LLM Reasoning

Large language models (LLMs) are increasingly used in the mental health domain, yet it remains unclear how well they capture related biomedical knowledge and how reliably they apply it to clinically salient structured judgments. Here, we present a knowledge-graph (KG)-grounded benchmark for assessing LLMs on mental-health entity recognition, relation judgment, and two-hop reasoning. The benchmark is derived from PrimeKG and comprises nine task families with KG-supported answers and controlled negative options. Experiments across 15 closed- and open-source LLMs reveal a persistent recognition-to-judgment gap: leading models achieve near-ceiling performance on entity typing and on the small relation-typing subset, yet they still struggle with relation prediction and two-hop reasoning. Additionally, short KG-derived snippets benefit some models but degrade performance for others. Moreover, output-format reliability can substantially influence measured performance under constrained multiple-choice settings, highlighting the critical role of response validity in benchmark-based evaluation. MHGraphBench should therefore be interpreted as evaluating agreement with a curated mental-health slice of PrimeKG under a constrained multiple-choice interface, rather than as a direct assessment of real-world clinical safety.

Parallel reasoning, where a generator samples many candidate solutions and an aggregator selects the best, is one of the most effective forms of test-time scaling in large language models, and pairwise self-verification has become its strongest aggregation primitive. Yet pairwise verification carries a heavy cost: each judgment reads two complete solutions in full, and existing methods perform tens of such judgments per problem regardless of whether the comparison is informative. We introduce CAPS (Cascaded Adaptive Pairwise Selection), an inference-only framework that allocates verifier compute non-uniformly along two orthogonal axes: an evidence axis that adapts how much of each candidate the judge sees, and a distribution axis that adapts how comparisons are spread across the pool. CAPS instantiates these into a four-stage cascade with an optional rescue subroutine, and admits a closed-form verifier-token cost in which the per-candidate marginal cost is roughly halved relative to uniform full-evidence schedules. On four self-verifying models (Qwen3-14B, GPT-OSS-20B, Qwen3-4B-Instruct/Thinking) and five reasoning benchmarks spanning code (LiveCodeBench-v5/v6, CodeContests) and math (AIME 2025, HMMT 2025), CAPS outperforms the leading pairwise verifier on 14 of 20 suites while using 25.4% of its verifier-token budget on code, and outperforms pointwise self-verification on all 20. The trade-off suites admit an interpretable diagnostic in terms of the verifier's accuracy at partial versus full evidence, providing a concrete pre-deployment check for cascade suitability.

Reasoning-trained language models often spend more tokens on harder problems, but longer chains of thought do not show whether a model is merely computing for more steps or following a different internal trajectory. We study this distinction through hidden-state trajectories during chain-of-thought generation across competitive programming, mathematics, and Boolean satisfiability. Raw trajectory geometry is strongly shaped by generation length: longer generations mechanically alter path statistics, so difficulty-dependent comparisons are misleading without adjustment. After residualizing trajectory statistics on length, difficulty remains systematically coupled to corrected trajectory geometry across all domains studied. The clearest reasoning-specific separation appears in the code domain, where harder problems show more direct corrected trajectories and less heterogeneous local curvature in reasoning-trained models than in matched instruction-tuned baselines. Corrected difficulty-geometry coupling is weaker, but still present, in mathematics and Boolean satisfiability. Prompt-stage linear probes do not mirror the code-domain separation, and behavioral annotations show that stronger corrected coupling co-occurs with strategy shifts and uncertainty monitoring. Together, these findings establish length correction as a prerequisite for generation-time trajectory analysis and show that reasoning training can be associated with distinct corrected trajectory geometry, with the strength of the effect depending on the domain.

Test-time compute scaling is a primary axis for improving LLM reasoning. Existing methods primarily scale depth by extending a single reasoning trace. Scaling breadth by sampling multiple candidates in parallel is straightforward, but introduces a selection bottleneck: choosing the best candidate without a ground-truth verifier, since pointwise LLM judging is noisy and biased. To address this, we introduce OpenDeepThink, a population-based test-time compute framework that selects via pairwise Bradley-Terry comparison. Each generation, the LLM judges random pairs of candidates and aggregates votes via Bradley-Terry into a global ranking; top-ranked candidates are preserved and the top three quarters are mutated using the natural-language critiques produced during comparison; the bottom quarter is discarded. OpenDeepThink raises Gemini 3.1 Pro's effective Codeforces Elo by +405 points in eight sequential LLM-call rounds (~27 minutes wall-clock). The pipeline transfers across weaker and stronger models without retuning, and on the multi-domain HLE benchmark, gains appear concentrated in objectively verifiable domains and reverse in subjective ones. We release CF-73, a curated set of 73 expert-rated Codeforces problems with International Grandmaster annotation and 99% local-evaluation agreement against the official verdict.

