptance rates improve monotonically, and the resulting samples exactly follow the constrained distribution. In experiments on a variety of domains -- e.g., program fuzzing and molecular generation -- CARS consistently achieves higher efficiency -- measured in the number of LM forward passes per valid sample -- while also producing stronger sample diversity than both GCD and methods that approximate the LM's distribution.

cs.AI↗

arXiv2025

Semantics of Sets of Programs

Applications like program synthesis sometimes require proving that a property holds for all of the infinitely many programs described by a grammar - i.e., an inductively defined set of programs. Current verification frameworks overapproximate programs' behavior when sets of programs contain loops, including two Hoare-style logics that fail to be relatively complete when loops are allowed. In this work, we prove that compositionally verifying simple properties for infinite sets of programs requires tracking distinct program behaviors over unboundedly many executions. Tracking this information is both necessary and sufficient for verification. We prove this fact in a general, reusable theory of denotational semantics that can model the expressivity and compositionality of verification techniques over infinite sets of programs. We construct the minimal compositional semantics that captures simple properties of sets of programs and use it to derive the first sound and relatively complete Hoare-style logic for infinite sets of programs. Thus, our methods can be used to design minimally complex, compositional verification techniques for sets of programs.

cs.PL↗

arXiv2025

Flexible and Efficient Grammar-Constrained Decoding

Large Language Models (LLMs) are often asked to generate structured outputs that obey precise syntactic rules, such as code snippets or formatted data. Grammar-constrained decoding (GCD) can guarantee that LLM outputs matches such rules by masking out tokens that will provably lead to outputs that do not belong to a specified context-free grammar (CFG). To guarantee soundness, GCD algorithms have to compute how a given LLM subword tokenizer can align with the tokens used by a given context-free grammar and compute token masks based on this information. Doing so efficiently is challenging and existing GCD algorithms require tens of minutes to preprocess common grammars. We present a new GCD algorithm together with an implementation that offers 17.71x faster offline preprocessing than existing approaches while preserving state-of-the-art efficiency in online mask computation.

cs.CL↗

arXiv2025

Constrained Sampling for Language Models Should Be Easy: An MCMC Perspective

Constrained decoding enables Language Models (LMs) to produce samples that provably satisfy hard constraints. However, existing constrained-decoding approaches often distort the underlying model distribution, a limitation that is especially problematic in applications like program fuzzing, where one wants to generate diverse and valid program inputs for testing purposes. We propose a new constrained sampling framework based on Markov Chain Monte Carlo (MCMC) that simultaneously satisfies three core desiderata: constraint satisfying (every sample satisfies the constraint), monotonically converging (the sampling process converges to the true conditional distribution), and efficient (high-quality samples emerge in few steps). Our method constructs a proposal distribution over valid outputs and applies a Metropolis-Hastings acceptance criterion based on the LM's likelihood, ensuring principled and efficient exploration of the constrained space. Empirically, our sampler outperforms existing methods on both synthetic benchmarks and real-world program fuzzing tasks.

cs.AI↗

arXiv2025

Verifying Solutions to Semantics-Guided Synthesis Problems

Semantics-Guided Synthesis (SemGuS) provides a framework to specify synthesis problems in a solver-agnostic and domain-agnostic way, by allowing a user to provide both the syntax and semantics of the language in which the desired program should be synthesized. Because synthesis and verification are closely intertwined, the SemGuS framework raises the problem of how to verify programs in a solver and domain-agnostic way. We prove that the problem of verifying whether a program is a valid solution to a SemGuS problem can be reduced to proving validity of a query in the `CLP calculus, a fixed-point logic that generalizes Constrained Horn Clauses and co-Constrained Horn Clauses. Our encoding into `CLP allows us to further classify the SemGuS verification problems into ones that are reducible to validity of (i) first-order-logic formulas, (ii) Constrained Horn Clauses, (iii) co-Constrained Horn Clauses, and (iv) `CLP queries. Furthermore, our encoding shines light on some limitations of the SemGuS framework, such as its inability to model nondeterminism and reactive synthesis. We thus propose a modification to SemGuS that makes it more expressive, and for which verifying solutions is exactly equivalent to proving validity of a query in the `CLP calculus. Our implementation of SemGuS verifiers based on the above encoding can verify instances that were not even encodable in previous work. Furthermore, we use our SemGuS verifiers within an enumeration-based SemGuS solver to correctly synthesize solutions to SemGuS problems that no previous SemGuS synthesizer could solve.

cs.PL↗

arXiv2025

Unrealizability Logic

We consider the problem of establishing that a program-synthesis problem is unrealizable (i.e., has no solution in a given search space of programs). Prior work on unrealizability has developed some automatic techniques to establish that a problem is unrealizable; however, these techniques are all black-box, meaning that they conceal the reasoning behind why a synthesis problem is unrealizable. In this paper, we present a Hoare-style reasoning system, called unrealizability logic for establishing that a program-synthesis problem is unrealizable. To the best of our knowledge, unrealizability logic is the first proof system for overapproximating the execution of an infinite set of imperative programs. The logic provides a general, logical system for building checkable proofs about unrealizability. Similar to how Hoare logic distills the fundamental concepts behind algorithms and tools to prove the correctness of programs, unrealizability logic distills into a single logical system the fundamental concepts that were hidden within prior tools capable of establishing that a program-synthesis problem is unrealizable.

