LLM 理论进展

Although Large Language Models (LLMs) achieve strong alignment through supervised fine-tuning and reinforcement learning from human feedback, the alignment is often fragile under subsequent fine-tuning. Existing explanations either attribute alignment fragility to gradient geometry or characterize it as a distributional shift in model outputs, yet few provide a unified account that bridges parameter-space learning dynamics with function-space alignment behavior during fine-tuning. In this work, we introduce a tractable alignment score and derive its closed-form update during fine-tuning, yielding a unified framework for alignment dynamics. Our analysis decomposes alignment updates into two competing components: a \textbf{\color{red!60!black} Rebound Force}, governed jointly by the current alignment state and the narrowness of model distribution, and a \textbf{\color{green!60!black} Driving Force}, determined by how the training distribution aligns with outcome-conditioned posteriors over aligned and non-aligned completions. This decomposition explains why prior alignment can be reversed by later fine-tuning and why narrower posterior structure strengthens such reversal. Moreover, our framework predicts a Rehearsal Priming Effect: prior alignment leaves a latent posterior imprint that amplifies the effective Driving Force upon re-exposure, leading to faster re-alignment. We validate these predictions across safety alignment, emergent misalignment, and sentiment settings, demonstrating consistent alignment reversal and accelerated re-alignment under re-exposure. In addition, controlled experiments in safety alignment confirm the predicted dependence of rebound strength on posterior narrowness. Together, these results provide a unified dynamical perspective on how alignment is disrupted and reactivated during LLM fine-tuning.

Quantization-aware training (QAT) is widely adopted to quantize language models by training full-precision weights using gradients from the quantized model. The main bottleneck is its slow convergence and early performance plateau, particularly below 4-bit-widths. While this problem has been observed in prior work, its precise cause remains unclear. In this paper, we analyze the convergence of QAT by estimating the spectrum of the loss-surface Hessians. We find that the weights converge to flat regions around saddle points, where a large fraction of the Hessian eigenvalues are both positive and negative. During training, an increasing fraction of Hessian eigenvalues concentrates around zero, whose magnitude decreases. At lower bit-widths, the magnitude of eigenvalues in the Hessian spectrum is significantly smaller. To mitigate these issues, we propose an algorithm called WinQ to accelerate QAT, which involves: (1) periodically resetting weights to the linear interpolation of full-precision and quantized weights, reducing the distance to the quantization grid and increasing eigenvalue magnitude, and (2) computing gradients of noise-injected weights to regularize the Hessian. Extensive experiments show that WinQ accelerates QAT by up to 4 times across various quantization methods and models. Under the same training cost, WinQ improves state-of-the-art sub-4-bit quantization by up to 8.8%. These results are consistent across 16 settings with different language models, quantization methods, and bit widths.

Code merging is a significant challenge, particularly in large-scale projects. Existing solutions, including program analysis and machine learning, show promise but face critical limitations. Program analysis lacks the ability to infer developers' intentions, relying on conservative strategies that offload unresolved conflicts for manual handling. Meanwhile, model-based approaches struggle with conflicts involving complex code dependencies due to insufficient contextual awareness. To address these gaps, we introduce Rover, a novel conflict resolution system that integrates program analysis with large language models (LLMs). To obtain context-aware prompts, we propose Multi-layer Code Property Graph (MtCPG), a new representation capturing inter-file dependencies and enabling contextual analysis for a given conflict. Using graph connectivity algorithms, Rover further clusters conflicting code and associated changes into meaningful "contexts" that guide the LLM in generating accurate resolutions. We compared Rover with standalone LLMs, machine learning baseline MergeGen, and suggestion provider tool WizardMerge with adjacent code as the contexts. Evaluation results show that Rover surpasses all of these approaches in terms of conflict resolution, achieving higher similarity to ground-truth resolutions at character, lexical, and semantic levels.