Harmonized System (HS) tariff classification is a high-stakes, expert-level task in which a free-form product description must be mapped to a specific six- or eight-digit code under the General Interpretive Rules (GIR), section notes, chapter notes, and Explanatory Notes. The difficulty lies not in knowledge volume but in multi-dimensional rule reasoning: a correct classification must satisfy competing priority rules along several axes simultaneously, including material, form, function, essential character, the part-versus-whole boundary, and specific listing versus residual headings. End-to-end prompting of large language models fails characteristically by resolving one axis while ignoring the priority constraints on the others. We present a deterministic agentic workflow in contrast to self-planning agents: the control flow is fixed, language model calls are confined to narrow stages, and reflection and verification are retained as local mechanisms. This design yields interpretability by construction--each decision is decomposed into stage-wise structured outputs with verbatim citation of the chapter or section notes that bear on it. The architecture combines offline knowledge-engineering of the Chinese HS tariff with an online six-stage pipeline. Evaluated on HSCodeComp at the six-digit level, the workflow reaches 75.0% top-1 and 91.5% top-3 at four digits, and 64.2% top-1 and 78.3% top-3 at six digits with Qwen3.6-plus; an open-weight Qwen3.6-27B-FP8 backbone in non-thinking mode achieves 84.2% four-digit and 77.4% six-digit top-1 agreement with the frontier model. A two-stage manual audit of 226 six-digit disagreements suggests that a non-trivial fraction of HSCodeComp ground-truth labels may deviate from HS general rules; full adjudication records are released in the appendix as preliminary findings for community review.

Multi-step LLM reasoning over structured tables fails because planning and execution share no explicit cell-grounding contract. Existing methods constrain the planner to a left-to-right factorization at odds with table permutation invariance, and score intermediate states by generated content alone, overlooking cell grounding. We conduct a pilot study showing that diffusion language models (DLMs) produce more human-aligned and permutation-stable cell attention on tables than autoregressive models, with a 40.2% median reduction in attention-AUROC variability under row reordering. Motivated by this, we propose TABALIGN, a planned table reasoning framework that operationalizes the contract. TABALIGN pairs a masked DLM planner, whose bidirectional denoising emits plan steps as binary cell masks, with TABATTN, a lightweight verifier trained on 1,600 human-verified attention standards to score each step by its attention overlap with the plan-designated mask. Across eight benchmarks covering table question answering and fact verification, TABALIGN improves average accuracy by 15.76 percentage points over the strongest open-source baseline at comparable 8B-class scale, with a matched-backbone ablation attributing 2.87 percentage points of this gain to the DLM planner over an AR planner on a fixed reasoner. Cleaner DLM plans also accelerate downstream reasoning execution by 44.64%.

Chain-of-Thought (CoT) reasoning has become a foundation for eliciting multi-step reasoning in large language models, but recent studies show that its benefits do not scale monotonically with chain length: while longer CoT generally enables a model to tackle harder problems, on a given problem, accuracy typically increases with CoT length up to a point, after which it declines. We identify a major cause of this phenomenon: as the CoT grows, the model's attention to critical insights produced earlier in the trace gradually weakens, making those insights progressively less accessible when they are most needed. Therefore, we propose \textbf{InsightReplay}, a stateful reasoning approach in which the model periodically extracts critical insights from its reasoning trace and replays them near the active generation frontier, keeping them accessible as the reasoning scales. Extensive experiments on a $\mathbf{2}\!\times\!\mathbf{3}\!\times\!\mathbf{4}$ benchmark grid, covering model scales $\{\text{8B}, \text{30B}\}$, model families $\{\text{Qwen3.5}, \text{DeepSeek-R1-Distill-Qwen}, \text{Gemma-4}\}$, and reasoning benchmarks $\{\text{AIME}, \text{HMMT}, \text{GPQA Diamond}, \text{LiveCodeBench v5}\}$, show that 3-round InsightReplay yields accuracy gains across \textbf{all 24 settings}, with an averaged improvement of $\mathbf{+1.65}$ points over standard CoT, and a largest single-setting gain of $\mathbf{+9.2}$ points on R1-Distill-32B's LiveCodeBench v5 subset. Our results suggest that the effectiveness of test-time scaling depends not only on how much a model reasons, but also on whether critical intermediate insights remain accessible throughout long reasoning trajectories.

