Tag
This paper proposes a dependency-graph framework to formalize compositional reasoning in language models and evaluates the impact of reinforcement learning post-training, finding an asymmetry where composed-skill training transfers more readily to decomposed tasks than vice versa.
This paper introduces a multi-stage rule-chaining framework for compositional and interpretable cognitive reasoning on the Abstraction and Reasoning Corpus (ARC) benchmark, achieving over 95% accuracy through integrated deterministic solvers and hierarchical abstraction.
CORE introduces a distillation method that transfers compositional ranking judgments from a reranker to an embedding model using a Rank-KL objective, enhancing compositional retrieval performance across benchmarks without compromising standard tasks.
The paper explores how looped language models, which use iterative latent computation, improve compositional tool calling in agentic systems, showing benefits for multi-step API interactions.
This paper introduces a framework that decomposes compound logical answer options into atomic judgments and uses an operator-constrained integer linear program to improve large language model reasoning over AND, OR, and NEITHER/NOR operators. It achieves significant F1 gains on LOGICAL-COMMONSENSEQA and a new benchmark LOGICAL-SATA.
MEGA is a self-evolving infrastructure for coding-agent optimization that distills reusable wisdom from sessions, composes it via a typed Wisdom Graph, and uses operational evidence to continuously improve both the agents and the knowledge guiding their optimization.
Introduces RECON, a benchmark for evaluating compositional reasoning over long contexts in LLM-based agents, spanning 24 case files across criminal, medical, and financial domains. The best non-oracle system achieves only 22.4% accuracy, revealing substantial limitations in current memory architectures.
This paper investigates whether reinforcement learning post-training can compose primitive skills into higher-level compositional strategies, using a fully observable rewrite-grammar environment. The authors find that RL reorganizes primitive competence through phased compositional mechanisms, while rejection fine-tuning plateaus due to producing many invalid shortcut-like rewrites.
Introduces CDR-Bench, a benchmark with 3,462 tasks to evaluate LLMs' ability to faithfully execute compositional, order-sensitive data refinement recipes. Experiments on 10+ LLMs reveal significant performance degradation in compositional and order-sensitive settings, highlighting a lack of procedural faithfulness.
This paper investigates holographic reduced representations for zero-shot compositional reasoning in knowledge graphs, finding that while single-hop performance is strong, composition fails due to retrieval capacity and interference effects in the superposed memory, not the bind-unbind algebra.
R-APS (Reflective Adversarial Pareto Search) is a novel method for constrained design tasks that addresses three structural failures in LLM-based agentic systems—error propagation, robustness evaluation, and knowledge invalidation—through reasoning-mode decomposition across three timescales, requiring no fine-tuning. Evaluated on planar mechanism synthesis, it achieves 3.5x tighter robustness certificates, 46% faster iterations-to-first-admission, and 2.1x Chamfer-distance reduction over baselines.
MAVEN is a lightweight symbolic reasoning scaffold that improves generalization in agentic tool calling by using modular verification and adaptive tool orchestration. It achieves significant accuracy gains on a new stress-test benchmark (MAVEN-Bench) and remains competitive with proprietary models at a fraction of the cost.
This paper introduces 'composition collapse', a phenomenon where language models with stable factual knowledge still fail to compose that knowledge into correct multi-hop reasoning, and proposes a double-gate protocol to isolate composition failure from atomic knowledge instability.
This research paper investigates how shortcut solutions learned by Transformer models, specifically BERT, impair their ability to perform continual compositional reasoning. It contrasts BERT with ALBERT, finding that ALBERT's recurrent nature offers better inductive bias for continual learning tasks.
The Amazing Agent Race (AAR) introduces a new benchmark with 1,400 directed acyclic graph (DAG) puzzle instances to evaluate LLM agents on fork-merge tool chains and Wikipedia navigation. Evaluations reveal agents excel at tool-use (errors <17%) but struggle with navigation (27-52% of failures), exposing a critical gap invisible to existing linear benchmarks.
Proposes Slipform, a training framework that uses lexical concreteness to select harder negatives and a margin-based Cement loss, boosting compositional reasoning in vision-language models.