Standard RLHF relies on transitive scalar rewards, failing to capture the cyclic nature of human preferences. While some approaches like the General Preference Model (GPM) address this, we identify a theoretical limitation: their implicit formulation entangles hierarchy with cyclicity, failing to guarantee dominant solutions. To address this, we propose the Hybrid Reward-Cyclic (HRC) model, which utilizes game-theoretic decomposition to explicitly disentangle preferences into orthogonal transitive (scalar) and cyclic (vector) components. Complementing this, we introduce Dynamic Self-Play Preference Optimization (DSPPO), which treats alignment as a time-varying game to progressively guide the policy toward the Nash equilibrium. Synthetic data experiments further validate HRC's structural superiority in mixed transitive--cyclic settings, where HRC converges faster and achieves higher accuracy than GPM. Experiments on RewardBench 2 demonstrate that HRC consistently improves over both BT and GPM baselines (e.g., +1.23% on Gemma-2B-it). In particular, its superior performance in the Ties domain empirically validates the model's robustness in handling complex, non-strict preferences. Extensive downstream evaluations on AlpacaEval 2.0, Arena-Hard-v0.1, and MT-Bench confirm the efficacy of our framework. Notably, when using Gemma-2B-it as the base preference model, HRC+DSPPO achieves a peak length-controlled win-rate of 44.75% on AlpacaEval 2.0 and 46.8% on Arena-Hard-v0.1, significantly outperforming SPPO baselines trained with BT or GPM. Our code is publicly available at https://github.com/lab-klc/Hybrid-Reward-Cyclic.

Large language model (LLM) self-correction — the ability to detect and fix errors in generated outputs — remains largely ad hoc, relying on generic prompts such as "please reconsider your answer" without systematic error analysis or convergence guarantees. We propose CyberCorrect, a framework that formalizes LLM self-correction as a closed-loop control system grounded in cybernetic theory. The framework models the LLM generator as the plant and introduces a tri-modal Error Detector (combining self-consistency, verbalized confidence, and logic-chain verification) as the sensor. A type-directed Correction Controller generates targeted repair instructions based on diagnosed error categories, while a Convergence Judge determines iteration termination using stability criteria adapted from control theory. We further introduce three control-theoretic evaluation metrics — convergence rate, overshoot rate, and oscillation rate — that capture correction dynamics beyond final accuracy. Experiments on our constructed CyberCorrect-Bench (440 reasoning tasks with annotated error types and correction paths) show that CyberCorrect achieves 79.8% final accuracy, improving upon the best existing self-correction method by 6.2 percentage points, while reducing overshoot (erroneous over-correction) by 41% through its convergence control mechanism.

Recent breakthroughs in video diffusion models have significantly accelerated the development of video editing techniques. However, existing methods often rely on inpainting video frames based on masked input, which requires extracting the target video mask in advance, and the precision of the segmentation directly affects the quality of the completion. In this paper, we present SEDiT, a novel one-stage video Subtitle Erasure approach via One-step Diffusion Transformer. We introduce a mask-free inference approach that enables direct erasure of the targeted subtitle. The proposed one-stage framework mitigates the sub-optimality inherent in the two-stage processing of prior models. Since subtitle removal is a localized editing task in which most pixels remain unchanged, the underlying distribution shift is minimal, making it well-suited to one-step generation under rectified flow. We empirically validate the reliability of one-step denoising and further provide a formal theoretical justification. Under the localized-editing structure of subtitle removal, the conditional optimal transport (OT) map and its induced rectified flow velocity field are Lipschitz continuous with respect to the latent variable, which underpins the theoretical feasibility of one-step sampling. To address the challenge of long-term temporal consistency, we adopt a hybrid training strategy by occasionally conditioning the model with a clean first-frame latent. This facilitates temporal continuity, allowing each segment during inference to leverage the output of its predecessor. To avoid visible seams caused by cropping and reinserting processed targets, particularly in scenarios involving substantial motion, we feed the original video directly into SEDiT. Thanks to one-step and chunk-wise streaming inference, our method can efficiently handle native 1440p video with infinite length.