We present Darwin Family, a framework for training-free evolutionary merging of large language models via gradient-free weight-space recombination. We ask whether frontier-level reasoning performance can be improved without additional training, by reorganizing latent capabilities already encoded in existing checkpoints. Darwin introduces three key ideas: (i) a 14-dimensional adaptive merge genome enabling fine-grained component- and block-level recombination; (ii) MRI-Trust Fusion, which adaptively balances diagnostic layer-importance signals with evolutionary search through a learnable trust parameter; and (iii) an Architecture Mapper that enables cross-architecture breeding between heterogeneous model families. Empirically, the flagship Darwin-27B-Opus achieves 86.9% on GPQA Diamond, ranking #6 among 1,252 evaluated models, and outperforming its fully trained foundation model without any gradient-based training. Across scales from 4B to 35B parameters, Darwin models consistently improve over their parents, support recursive multi-generation evolution, and enable a training-free evolutionary merge that combines Transformer- and Mamba-based components. Together, the Darwin Family demonstrates that diagnostic-guided evolutionary merging is a practical and reproducible alternative to costly post-training pipelines for reasoning-centric language models.

Language models often generate long chain-of-thought traces, but it remains unclear how much of this reasoning is necessary for preserving the final prediction. We study this through the lens of overcomplete reasoning traces: generated traces that contain more intermediate steps than are needed to support the model's answer. We define the minimal core as the smallest subset of steps that preserves either the final answer or predictive distribution, and introduce metrics for compression ratio, redundancy mass, step necessity, and necessity concentration. Across six deliberative reasoning benchmarks spanning arithmetic, competition mathematics, expert scientific reasoning, and commonsense multi-hop QA, we find substantial overcompleteness: on average, 46% of steps are removable under greedy minimal-core extraction while preserving the original answer in 86% of cases. We also find that predictive support is concentrated: the top three steps account for 65% of measured necessity mass on average. Beyond compression, minimal cores expose a cleaner geometry of reasoning: compared with full traces, they improve correct-incorrect trace separation by 11 points, reduce estimated intrinsic dimensionality by 34%, and transfer across model families with 85% off-diagonal answer retention. Theoretically, we establish existence of minimal sufficient subsets, local irreducibility guarantees for greedy elimination, and certificates of overcompleteness and sparse necessity. Together, these results suggest that full reasoning traces are often verbose and overcomplete, while minimal cores isolate the effective support underlying language-model predictions.

Large language models (LLMs) often expose useful signals of self-monitoring: before solving a problem, they can estimate whether they are likely to succeed, and after solving it, they can judge whether their answer is likely to be correct. However, these signals are typically measured or elicited in isolation, rather than used to control inference. In this work, we ask whether LLMs possess latent metacognitive ability that can be turned into effective test-time control. Inspired by the Nelson--Narens theory from cognitive psychology, we propose a metacognitive harness that separates monitoring from reasoning. For each problem, the model first reports a pre-solve feeling-of-knowing (FOK) signal; after each solve attempt, it reports a post-solve judgment-of-learning (JOL) signal. Rather than treating these signals as passive confidence estimates, the harness turns them into an explicit control interface for reasoning: it decides when to trust the current solution, when to retry with compact metacognitive feedback, and when to pass multiple attempts to a final aggregator. Across text, code, and multimodal reasoning benchmarks, our harness substantially improves a fixed Claude Sonnet-4.6 base model without parameter updates or benchmark-specific fine-tuning. On the evaluated public benchmark snapshots, it raises pooled accuracy from 48.3 to 56.9 and exceeds the strongest listed leaderboard entries on the three primary evaluation settings: HLE-Verified, LiveCodeBench v6, and R-Bench-V. These results suggest that strong LLMs may already possess useful metacognitive ability, but require an explicit control harness to act on it during reasoning.

