LLM benchmark scores can be precise even when the observation protocol does not identify the behavioral property they are intended to measure. In a controlled, solver-grounded setting, we formalize a protocol-level identifiability audit over a finite behavioral policy class: given policies H, observation support O, and estimand $τ$, we test whether O separates every pair with different $τ$. The audit requires zero model calls and resolves our diagnostic case: base-only observation collapses seven frozen deterministic policies into one equivalence class; full support yields seven classes and no cross-estimand collisions; every leave-one-out support retains a constructive collision witness. Empirically, both constrained-generation variants have pair-validity 1.0, yet base accuracy and selective-response fidelity diverge - 0.620 versus 0.324 across six balanced oracle-transition directions (cluster-bootstrap 95% CI [0.600, 0.642] vs. [0.304, 0.345]) - and the gap recurs on a second deterministic source (0.646 vs. 0.331). The audit also synthesizes a minimum identifying support $O^*$ for the frozen policy class: two cells instead of the full 36-cell tensor. This case shows how evaluation-design validity can be checked structurally before model inference and why base correctness does not determine intervention-response fidelity.
Scientific reasoning benchmarks typically evaluate large language models (LLMs) using final-answer accuracy. However, a correct answer does not necessarily demonstrate the reasoning capability targeted by the problem. We identify Solution Hacking, a failure mode in which an LLM reaches the correct answer through invalid shortcuts, such as numerical search, enumeration, guessing, or answer-first verification, without providing a valid task-targeted derivation. We systematically analyze this phenomenon across difficulty levels, scientific domains, and frontier models. Solution hacking increases sharply with benchmark difficulty, from 2.2\% on common problems to 28.3\% on Olympiad-level problems and 37.4\% on HLE. Moreover, 8.2\%-44.1\% of answers credited as correct across frontier models are identified as hacked solutions. We further develop expert-inspired anti-hacking strategies, including an automatic judge and a test-time instruction. The results show that suppressing shortcut behavior substantially reduces reported accuracy while having a smaller effect on correct and non-hacked accuracy. These findings reveal that answer-only evaluation can overestimate the scientific reasoning capabilities of frontier LLMs.
Language can be viewed as a formalized subset of thought: a consequence-governed symbolic structure projected from wider situated cognition. Large language models trained at scale exhibit compensatory emergence: sparse architectural primitives support in-context learning, multi-step reasoning, tool use, and chain of thought. Yet a language-first probabilistic architecture inherits substantive, substrate, and high-level incompletenesses relative to human cognition. Their coexistence makes an LLM a human-like thought-form generator that reconstructs increasingly human-like reasoning forms from an incomplete substrate. We ask whether emergence can compensate for every missing distinction. We formalize the philosophical premise as the Symbolization--Substructure Thesis and introduce emergence invariance. For a scale-indexed family acting through a shared task interface $φ$, $\mathcal{R}_s^*=\mathcal{R}_φ^*+C_s$: scale can reduce the compensation gap $C_s$, while a positive interface floor $\mathcal{R}_φ^*$ persists. We prove that, under a fixed input law, one interface is universally no less informative exactly when its completed information $σ$-field refines the other, and that total compensation occurs exactly when both the interface floor and asymptotic compensation gap vanish. The framework unifies existing results on grounding, memory, position, attention, Bayesian inheritance, scientific abduction, and reasoning control. In a matched DeepSeek V4-Flash API study, thinking improves pointer chasing from $0/16$ to $14/16$ when relevant distinctions are available; exact observational twins remain at their $50\%$ construction floor; and restoring decisive memory moves matched performance from $50\%$ to $100\%$. These results provide initial evidence for the predicted separation between scaling within an interface and refining the interface itself.
When LLMs exhibit uneven performance across planning tasks, these gaps are often attributed to task difficulty. We argue that this explanation is incomplete, as task-level variation may reflect distinct latent planning competencies rather than differences along a single ability spectrum. We study this question on ACPBench-Hard by evaluating multiple LLM families under varying test-time reasoning budgets and applying a multidimensional item response theory model to uncover the latent competency structure underlying LLM planning. The analysis reveals two principal dimensions that shape planning performance: operational reasoning, the ability to evaluate local action applicability and immediate state transitions, and structural enumeration, the ability to reason about goal reachability and landmark structure. Operational reasoning improving under model scaling and longer reasoning traces, while structural enumeration remains comparatively insensitive. Our findings motivate competency-level evaluation of LLM planning, shifting the focus from whether models improve overall to which planning competencies improve, under what conditions, and why.