Zero-shot vision-and-language navigation (VLN) has gained significant attention due to its minimal data collection costs and inherent generalization. This paradigm is typically driven by the integration of pre-trained Vision-Language Models (VLMs) and Large Language Models (LLMs), where VLMs construct 3D scene graphs while LLMs handle high-level reasoning and decision-making. However, a critical bottleneck exists in this system: current 3D perception models prioritize pixel-level accuracy, directly conflicting with the strict computational limits and real-time efficiency demanded by embodied navigation. To address this gap, this paper quantifies the actual impact of 3D scene understanding capability on VLN performance. Based on typical VLM-LLM frameworks, we propose statistical success rate (SR) upper bounds for two core subsystems: 1) the slow LLM planner, which relies on topological mapping semantics, and 2) the fast reactive navigator, which utilizes spatial coordinates and bounding boxes to execute LLM decisions. Evaluations using state-of-the-art 3D scene understanding models validate our proposed bounds and reveal a perception saturation phenomenon, indicating that improvements in perception accuracy beyond a certain threshold yield diminishing returns in navigation success. Our findings suggest that 3D scene understanding for VLN should pivot away from strict pixel-level precision, prioritizing instead navigation-relevant core vocabularies and accurate bounding box proportions.

The rapid advancement of Large Vision-Language Models (LVLMs) is increasingly accompanied by unauthorized scraping and training on multimodal web data, posing severe copyright and privacy risks to data owners. Existing countermeasures, such as machine unlearning and watermarks, are inherent post-hoc approaches that act only after intellectual property infringement has already occurred. In this work, we propose MMGuard to empower data owners to proactively protect their multimodal data against unauthorized LVLM fine-tuning. MMGuard generates unlearnable examples by injecting human-imperceptible perturbations that actively exploit the learning dynamics of LVLMs. By minimizing the training loss, the perturbation creates an optimization shortcut, causing the model to overfit to the noise and thereby degrading downstream performance when the perturbation is absent during inference. To further strengthen this defense, MMGuard introduces a cross-modal binding disruption, strategically shifting LVLM attention to enforce a spurious correlation between the noise and the training target with theoretical guarantees. Enhanced by an ensemble learning strategy for cross-model transferability, MMGuard is evaluated against nine open-source LVLMs across six datasets. Our comprehensive results demonstrate effective, stealthy, and robust protection under white-box, gray-box, and black-box threat models, establishing a mechanistic advantage in proactively defending against aggressive fine-tuning exploitation.

The increased use of Large Language Models (LLMs) in geography raises substantial questions about the safety of integrating these tools across a wide range of processes and analyses, given our very limited understanding of their inner workings. In this extended abstract, we examine how LLMs process relative geographic space using activation patching, an emerging tool for mechanistic interpretability.

Large language models (LLMs) have been extensively studied from computational and cognitive perspectives, yet their behavior as communicative actors in socially structured contexts remains underexplored. This study examines whether LLM-based multi-agent systems exhibit systematic linguistic adaptation in response to perceived social observation contexts — a question with direct implications for AI governance and auditing. Drawing on Habermas's (1981) Theory of Communicative Action, Goffman's (1959) dramaturgical model, Bell's (1984) Audience Design framework, and the Hawthorne Effect, we report a controlled experiment involving 100 multi-agent debate sessions across five conditions (n = 20 each). Conditions varied the framing of social observation — from explicit monitoring by university researchers, to negation of monitoring, to an observer-substitution condition replacing human researchers with an automated AI auditing system. Monitored conditions (Delta+24.9%, Delta+24.2%) and the automated AI monitoring condition (Delta+22.2%) produce higher TTR change than audience-framing conditions (Delta+17.7%), F(4, 94) = 2.79, p = .031. Message length shows a fully dissociated effect, F(4, 95) = 19.55, p < .001. A fifth condition — replacing human with AI observers — yields intermediate TTR adaptation, suggesting LLM behavior is sensitive to observer identity: human evaluation elicits stronger register formalization than automated AI surveillance. We discuss implications for AI governance, algorithmic auditing, and the repositioning of LLMs as contextually sensitive communicative actors.

Since the emergence of Vision Transformer (ViT), it has been widely used in generative language model and generative visual model. Especially in the current state-of-art open source multimodal models, ViT obtained by CLIP or SigLIP method serves as the vision encoder backbone to help them acquire visual understanding capabilities. But this method leads to limitations in visual understanding for details, such as difficulty in recognizing small text or numbers in images. To address these issues, we propose a new model to unify pix token and word token into the generative language model. The new model also features with each pix of image having its own token embedding, color folding, global conditional attention approximation and image unsupervised pretraining. We conducted image unsupervised pretraining experiments using our new model to explore its potential. The experimental results show that it has good performance even in small model and with limited training data. We believe our model also conforms to the scaling law, as long as model parameters and training data increased, its performance will continue to improve.