How much of the capability of a reasoning model exposed by repeated sampling can an imperfect selector recover? We study this problem as {\em inference-time amplification}, separating {\em coverage}---whether a useful candidate is generated---from {\em identifiability}---whether it can be recognized among the alternatives. We show that better coverage alone does not guarantee better selection, while even imperfect local selection signals can be amplified through repetition. We also characterize proposal-side blind spots that additional sampling cannot overcome. On SWE-bench Verified, a single GPT-5.4 nano trajectory solves 67.0% of tasks, while eight proposals contain a correct patch on 79.0%; critic--comparator selection reaches 76.4%, recovering 78.3% of the oracle-exposed gain. The same pattern appears across proposer families and in mathematical reasoning, competitive programming, and formal proof. On MATH L4--5 and GSM-Plus, our selector recovers 85% and 77% of the oracle gap and substantially outperforms majority vote. Holding proposal coverage fixed, providing the selector with additional repository information also improves recovery. Conversely, when the available selection signal is weak, redundant, or misaligned, amplification can diminish or fail. These results identify selection quality as the key factor governing how much capability exposed by repeated sampling can actually be recovered.

Chain-of-thought (CoT) reasoning with self-consistency improves performance by aggregating multiple sampled reasoning paths. In this setting, correctness is no longer tied to a single reasoning trace but to the aggregation rule over a pool of candidate paths, making aggregation uncertainty the central challenge. This issue is critical where confidently incorrect answers are far more costly than abstentions. We introduce a conformal procedure for CoT reasoning that directly addresses aggregation uncertainty. Our approach replaces majority voting with weighted score aggregation over reasoning paths and calibrates an abstention rule using conformal risk control. This approach leads to finite-sample guarantees on the confident-error rate--the probability that the system answers and is wrong. We further identify score separability as the key condition under which abstention provably improves selective accuracy, and derive closed-form expressions that predict accuracy gains from calibration data alone. The method is fully inference-time, and requires no retraining. Across four benchmarks, four open-source models, and three score classes, realized confident-error rates are consistent with the prescribed targets up to calibration-split and test-set variability. Our method achieves $90.1\%$ selective accuracy on GSM8K by abstaining on less than $5\%$ of problems, compared with $82\%$ accuracy under majority-voting baseline.

In current Large Language Models we can trust the production of smoothly flowing prose on the basis of the principles of machine learning. However, there is no comparably principled basis to justify trust in the content of the text produced. It appears to be conventional wisdom that addressing this issue by adding more principled reasoning is not computationally affordable. Here we propose a principled method of reasoning that is efficient enough to be practical for large language models. Further, the method allows the retention of much of the currently used software and hardware base. Our method for improving the functioning of large language models consists of a first stage of preprocessing that recodes the data to a Unary Relational Integracode that is more explicit about the relationships among the objects described in the text, followed as a second stage by a standard but possibly streamlined machine learning process that then also learns to predict these relationships. The method may be viewed as realizing a world model and applying beyond natural language, to vision and actions, for example, where the multiple properties of an object referred to in an input are brought together explicitly, rather than remaining distributed in the various references to it in the input. We articulate its advantages in terms of Robust Logic, a system for performing principled chaining on learned, and hence uncertain, information. We show that this recoding has the surprising and fortuitous property that, while succinct, it makes the task of learning a core subset of relational rules that hold in the world described in the training data polynomial time learnable in a defined sense, the polynomial depending on the complexity of the rule. This gives support for sound reasoning within each single call of the learned classifier as well as between multiple calls.

Large Language Models (LLMs) are being employed widely to automate tasks across the software development life-cycle. It is, however, unclear whether these tasks are performed consistently with respect to the semantics of the artefacts being handled. This question is particularly under-researched concerning architectural design specification. In this paper, we address this question for High-Level Message Sequence Charts (HMSCs). These are visual models with a rigorous formal semantics that have been used for various purposes, including as a foundation for Sequence Diagrams in the Unified Modelling Language (UML). We examine whether LLMs "understand" the semantics of HMSCs by examining three LLMs (Gemini-3, GPT-5.4, and Qwen-3.6) on how they perform 129 semantic tasks ranging from querying basic semantic constructs in HMSCs (i.e., events and their ordering) to semantic-preserving abstractions and compositions, and calculating the set of traces and trace-equivalent labelled transition systems. The results show that LLMs only have a modest understanding of the formal semantics of HMSCs (ca. 52% overall accuracy), with great variability across different semantic concepts: while LLMs seem to understand the basic semantic concepts of MSCs (ca. 88% accuracy), they struggle with semantic reasoning in tasks involving abstraction and composition (ca. 36% accuracy) and traces and LTSs (ca. 42% accuracy). In particular, all three LLMs struggle with the notions of co-region and explicit causal dependencies and never employed them in semantic-preserving transformations.