cs.PL↗

arXiv2025

LOUD: Synthesizing Strongest and Weakest Specifications

This paper tackles the problem of synthesizing specifications for nondeterministic programs. For such programs, useful specifications can capture demonic properties, which hold for every nondeterministic execution, but also angelic properties, which hold for some nondeterministic execution. We build on top of a recently proposed a framework by Park et al. in which given (i) a quantifier-free query posed about a set of function definitions (i.e., the behavior for which we want to generate a specification), and (ii) a language L in which each extracted property is to be expressed (we call properties in the language L-properties), the goal is to synthesize a conjunction of L-properties such that each of the conjunct is a strongest L-consequence for the query: each property is an over-approximation of the query and there is no other L-property that over-approximates the query and is strictly more precise than each property. This framework does not apply to nondeterministic programs for two reasons: it does not support existential quantifiers in queries (which are necessary to expressing nondeterminism) and it can only compute L-consequences, i.e., it is unsuitable for capturing both angelic and demonic properties. This paper addresses these two limitations and presents a framework, LOUD, for synthesizing both strongest L-consequences and weakest L-implicants (i.e., under-approximations of the query) for queries that can involve existential quantifiers. We implement a solver, ASPIRE, for problems expressed in LOUD which can be used to describe and identify sources of bugs in both deterministic and nondeterministic programs, extract properties from concurrent programs, and synthesize winning strategies in two-player games.

cs.PL↗

arXiv2025

Perceptions of the Fairness Impacts of Multiplicity in Machine Learning

Machine learning (ML) is increasingly used in high-stakes settings, yet multiplicity - the existence of multiple good models - means that some predictions are essentially arbitrary. ML researchers and philosophers posit that multiplicity poses a fairness risk, but no studies have investigated whether stakeholders agree. In this work, we conduct a survey to see how multiplicity impacts lay stakeholders' - i.e., decision subjects' - perceptions of ML fairness, and which approaches to address multiplicity they prefer. We investigate how these perceptions are modulated by task characteristics (e.g., stakes and uncertainty). Survey respondents think that multiplicity threatens the fairness of model outcomes, but not the appropriateness of using the model, even though existing work suggests the opposite. Participants are strongly against resolving multiplicity by using a single model (effectively ignoring multiplicity) or by randomizing the outcomes. Our results indicate that model developers should be intentional about dealing with multiplicity in order to maintain fairness.

cs.CY↗

arXiv2024

Synthesizing Formal Semantics from Executable Interpreters

Program verification and synthesis frameworks that allow one to customize the language in which one is interested typically require the user to provide a formally defined semantics for the language. Because writing a formal semantics can be a daunting and error-prone task, this requirement stands in the way of such frameworks being adopted by non-expert users. We present an algorithm that can automatically synthesize inductively defined syntax-directed semantics when given (i) a grammar describing the syntax of a language and (ii) an executable (closed-box) interpreter for computing the semantics of programs in the language of the grammar. Our algorithm synthesizes the semantics in the form of Constrained-Horn Clauses (CHCs), a natural, extensible, and formal logical framework for specifying inductively defined relations that has recently received widespread adoption in program verification and synthesis. The key innovation of our synthesis algorithm is a Counterexample-Guided Synthesis (CEGIS) approach that breaks the hard problem of synthesizing a set of constrained Horn clauses into small, tractable expression-synthesis problems that can be dispatched to existing SyGuS synthesizers. Our tool Synantic synthesized inductively-defined formal semantics from 14 interpreters for languages used in program-synthesis applications. When synthesizing formal semantics for one of our benchmarks, Synantic unveiled an inconsistency in the semantics computed by the interpreter for a language of regular expressions; fixing the inconsistency resulted in a more efficient semantics and, for some cases, in a 1.2x speedup for a synthesizer solving synthesis problems over such a language.

cs.PL↗

arXiv2024

Automating Pruning in Top-Down Enumeration for Program Synthesis Problems with Monotonic Semantics

In top-down enumeration for program synthesis, abstraction-based pruning uses an abstract domain to approximate the set of possible values that a partial program, when completed, can output on a given input. If the set does not contain the desired output, the partial program and all its possible completions can be pruned. In its general form, abstraction-based pruning requires manually designed, domain-specific abstract domains and semantics, and thus has only been used in domain-specific synthesizers. This paper provides sufficient conditions under which a form of abstraction-based pruning can be automated for arbitrary synthesis problems in the general-purpose Semantics-Guided Synthesis (SemGuS) framework without requiring manually-defined abstract domains. We show that if the semantics of the language for which we are synthesizing programs exhibits some monotonicity properties, one can obtain an abstract interval-based semantics for free from the concrete semantics of the programming language, and use such semantics to effectively prune the search space. We also identify a condition that ensures such abstract semantics can be used to compute a precise abstraction of the set of values that a program derivable from a given hole in a partial program can produce. These precise abstractions make abstraction-based pruning more effective. We implement our approach in a tool, Moito, which can tackle synthesis problems defined in the SemGuS framework. Moito can automate interval-based pruning without any a-priori knowledge of the problem domain, and solve synthesis problems that previously required domain-specific, abstraction-based synthesizers -- e.g., synthesis of regular expressions, CSV file schema, and imperative programs from examples.

cs.PL↗