The study examines whether language models maintain invariant internal representations when the order of mathematical rules is permuted, even though the correct answer remains unchanged. Using synthetic multi-step function‑composition tasks, the authors measure accuracy and a permutation signal‑to‑noise ratio (SNR) across 16 models (1B–8B parameters). They find a strong positive correlation (up to 0.86 Spearman) between higher accuracy on reordered problems and more distinct internal representations of rule orderings, indicating that successful reasoning can coexist with representation sensitivity.
arXiv:2607. 20520v1 Announce Type: new Abstract: Large language models (LLMs) are increasingly evaluated on mathematical problem solving, yet prior work often treats representationally equivalent formulations as interchangeable and conflates reasoning errors with interface failures.
arXiv:2510.00628v3 Announce Type: replace-cross Abstract: Large audio-language models (LALMs) are often used in tasks that involve reasoning over ordered options. An open question is whether their pr...
The paper investigates how large language models (LLMs) organize reasoning operations—such as problem formulation, goal decomposition, and deduction—within their hidden representation spaces. It shows that these operations are separable in held‑out representations, with peak separability in middle layers, and that token‑wise alignment of operations becomes more distributed across spans as layers deepen. Attention‑masking experiments reveal that representations aligned to operations at chunk onsets depend on prior reasoning context, indicating a geometric correspondence between linguistic reasoning expressions and internal model structure.
The paper introduces a benchmark of 3,600 exact‑rational word problems and 8,600 prompts that test whether language models give the same canonical answer when the same quantity is expressed in different numeric forms (decimal, fraction, percentage, number word, scientific notation, or unit‑converted). After normalizing answer syntax, canonical accuracy is high (0.969–0.996), but correctness across equivalent representations drops to 0.848–0.981, revealing that many errors stem from the evaluator’s number grammar rather than the models’ reasoning. The study also finds that representation consensus does not outperform paraphrase consensus on low‑error subsets and that certain models (e.g., Mistral Small 4) exhibit systematic unit‑conversion errors.
The study investigates whether large language models (LLMs) can generalize arithmetic reasoning across different input formats, such as numeric and verbal expressions in English, Spanish, and Italian. By using attribution patching, the researchers identified the specific internal circuits each model employs for numeric versus verbal arithmetic tasks. They found that the degree of overlap between a model’s numeric circuit and its verbal circuit predicts how well the model generalizes to verbal formats, explaining variations in performance across languages and formats without requiring labeled data.
The paper investigates the so‑called ‘structure tax’ that is often attributed to constraining large language model outputs to structured formats like JSON or XML. By evaluating various models, datasets, and schema designs, the authors find that the accuracy loss depends more on how the schema is organized than on the presence of structure itself. Specifically, reasoning‑first field ordering can match or surpass free‑form accuracy, whereas answer‑first ordering leads to significant drops, especially in smaller models. The study also shows that well‑designed schemas improve confidence calibration and the separability of correct versus incorrect representations in transformer layers.
arXiv:2608. 10420v1 Announce Type: new Abstract: Reasoning shortcuts are solutions of a neurosymbolic system's rules that produce correct predictions through unintended concepts.
The paper investigates whether reasoning representations—explanations for large language model outputs—aid humans in evaluating those outputs. A controlled human study tested six reasoning formats across tasks of varying complexity, measuring structural understanding, error detection, and trust calibration. Results revealed a mismatch: participants favored planning- and decomposition-based representations, yet simpler chain-of-thought traces better supported verification, trust, and interpretability, while preferred formats increased calibration risks.
Recursive reasoning models apply a small shared Transformer block many times to refine a latent state. This gives them large effective depth with few parameters and makes them strong on algorithmic ta...
arXiv:2606. 05025v1 Announce Type: cross Abstract: Large language models (LLMs) suffer from shortcut learning: they systematically fail on out-of-distribution (OOD) inputs whose semantic surface differs from training data, even when the logical structure is identical.
The paper introduces IndicReStruct, a benchmark dataset comprising two variants—GSM8K-Reordered and GSM8K-Voice—derived from GSM8K to test multilingual LLMs on Hindi and Malayalam. It evaluates six state‑of‑the‑art LLMs under constrained constituent reordering and active‑passive voice transformation, finding consistent and significant drops in mathematical reasoning performance. Qualitative error analysis and residual‑stream activation patching reveal that failures often stem from disrupted entity‑quantity alignment, with intermediate transformer layers playing a key role in restoring reasoning.
arXiv:2610.03551v1 Announce Type: new Abstract: Large language models (LLMs) have reached expert-level performance on competition mathematics largely through the volume of search placed around them:...