Validating training data for reasoning models typically requires expensive trial-and-error fine-tuning cycles. In this work, we investigate whether the utility of a reasoning dataset can be reliably predicted prior to training using intrinsic data metrics. We propose a suite of quantitative measures and evaluate their predictive power by fine-tuning 8B and 11B models on semantically distinct variants of a Polish reasoning dataset. Our analysis reveals that these intrinsic metrics demonstrate strong and significant correlations with downstream model performance. Crucially, we find that the predictors of utility are scale-dependent: smaller models rely on alignment-focused metrics to ensure precision, whereas larger models benefit from high redundancy, utilizing verbose traces to solve complex tasks. These findings establish a scale-aware framework for validating reasoning data, enabling practitioners to select effective training sets without the need for exhaustive empirical testing.

On-policy self-distillation trains a reasoning model on its own rollouts while a teacher, often the same model conditioned on privileged context, provides dense token-level supervision. Existing objectives typically weight the teacher's token-level signal uniformly across a chain-of-thought sequence, despite substantial variation in the entropy of the teacher's predictive distribution. We propose EGRSD (Entropy-Guided Reinforced Self-Distillation), which unifies token-level updates through three signals: a reward-grounded direction, a teacher-student likelihood-ratio magnitude, and the proposed teacher-entropy confidence gate that down-weights high-entropy token positions while maintaining a nonzero lower bound on every token weight. We further introduce CL-EGRSD, a causal-lookahead variant that distinguishes sustained high-entropy spans from transient high-entropy positions whose following context rapidly becomes low entropy. Experiments with Qwen3-4B and Qwen3-8B in thinking mode show that EGRSD and CL-EGRSD advance the accuracy-length frontier among the compared trainable methods.

Distilling multi-step reasoning abilities from large language models (LLMs) into compact student models remains challenging due to noisy rationales, hallucinated supervision, and static teacher-student interactions. Existing reasoning distillation methods, including mentor-based approaches, predominantly operate in an open-loop manner, implicitly assuming uniform teacher reliability and consequently propagating erroneous intermediate reasoning. We propose GateKD, a confidence-gated closed-loop distillation framework that enables robust reasoning transfer by treating the teacher as a dynamic gatekeeper rather than a static oracle. GateKD introduces three complementary mechanisms: (i) confidence-gated soft supervision that selectively distills reliable predictive signals, (ii) gated hidden-state evolution that aligns intermediate representations only when teacher confidence is high, and (iii) reliability-filtered attention distillation that preserves stable reasoning structures while suppressing noisy patterns. These components jointly form a closed feedback loop in which teacher confidence continuously modulates the distillation process, reducing hallucination transfer and stabilizing student reasoning. Extensive experiments across commonsense, logical, and symbolic reasoning benchmarks, using T5 and Flan-T5 backbones of varying sizes, demonstrate that GateKD consistently outperforms strong open-loop distillation baselines. Notably, GateKD yields substantial gains in logical and symbolic reasoning, remains robust under low-resource distillation settings, and shows clear performance degradation when any gating component is removed. Our results highlight that confidence-gated closed-loop supervision is critical for building reliable and scalable small reasoning models.

To effectively interact with the real world, Large Language Models (LLMs) require entity-based commonsense reasoning, a challenging task that necessitates integrating factual knowledge about specific entities with commonsense inference. Existing datasets for evaluating LLM entity-based commonsense reasoning have largely focused on True/False or multiple-choice questions, leaving the explicit assessment of the model's ability in abductive reasoning about causes and effects and generating explanations largely unexamined. In this work, we introduce CommonWhy, a dataset of 15,000 why questions designed to evaluate entity-based commonsense reasoning about causal relationships in LLMs. CommonWhy also serves as a Knowledge Graph Question Answering (KGQA) benchmark, as all supporting knowledge required to answer its queries is available in the Wikidata knowledge graph. Unlike existing KGQA datasets, which primarily test fact retrieval, CommonWhy targets causal commonsense reasoning, establishing a new paradigm for KGQA evaluation. Experiments with state-of-the-art LLMs and LLM-based KGQA methods reveal their significant shortcomings, including frequent factual hallucinations and failures in causal reasoning.

