Tag
This paper identifies the verification gap as a key constraint in AI reasoning progress, provides theoretical and empirical analysis of verifier soundness trade-offs, and introduces proof-carrying cognition to enhance reward settlement in reinforcement learning.
A repair tool that infers fixes from a single worked example and successfully repairs 5 out of 5 bugs in unseen code using only 4 CPU instructions and zero tokens, with exhaustive verification.
ClosureBench is a constructive benchmark for evaluating language models on compositional graph reasoning tasks, revealing that frontier models degrade with complexity while a fine-tuned model maintains performance through program synthesis.
The paper introduces aDSL, a co-designed domain-specific language and multi-agent system that improve LLM-driven 3D program synthesis through relational operators and iterative feedback, enhancing robustness and controllability in 3D content creation.
The article investigates whether recurring LLM workloads can be replaced by automatically synthesized deterministic pipelines of typed ML/NLP operators, and asks for feedback on feasibility and approaches.
This paper introduces two library combinators, prune and defrel/bank, that bring bottom-up enumeration with observational deduplication to miniKanren relational programming, improving program-by-example synthesis performance.
This paper introduces Brevis, a lossless tensor compression method that formulates compression as program synthesis using a typed DSL. It achieves 33.93% storage reduction on public checkpoints and outperforms general-purpose and tensor-specific compressors.
SpecFirst proposes making behavioral specification elicitation a first-class step in agent-based program synthesis, separating it from exploration and coding. Evaluated on 200 program instances, it significantly improves test pass rates and binary exploration coverage across multiple models.
MindForge is an automated pipeline that converts open-source command-line programs into source-free training environments for small language models. Fine-tuning Qwen3.6-27B on trajectories generated by GLM-5.2 significantly improves its performance on software engineering benchmarks, achieving results comparable to larger models.
ARCANA is a reflective multi-agent framework that decomposes ARC-AGI-2 abstract reasoning tasks into iterative perception, hypothesis generation, symbolic execution, and reflective refinement, improving reasoning efficiency under strict constraints.
Anima Anandkumar announces four Lean-related papers from their group at ICML workshops, covering verified ML systems, functional program synthesis, proof assistant interoperability, and scientific reasoning, positioning Lean as infrastructure for AI.
This paper presents cost-effective agent harnesses for ARC-AGI-1 that achieve strong performance using DeepSeek V3.2 without fine-tuning, via an Explorer-Definer Pipeline and a Reflective Orchestrator, achieving 67.25% pass@2 at low cost.
François Chollet predicts AI will converge towards intuition-guided symbolic world modeling, specifically deep learning-guided program synthesis, which allows systems to build compact, reusable, and generalizable mental models from minimal data.
Introduces AlgoEvolve, an LLM-driven evolutionary framework that generates and iteratively improves algorithmic trading strategies, with a meta-evolutionary outer loop that evolves prompts to guide the inner loop synthesis.
Introduces Evolving Programmatic Bottlenecks (EPB), a framework for interpreting neural combinatorial optimization policies by distilling black-box models into human-readable program portfolios using LLM-guided evolution.
Discusses the need for formal specifications in AI-generated code and introduces PICK, a tool that leverages human judgment to help programmers specify desired properties for LLM-generated regular expressions.
The article presents 'Lifting E-Graphs', a refined approach to e-graphs that explicitly encodes the context (dimension) of functions to resolve issues with variable naming, missed sharing, and accidental over-sharing, based on a semantic model of functions from R^n to R.
ReSyn is a generalized recursive framework for synthesizing regular expressions from examples, aiming to improve upon existing synthesis techniques.
This paper proposes using counterexample-guided learning for LLMs to perform regular-expression induction, where a verifier provides counterexamples to refine candidate expressions. The method significantly improves sample efficiency and success rates on challenging tasks, demonstrating that LLMs can benefit from structured feedback beyond treating it as additional data.
Introduces Semantic Reification, a new paradigm for random program generation, as detailed in this research paper.