Dmitry Beresnev, Vladimir Makharev, Roman Khalikov +2cs.AI
Many reasoning tasks are not well described by a single left-to-right chain: a solver may need to pursue a plausible branch, observe delayed failure, and return to the latest prefix that can still be completed. We introduce Pyligent, a training and inference framework inspired by the Diligent Learner formulation that represents reasoning as validated search over partial solution chains. A task validator labels generated continuations and failures, and the resulting search trees are converted into supervised targets for three actions: continue, finish, and backtrack, with optional traces that summarize abandoned branches. We evaluate Pyligent on a hidden directed graph task designed to isolate delayed-failure recovery, and on structured reasoning domains with exact validators, including $4{\times}4$ Sudoku, Sudoku with reasoning traces, and Blocksworld. Compared with gold-only supervised fine-tuning, Pyligent improves solve rate by $72.7$ percentage points on hidden graphs, by $17$ and $18$ points on mixed and expert Sudoku, by $27$ and $14$ points on mixed and expert Sudoku with reasoning traces, and by $13$ points on Blocksworld. These results suggest that explicit failed-branch supervision can teach useful recovery behavior beyond imitation of polished solution chains.
Knowledge graphs can guide large language models (LLMs) reasoning, but the graph seen by a system is usually a retrieved, linked, temporally scoped, and incomplete evidence state rather than a complete account of truth. We develop a theoretical perspective on grounding observable LLM trajectories under such incomplete graph evidence.The evidence state induces entity anchors, typed relation residuals, path energies, and support regions, while the language model supplies a prior over candidate trajectories. We show that, under open-world incompleteness, no hard rule based only on the observed state can both reject every false unsupported trajectory and retain every true-but-unobserved one.We then characterize soft grounding as a KL-regularized deformation of the LLM prior: finite slack preserves support for unsupported but non-contradicted trajectories, whereas hard conditioning appears as an infinite-penalty limit.The framework also yields stability bounds under evidence perturbations and clarifies the constraint regimes appropriate for GraphRAG, KGQA, graph agents, constrained decoding, and faithful generation. The claims are evidence-relative: KG compatibility is treated as declared support, not factual truth.
The Abstraction and Reasoning Corpus (ARC;~\citealp{chollet2019measure}) contains tasks that require summarizing patterns from limited grid samples and predicting output grids. Recently, many large language model based approaches have attempted to transform it into a text-based reasoning task. However, methods based on open-source models have generally yielded unsatisfactory results, while those relying on closed-source models are too costly. Current efforts mainly focus on data augmentation, constructing ARC-like data for more comprehensive supervised fine-tuning. In this work, we argue that solving ARC-like problems requires not only \textit{positive} sample supervision but also the ability to improve model reasoning by distinguishing \textit{negative} samples. To this end, we draw on the idea of preference alignment and propose \textsc{DiARC}, a method that constructs preference pairs to enable the model to distinguish between them. Specifically, we propose three ways to construct negative samples, including output-level visual transformations, DSL-level rule inversion, and task-specific rule editing. The resulting negative samples provide informative near-miss alternatives while keeping the observed demonstrations unchanged. Experimental results across multiple ARC-like benchmarks show that \textsc{DiARC} consistently improves performance over baseline models. The code is released at https://github.com/szu-tera/DiARC.
Yuxu Zhou, Ondřej Kuželka, Yuyi Wang +2cs.AI cs.CL
We present CombEval, a dynamic benchmark for evaluating combinatorial counting in large language models. CombEval represents each problem as a typed Cofola specification over entities, combinatorial objects, object dependencies, and constraints, enabling controlled generation of natural-language counting problems with exact solver-verified answers. Unlike static collections, CombEval supports systematic variation of object type, entity scale, constraint count, and reasoning depth. We evaluate 11 LLMs under direct and code-augmented settings and find that models remain brittle on ordered objects, indistinguishable elements, relatively positional constraints, and nested object dependencies. Error analysis further identifies failures in constraint interpretation and counting principles. CombEval provides a diagnostic testbed for studying when and why LLMs fail at combinatorial reasoning. The code and generated benchmark suites are publicly available at \url{https://github.com/YuxuZhou-CN/combination-problem-generation}.