Large Language Models (LLMs) achieve strong performance on logical reasoning benchmarks, yet their reliability remains uncertain. Existing evaluations rely on static benchmarks, which fail to assess robustness under logically equivalent transformations and often overestimate reasoning capability. We propose LGMT (Logic-Grounded Metamorphic Testing), an oracle-free framework that leverages first-order logic (FOL) to evaluate LLM reasoning. By deriving metamorphic relations from formal logical equivalences, LGMT constructs semantically invariant test cases and detects reasoning defects through cross-case consistency checking. Experiments on six state-of-the-art LLMs show that LGMT exposes substantial hidden defects missed by traditional reference-based evaluations. We further find that models are particularly sensitive to symbol-level and conclusion-level variations, and that advanced prompting such as Few-shot CoT only partially mitigates these issues. These results suggest that LLM evaluation should move beyond isolated correctness toward robustness under logical invariance. LGMT provides a principled and scalable approach for diagnosing reasoning failures.

Looped Transformers offer a promising alternative to purely feed-forward computation by iteratively refining latent representations, improving language modeling and reasoning. Yet recurrent architectures remain unstable to train, costly to optimize and deploy, and constrained to small, fixed recurrence depths. We introduce Attractor Models, in which a backbone module first proposes output embeddings, then an attractor module refines them by solving for the fixed point, with gradients obtained through implicit differentiation. Thus, training memory remains constant in effective depth, and iterations are chosen adaptively by convergence. Empirically, Attractor Models outperform existing models across two regimes, large-scale language-model pretraining and reasoning with tiny models. In language modeling, Attractor Models deliver a Pareto improvement over standard Transformers and stable looped models across sizes, improving perplexity by up to 46.6% and downstream accuracy by up to 19.7% while reducing training cost. Notably, a 770M Attractor Model outperforms a 1.3B Transformer trained on twice as many tokens. On challenging reasoning tasks, we show that our model with only 27M parameters and approximately 1000 examples achieves 91.4% accuracy on Sudoku-Extreme and 93.1% on Maze-Hard, scaling favorably where frontier models like Claude and GPT o3, fail completely, and specialized recursive reasoners collapse at larger sizes. Lastly, we show that Attractor Models exhibit a novel phenomenon, which we call equilibrium internalization: fixed-point training places the model's initial output embedding near equilibrium, allowing the solver to be removed at inference time with little degradation. Together, these results suggest that Attractor Models make iterative refinement scalable by turning recurrence into a computation the model can learn to internalize.

We study on-policy self-distillation (OPSD), where a language model improves its reasoning ability by distilling privileged teacher distributions along its own on-policy trajectories. Despite its promise, OPSD can suffer from training instability due to a pattern mismatch between teacher and student responses. Self-reflected teacher responses may introduce reflection-induced biases and response templates that miscalibrate token-level supervision, ultimately harming the student's reasoning ability. To mitigate this issue, we propose OGLS-SD, an outcome-guided logit-steering framework that leverages verifiable outcome rewards to calibrate privileged teacher logits. Specifically, OGLS-SD contrasts teacher logits induced by successful and failed on-policy trajectories, constructing an outcome-discriminative steering direction for token-level guidance. Experiments on mathematical reasoning benchmarks show that OGLS-SD stabilizes self-distillation and improves performance over standard OPSD and other variants.

While Large Language Models (LLMs) are widely used in open-domain Question Answering (QA), their ability to handle inferential questions-where answers must be derived rather than directly retrieved-remains still underexplored. This study investigates how the structure and quality of passages influence LLM performance on such questions. We focus on convergence, a measure of how effectively sentences (hints) eliminate incorrect answers, as a criterion for constructing passages. Using subsets of the TriviaHG dataset, we form passages by combining sentences with varying convergence levels and evaluate six LLMs of different sizes and architectures. Our results show that passages built from higher convergence sentences lead to substantially better answer accuracy than those selected by cosine similarity, indicating that convergence captures meaningful relevance for inferential reasoning. Additionally, ordering sentences by descending convergence slightly improves performance, suggesting that LLMs tend to prioritize earlier, information-rich cues. These findings highlight convergence as a practical signal for guiding passage construction and analyzing inferential reasoning behavior in LLMs.

