What Would You Propose for Better LLMs Reasoning?)
Yet Another Study on LLM Reasoning Problems: There Is No True Reasoning (Like in Humans), Just Sequence-to-Sequence Modeling
Current LLMs don’t actually reason — they perform sequence generation based on pre-trained patterns. While this works well for familiar tasks, novel or unseen challenges often lead to failures. All these “thinking” models do internal context engineering (you can see that intermediate generations) and some optimization techniques for “better” results.
How do humans solve novel tasks? Simple tactics: decomposition, transformation, reflection, analogies, research, heuristics, etc. Do the same, guys.
Any task/sub-task solvable through computation can be transformed into code and executed to get a solution.
Nice picture from Grigory:
While state-of-the-art large language models (LLMs) demonstrate advanced reasoning capabilities — achieving remarkable performance on challenging competitive math and coding benchmarks — they also frequently fail on tasks that are easy for humans. This work studies the performance of frontier LLMs on a broad set of such “easy” reasoning problems. By extending previous work in the literature, we create a suite of procedurally generated simple reasoning tasks, including counting, first-order logic, proof trees, and travel planning, with changeable parameters (such as document length. or the number of variables in a math problem) that can arbitrarily increase the amount of computation required to produce the answer while preserving the fundamental difficulty. While previous work showed that traditional, non-thinking models can be made to fail on such problems, we demonstrate that even state-of-the-art thinking models consistently fail on such problems and for similar reasons (e.g., statistical shortcuts, errors in intermediate steps, and difficulties in processing long contexts). To further understand the behavior of the models, we introduce the Unpuzzles dataset, a different “easy” benchmark consisting of trivialized versions of well-known math and logic puzzles. Interestingly, while modern LLMs excel at solving the original puzzles, they tend to fail on the trivialized versions, exhibiting several systematic failure patterns related to memorizing the originals. We show that this happens even if the models are otherwise able to solve problems with different descriptions but requiring the same logic. Our results highlight that out-of-distribution generalization is still problematic for frontier language models and the new generation of thinking models, even for simple reasoning tasks, and making tasks easier does not necessarily imply improved performance.
Structured reasoning over natural language inputs remains a core challenge in artificial intelligence, as it requires bridging the gap between unstructured linguistic expressions and formal logical representations. In this paper, we propose a novel bi-level framework that maps language to logic through a two-stage process: high-level task abstraction and low-level logic generation. At the upper level, a large language model (LLM) parses natural language queries into intermediate structured representations specifying the problem type, objectives, decision variables, and symbolic constraints. At the lower level, the LLM uses these representations to generate symbolic workflows or executable reasoning programs for accurate and interpretable decision making. The framework supports modular reasoning, enforces explicit constraints, and generalizes across domains such as mathematical problem solving, question answering, and logical inference. We further optimize the framework with an end-to-end bi-level optimization approach that jointly refines both the high-level abstraction and low-level logic generation stages. Experiments on multiple realistic reasoning benchmarks demonstrate that our approach significantly outperforms existing baselines in accuracy, with accuracy gains reaching as high as 40%. Moreover, the bi-level design enhances transparency and error traceability, offering a promising step toward trustworthy and systematic reasoning with LLMs.
In this paper, we proposed Lang2Logic, a bilevel structured reasoning framework that aims to bridge the gap between natural language understanding and logical formalization. Our approach explicitly separates task abstraction from logic generation, allowing LLMs to construct structured formal models and then generate executable symbolic workflows. This design captures the underlying logic of complex problems more effectively, facilitating more accurate, interpretable, and generalizable reasoning. We introduced a two-stage training strategy, including a supervised cold-start phase using a model-augmented dataset and a bilevel reinforcement learning algorithm for joint optimization. This training regimen enables the framework to effectively coordinate abstraction and execution, significantly enhancing its reasoning performance. Extensive experiments across diverse reasoning benchmarks, demonstrated that our framework consistently outperforms different baselines, achieving over 10% average accuracy gains. These results confirm that our Lang2Logic can offer a promising direction for the future of systematic reasoning with LLMs.
Human intelligence emerges from the interaction of specialized brain networks, each dedicated to distinct cognitive functions such as language processing, logical reasoning, social understanding, and memory retrieval. Inspired by this biological observation, we introduce the Mixture of Cognitive Reasoners (MiCRo) architecture and training paradigm: a modular transformer-based language model with a training curriculum that encourages the emergence of functional specialization among different modules. Inspired by studies in neuroscience, we partition the layers of a pretrained transformer model into four expert modules, each corresponding to a well-studied cognitive brain network. Our Brain-Like model has three key benefits over the state of the art: First, the specialized experts are highly interpretable and functionally critical, where removing a module significantly impairs performance on domain-relevant benchmarks. Second, our model outperforms comparable baselines that lack specialization on seven reasoning benchmarks. And third, the model’s behavior can be steered at inference time by selectively emphasizing certain expert modules (e.g., favoring social over logical reasoning), enabling fine-grained control over the style of its response. Our findings suggest that biologically inspired inductive biases involved in human cognition lead to significant modeling gains in interpretability, performance, and controllability.
AI systems that “think” in human language offer a unique opportunity for AI safety: we can monitor their chains of thought (CoT) for the intent to misbehave. Like all other known AI oversight methods, CoT monitoring is imperfect and allows some misbehavior to go unnoticed. Nevertheless, it shows promise and we recommend further research into CoT monitorability and investment in CoT monitoring alongside existing safety methods. Because CoT monitorability may be fragile, we recommend that frontier model developers consider the impact of development decisions on CoT monitorability.
This offers a unique opportunity for AI safety: chain of thought monitoring. A CoT monitor is an automated system that reads the CoT of a reasoning model and other relevant information and flags suspicious or potentially harmful interactions. These responses could then be blocked, or replaced with safer actions, or reviewed in more depth (Greenblatt et al., 2024b). More generally, by studying the CoT, we can gain some insight into how our AI agents think and what goals they have (see Section 3).