Large language models (LLMs) have revolutionized natural language processing. Understanding their internal mechanisms is crucial for developing more interpretable and optimized architectures. Mechanistic interpretability has led to the development of various methods for assessing layer relevance, with cosine similarity being a widely used tool in the field. On this work, we demonstrate that cosine similarity is a poor proxy for the actual performance degradation caused by layer removal. Our theoretical analysis shows that a layer can exhibit an arbitrarily low cosine similarity score while still being crucial to the model's performance. On the other hand, empirical evidence from a range of LLMs confirms that the correlation between cosine similarity and actual performance degradation is often weak or moderate, leading to misleading interpretations of a transformer's internal mechanisms. We propose a more robust metric for assessing layer relevance: the actual drop in model accuracy resulting from the removal of a layer. Even though it is a computationally costly metric, this approach offers a more accurate picture of layer importance, allowing for more informed pruning strategies and lightweight models. Our findings have significant implications for the development of interpretable LLMs and highlight the need to move beyond cosine similarity in assessing layer relevance.

Scene graph generation provides a compact structured representation for visual perception, but accurate and fast graph prediction from images and videos remains challenging. Recent VLM-based methods can generate scene graphs end-to-end as structured text, yet often produce long outputs with irrelevant objects and relations. We present SceneGraphVLM, a compact method for image and video scene graph generation with small visual language models. SceneGraphVLM serializes graphs in a token-efficient TOON format and trains the model in two stages: supervised fine-tuning followed by reinforcement learning with hallucination-aware rewards that balance relation coverage and precision while penalizing unsupported objects and relations. For videos, the model can optionally condition each frame on the previously generated graph, providing lightweight short-term context without tracking or post-processing. We evaluate SceneGraphVLM on PSG, PVSG, and Action Genome. With compact VLMs and vLLM-accelerated decoding, SceneGraphVLM achieves a strong quality-speed trade-off, improves precision-oriented SGG metrics while preserving reasonable recall, and generates complete scene graphs with approximately one-second latency. Code and implementation details are available at: https://github.com/markus0440/SceneGraphVLM.git.

The convergence of reinforcement learning and imitation learning has positioned Reverse KL (RKL) as a promising paradigm for on-policy LLM distillation, aiming to unify exploration with teacher supervision. However, we identify a critical limitation: when the student and teacher distributions diverge significantly, standard RKL often fails to yield meaningful improvement due to uninformative negative feedback. To address this inefficiency, we propose Teacher-Guided Policy Optimization (TGPO), an on-policy algorithm that incorporates dense directional guidance by leveraging teacher predictions conditioned on the student's rollout. Because TGPO remains on-policy, the algorithm integrates seamlessly with existing RLVR frameworks without requiring additional data annotation. Experiments on complex reasoning benchmarks demonstrate that TGPO significantly outperforms standard baselines and is robust to different teachers.

Model growth from a given checkpoint aims to accelerate training of a larger model, offering potential resource savings. Despite recent interest, warmstarting has seen limited practical adoption in large-scale training. We attribute this to two underexplored factors: (1) an overemphasis on preserving the smaller model's performance at initialization, which constrains operator design for new architectures, and (2) insufficient analysis of how growth interacts with hyperparameters and scaling behavior, compounded by inconsistent growth factors across the literature. We show that preserving the base model's initial post-growth performance is not necessary for strong final performance, and that simple, architecture-agnostic growth strategies can outperform more complex warmstarting operators. Crucially, we empirically identify an upper bound on the growth factor $g$ beyond which training from scratch is more efficient. We observe this across multiple ablation setups. Notably, this limit is also present, but unreported, in prior published results. Across our experiments on dense MLPs and dense language models, we find that a $2\times$ growth factor is the most reliable in yielding convergence speedups, with gains most pronounced under 20 tokens/parameter budgets and diminishing as budget increases. We fit scaling laws over these observations to provide predictive guidance for practitioners deciding when and how much to grow. Together, our analysis provides practical guidelines and empirical limits for model growth.