LLM-based conversational AI agents struggle to maintain coherent behavior over long horizons due to limited context. While RAG-based approaches are increasingly adopted to overcome this limitation by storing interactions in external memory modules and performing retrieval from them, their effectiveness in answering challenging questions (e.g., multi-hop, commonsense) ultimately depends on the agent's ability to reason over the retrieved information. However, existing methods typically retrieve memory based on semantic similarity to the raw user utterance, which lacks explicit reasoning about missing intermediate facts and often returns evidence that is irrelevant or insufficient for grounded reasoning. In this work, we introduce Goal-Mem, a goal-oriented reasoning framework for RAG-based agentic memory that performs explicit backward chaining from the user's utterance as a goal. Rather than progressively expanding from retrieved context, Goal-Mem decomposes each goal into atomic subgoals, performs targeted memory retrieval to satisfy each subgoal, and iteratively identifies what information from memory should be retrieved when intermediate goals cannot be resolved. We formalize this process in Natural Language Logic, a logical system that combines the verifiability of reasoning provided by FOL with the expressivity of natural language. Through extensive experiments on two datasets and comparing to nine strong memory baselines, we show that Goal-Mem consistently improves performance, particularly on tasks requiring multi-hop reasoning and implicit inference.

Recent soft prompt research has tried to improve reasoning by inserting trained vectors into LLM inputs, yet whether the gain comes from the learned content or from the act of injection itself has not been carefully separated. We study Random Soft Prompts (RSPs), which drop the training step entirely and append a freshly drawn sequence of random embedding vectors to the input. Each RSP vector is sampled from an isotropic Gaussian fitted to the entrywise mean and variance of the pretrained embedding table; the sequence carries no learned content, and yet reaches accuracy comparable to optimized soft prompts on math reasoning benchmarks in several settings. The mechanism unfolds in two stages: because attention has to absorb a never-seen-before random position, the distribution over the first few generated tokens flattens and reasoning trajectories branch, and as generation continues this influence dilutes naturally so the response commits to a single completion. We show that during inference RSPs lift early-stage token diversity and, combined with temperature sampling, widen Pass@N, the probability that at least one out of N attempts is correct. Beyond inference, we carry the same effect into DAPO training and demonstrate practical gains. Our contributions are: (i) RSP isolates the simplest form of soft prompt -- training-free, freshly resampled -- providing a unified lens for the structural effect of injection that variants otherwise differing in training and form all share; (ii) a theoretical and empirical validation of the underlying mechanism; and (iii) an extension from inference to training.

The reasoning frontier of Large Language Models (LLMs) has advanced significantly through modern post-training paradigms (e.g., Reinforcement Learning from Verifiable Rewards (RLVR)). However, the efficacy of these methods remains fundamentally constrained by the diversity and complexity of the training data. One practical solution is data synthesis; yet, prevalent methods relying on unstructured mutation or exploration suffer from homogeneity collapse, failing to systematically expand the reasoning frontier. To overcome this, we propose Evoutionary Task Discovery (EvoTD), a framework that treats data synthesis as a directed search over a dual-axis manifold of Algorithmic Skills and Complexity Attributes. We introduce structured evolutionary operators to navigate this space: a Crossover operator that synthesizes novel skill compositions to enhance diversity, and a Parametric Mutation operator that scales structural constraints (e.g., input size, tree depth) to drive robust generalization. Crucially, we integrate a dynamic Zone of Proximal Development filter, ensuring tasks lie within the learnable region of the model. Empirically, EvoTD delivers substantial reasoning gains that generalize consistently across model architectures, pretraining regimes, and scales, demonstrating that structured evolutionary curricula can effectively support reasoning improvement. We release our code on https://github.com/liqinye/EvoTD.

Fixed reasoning benchmarks evaluate canonical prompts, but semantically valid changes in presentation can still change model behavior. Studies of prompt variation can reveal such failures, but without audit they can mix genuine model errors with invalid perturbations, extraction artifacts, and unmatched search procedures. We propose an audit-constrained protocol for targeted reasoning evaluation. Prompt variants are generated from a finite component grammar, rendered deterministically, evaluated under a fixed query budget, and counted as model errors only after semantic and extraction audit. Within this protocol we instantiate Component-Adaptive Prompt Sampling (CAPS), a score-based sampler over prompt components, and compare it with equal-budget uniform component sampling under the same task bank, renderer, model interface, decoding settings, and audit procedure. Across three audited slices, the protocol identifies confirmed model-error prompt keys while excluding formatting and extraction artifacts, but matched comparisons do not show that CAPS improves audited yield or unique prompt-key discovery over uniform sampling. The contribution is methodological: targeted prompt variation can be studied under a reconstructable, reviewable, budget-matched protocol, and proxy-guided policies should be judged by audited yield rather than raw mismatch counts or selected examples alone.

Humans intuitively solve complex problems by flexibly shifting among reasoning modes: they plan, execute, revise intermediate goals, resolve ambiguity through associative judgment, and apply formal procedures to well-specified subproblems. Current LLM agents lack this flexibility, as their scaffolds hard-code such reasoning decisions in advance. These scaffolds are effective when their prescribed structure matches the task, but brittle when solving the task requires adapting the structure of reasoning itself. We introduce Deep Reasoning -- an inference-time approach for constructing task-specific scaffolds through structured meta-reasoning. Deep Reasoning uses a formal language that represents meta-reasoning as executable decompositions over associative inference, formal computation, and recursive subproblem solving, enabling decomposition principles to be encoded as in-context examples that guide test-time scaffold construction. We instantiate this approach in a general-purpose agent (DOLORES) that distributes complex tasks across more controlled reasoning threads. We evaluate it against state-of-the-art scaffolding methods across four hard benchmarks: multi-hop reasoning, long-chain question answering, long-context aggregation, and deep research-style information seeking. DOLORES outperforms all evaluated scaffolds across three model sizes and two model families, improving over the strongest evaluated scaffold baseline by 24.8% on average. DOLORES distributes cognition across structured, lower-load reasoning threads, thereby reducing premature termination and hallucinations. This advantage can even bridge the scaling gap, with an 8B version surpassing all evaluated 32B baselines from the same family in more than half the settings. These results point toward future agentic systems that treat scaffolding as adaptive reasoning, constructing the structure each task requires just-in-time.

Although Large Language Models (LLMs) have made remarkable progress, current preference optimization methods still struggle to align directional consistency while preserving reasoning diversity. To address this limitation, we propose Directional-Groupwise Preference Optimization (DGPO), a lightweight framework that aggregates supervision signals at the group level and explicitly models direction-aware alignment through multi-candidate comparisons. DGPO organizes forward and reverse question-answer instances into structured sets and optimizes a margin-based likelihood objective that separates coherent reasoning paths from inconsistent alternatives. This group-wise formulation captures richer relative information than pairwise objectives and reinforces consistency across diverse reasoning pathways. Empirical results show that our constructed reverse data yields a 3.2% average improvement across five benchmarks, while DGPO further delivers consistent gains across multiple datasets and model families, achieving average accuracy improvements of up to 3.6%.

Large Language Models (LLMs) show remarkable semantic understanding but often struggle with structural understanding when processing graph topologies in a serialized format. Existing solutions rely on training external graph-based adapters or fine-tuning, which incur high costs and lost generalizability. In this work, we investigate the internal mechanisms of LLMs and present a critical finding: LLMs spontaneously reconstruct the graph's topology internally, evidenced by a distinct "sawtooth" pattern in their attention maps that structurally aligns with the "token-level adjacency matrix". However, this intrinsic structural understanding is diluted by the attention sink. We theoretically formalize this dilution as a representation bottleneck, stemming from a fundamental conflict: the model's anisotropic bias, essential for language tasks, suppresses the topology-aware local aggregation required for graph reasoning. To address this, we propose a training-free solution, named StructuraL Attention SHarpening (SLASH), which amplifies this internal structural understanding via a plug-and-play attention redistribution. Experiments on pure graph tasks and molecular prediction validate that SLASH delivers significant and consistent performance gains across diverse LLMs.

Using Large Language Models (LLMs) to process graph-structured data is an active research area, yet current state-of-the-art approaches typically rely on multi-step pipelines with Graph Neural Network (GNN) encoders that compress rich textual attributes into solitary tokens, creating a significant semantic bottleneck. In this paper, we introduce the Graph Transformer Language Model (GTLM), a novel architecture that enables pretrained LLMs to natively process graph topologies while entirely eliminating this compressive bottleneck. GTLM is exceptionally parameter-efficient: by injecting graph-aware attention biases directly into the LLM's attention modules, it introduces only 0.015% additional parameters relative to the base model. We theoretically prove that our bidirectional attention prefix preserves node permutation equivariance while maintaining exact backward compatibility with the pretrained base model. Extensive evaluations demonstrate that a 1B-parameter GTLM matches or exceeds the performance of 7B-parameter state-of-the-art models on standard Text-Attributed Graph benchmarks, while significantly surpassing baselines on GraphQA. Finally, we demonstrate that GTLM attention heads implicitly learn to simulate message passing, explaining its superior performance on algorithmic tasks. This paradigm shift enables true algorithmic reasoning within LLMs and provides a scalable foundation for next-generation GraphRAG and relational deep learning.