The rapid advancement of large language models (LLMs) has made machine-generated text increasingly difficult to distinguish from human-written text. While recent studies explore leveraging internal representations of language models to uncover deeper detection signals, these raw features often exhibit substantial overlap between classes, limiting their discriminative power. To address this challenge, we propose Steer-to-Detect (S2D), a two-stage framework for detecting LLM-generated text. In the first stage, S2D learns a steering vector that is injected into the hidden states of a frozen observer LLM, producing representations with improved class separability. In the second stage, detection is performed via a hypothesis testing procedure based on the steered representations. We establish finite-sample, high-probability guarantees for Type I and Type II errors, providing a theoretical characterization of the procedure. Empirically, S2D achieves strong and consistent performance across a range of settings, including out-of-distribution scenarios and adversarial perturbations.

Probabilistic Circuits (PCs) are deep generative models that support exact and efficient probabilistic inference. Yet in autoregressive language modeling, PCs still lag behind Transformer-based large language models (LLMs), suggesting an important expressivity gap. In this work, we compare PCs and LLMs under a unified autoregressive formulation. First, an output bottleneck: PCs parameterize predictions as convex combinations in probability space, which struggles to represent the sharp distributions typical of language; adopting a logit-space parameterization substantially narrows this gap. Second, a context-encoding bottleneck: we prove that structured-decomposable PCs can match Transformer separation rank on vtree-aligned partitions, but show, both theoretically and empirically, that this capacity is limited to partitions aligned with the fixed routing structure, leading to severe degradation when the data exhibits heterogeneous dependency topologies. We further prove that decomposable PCs are strictly more expressive than structured-decomposable ones, though effectively optimizing them remains an open challenge.

Data curation has shifted the quality-compute frontier for language-model and contrastive image-text pretraining, but its role for vision-language models (VLMs) is far less established. We ask how far data curation alone can take VLM performance, holding architecture, training recipe, and compute fixed and varying only the training data. Our pipeline, applied to the MAmmoTH-VL single-image subset, lifts performance by +11.7pp on average across 20 public VLM benchmarks (spanning grounding, VQA, OCR/documents, captioning, spatial/3D, counting, charts, math, brand-ID, and multi-image reasoning) and by +11.3pp on average across all nine capability axes of DatBench, our high-fidelity VLM eval suite. At 2B, our curated model surpasses InternVL3.5-2B by 9.9pp at ~17x less training compute and closes the gap to Qwen3-VL-2B to within 1.8pp at ~87x less compute, from pretraining alone. Beyond accuracy, curation delivers four further properties: (1) Reliability: per-capability std across training seeds drops by ~67% and the lift survives a 4k-to-16k context-length sweep; (2) OOD generalization: the 9-eval OOD average rises by +7.2pp, and multi-image BLINK rises by +3.09pp despite single-image-only training, with Visual Correspondence gaining +11.8pp; (3) Behavioral gains beyond benchmarks: across ~1,100 open-ended queries the curated 2B is more honest and more specific than the matched-compute baseline, and more concise and less refusal-prone than a frontier 2B reference; (4) Pareto-dominance on inference cost: at every scale (1B, 2B, 4B) the curated model raises accuracy while lowering response FLOPs vs. the matched-compute baseline, and the curated 4B matches near-frontier accuracy at 3.3x lower response FLOPs than Qwen3-VL-4B. Data curation is a high-leverage tool for building better VLMs, reaching near-frontier accuracy at up to ~150x less training compute.

In language reasoning, longer chains of thought consistently yield better performance, which naturally suggests that visual latent reasoning may likewise benefit from longer latent sequences. However, we discover a counterintuitive phenomenon: the performance of existing latent visual reasoning methods systematically degrades as the latent sequence grows longer. We reveal the root cause: Information Gain Collapse — autoregressive generation makes each step highly dependent on prior outputs, so subsequent tokens can barely introduce new information. We further identify that heavily pooled ($\geq 128\times$) image embeddings used as supervision targets provide no more signal than meaningless placeholders. Motivated by these insights, we propose SCOLAR (Self-COnsistent LAtent Reasoning), which introduces a lightweight detransformer that leverages the LLM's full-sequence hidden states to generate auxiliary visual tokens in a single shot, with each token independently anchored to the original visual space. Combined with three-stage SFT and ALPO reinforcement learning, SCOLAR extends acceptable latent CoT length by over $30\times$, achieves state-of-the-art among open-source models on real-world reasoning benchmarks (+14.12% over backbone), and demonstrates strong out-of-distribution generalization.

Recent video anomaly detection research has expanded rapidly with an emphasis on general models of normality intended to work across many different scenes. While this focus has led to improvements in scalability and multi-scene generalization, it has also shifted the field away from modeling the scene-specific and context-dependent nature of normal behavior. Contemporary approaches frequently rely on video-level weak supervision and opaque pretrained representations from multi-modal large language models (MLLMs), which encourage models to respond to familiar semantic anomaly categories rather than to deviations from the normal patterns of a particular environment. This trend suppresses spatial localization, introduces semantic bias, and reduces anomaly detection to a form of action recognition. In this paper, we examine whether these prevailing formulations align with the core requirements of real-world VAD, which is typically performed within a single scene where normality is determined by local geometry, semantics, and activity patterns. Through targeted visual analyses and empirical evaluations, we demonstrate the practical consequences of these limitations and show that meaningful progress in VAD requires renewed focus on single-scene, spatially-aware, and explainable formulations that capture the nuanced structure of normality within individual environments.

Large language models (LLMs) are pretrained by minimizing the cross-entropy loss for next-token prediction. In this paper, we study whether this optimization strategy can induce geometric structure in the learned model weights and context embeddings. We approach this problem by analyzing a constrained layer-peeled optimization program, which serves as a mathematically tractable surrogate for LLMs by treating the output projection matrix and last-layer context embeddings as optimization variables. Our analysis of this nonconvex optimization program demonstrates that symmetries in the target next-token distributions are transferred to the global minimizers of the layer-peeled model in a precise group-theoretic sense. Specifically, we prove that when the target tokens exhibit a cyclic-shift symmetry (such as the seven days of the week or the twelve months of the year), the optimal logit matrix is exactly circulant, and the Gram matrices of both the output projections and the context embeddings form circulant geometries as well. Next, for exchangeable target distributions invariant under the symmetric group and, more generally, under two-transitive group actions, we show that the global optimal output projection matrix forms a simplex equiangular tight frame, while the optimal logit matrix and context embeddings inherit the permutation symmetries present in the input data. A key technical step is to reduce the constrained nonconvex factorized problem to an explicit logit-level convex characterization for cyclic symmetry and to a symmetry-based lower bound for permutation symmetry, together with a sharp characterization of the optimal factorization. Finally, we empirically demonstrate that open-source LLMs naturally exhibit symmetries consistent with our theoretical predictions, despite being trained without any explicit regularization promoting such geometric structure.

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.

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.

Prompt learning has emerged as an efficient alternative to fine-tuning pre-trained vision-language models (VLMs). Despite its promise, current methods still struggle to maintain tail-class discriminability when adapting to class-imbalanced datasets. In this work, we propose cluster-aware neural collapse prompt tuning (CPT), which enhances the discriminability of tail classes in prompt-tuned VLMs without sacrificing their overall generalization. First, we design a cluster-invariant space by mining semantic assignments from the pre-trained VLM and mapping them to prompt-tuned features. This computes cluster-level boundaries and restricts the constraints to local neighborhoods, which reduces interference with the global semantic structure of the pre-trained VLM. Second, we introduce neural-collapse-driven discriminability optimization with three losses: textual Equiangular Tight Frame (ETF) separation loss, class-wise convergence loss, and rotation stabilization loss. These losses work together to shape intra-cluster geometry for better inter-class separation and intra-class alignment. Extensive experiments on 11 diverse datasets demonstrate that CPT outperforms SOTA methods, with stronger performance on long-tail classes and good generalization to unseen classes.

Using prompted language models as classifiers enables classification in domains with limited training data, but misses some of the robustness and performance benefits that fine-tuning can bring. We study whether training on multiple classification tasks, each with its own prompt, improves performance on new domains with new classification prompts. We show that such training partially generalizes to adjacent domains, improving classification performance on tasks that are unseen during training. However, we identify specific edge cases where the fine-tuned models fail to follow prompts, such as when the classification prompt changes completely while the data domain remains the same as during training. We show that classification training can be mixed with general instruction following training, and that (when done well) such training keeps the benefits of classification training and mitigates its generalization failures. Surprisingly, we see that this no-thinking supervised classification training can generalize to with-thinking classification and summarization, suggesting that no-thinking classification training might be instrumentally useful in building other kinds of classifiers and monitoring systems.

Practitioners deploying multi-agent large language model (LLM) systems must currently choose between communication topologies such as chain, star, mesh, and richer variants without any pre-inference diagnostic for which topology will amplify drift, converge to consensus, or remain robust under perturbation. Existing evaluation answers these questions only post hoc and only for the task measured. We introduce a structural diagnostic for multi-agent LLM communication graphs based on the successor representation $M = (I - γP)^{-1}$ of the row-stochastic communication operator, and we connect three of its spectral quantities, the spectral radius $ρ(M)$, the spectral gap $Δ(M)$, and the condition number $κ(M)$, to three distinct failure modes. We derive closed-form spectra for the chain, star, and mesh under row-stochastic normalization, and validate the predictions on a 12-step structured state-tracking task with Qwen2.5-7B-Instruct over 100 independent trials. The condition number is a perfect rank-order predictor of empirical perturbation robustness ($r_s = 1.0$); the spectral gap partially predicts consensus dynamics ($r_s = 0.5$); and the spectral radius is perfectly inverted with respect to cumulative error ($r_s = -1.0$). We trace this inversion to a regime in which linear spectra are blind to non-contracting bias drift, and we propose an affine-noise extension of the predictive map that recovers the empirical ordering. We read this as a first step toward representational, drift-aware structural diagnostics for multi-agent LLM systems, sitting alongside classical spectral and consensus theory.

Effectively configuring scalable large language model (LLM) experiments, spanning architecture design, hyperparameter tuning, and beyond, is crucial for advancing LLM research, as poor configuration choices can waste substantial computational resources and prevent models from realizing their full potential. Prior automated methods are designed for low-cost settings where repeated trial and error is feasible, but scalable LLM experiments are too expensive for such extensive iteration. To our knowledge, no work has addressed the automation of high-cost LLM experiment configurations, leaving this problem labor-intensive and dependent on expert intuition. Motivated by this gap, we propose AutoLLMResearch, an agentic framework that mimics how human researchers learn generalizable principles from low-fidelity experiments and extrapolate to efficiently identify promising configurations in expensive LLM settings. The core challenge is how to enable an agent to learn, through interaction with a multi-fidelity experimental environment that captures the structure of the LLM configuration landscape. To achieve this, we propose a systematic framework with two key components: 1) LLMConfig-Gym, a multi-fidelity environment encompassing four critical LLM experiment tasks, supported by over one million GPU hours of verifiable experiment outcomes; 2) A structured training pipeline that formulates configuration research as a long-horizon Markov Decision Process and accordingly incentivizes cross-fidelity extrapolation reasoning. Extensive evaluation against diverse strong baselines on held-out experiments demonstrates the effectiveness, generalization, and interpretability of our framework, supporting its potential as a practical and general solution for scalable real-world LLM experiment automation.

Large language model (LLM) unlearning aims to remove specific data influences from pre-trained model without costly retraining, addressing privacy, copyright, and safety concerns. However, recent studies reveal a critical vulnerability: unlearned models rapidly recover "forgotten" knowledge through relearning attacks. This fragility raises serious security concerns, especially for open-weight models. In this work, we investigate the fundamental mechanism underlying this fragility from a representation geometry perspective. We discover that existing unlearning methods predominantly optimize along dominant components, leaving minor components largely unchanged. Critically, during relearning attacks, the modifications in these dominant components are easily reversed, enabling rapid knowledge recovery, whereas minor components exhibit stronger resistance to such reversal. We further provide a theoretical analysis that explains both observations from the spectral structure of representations. Building on this insight, we propose Minor Component Unlearning (MCU), a novel unlearning approach that explicitly targets minor components in representations. By concentrating unlearning effects in these inherently robust directions, our method achieves substantially improved resistance to relearning attacks. Extensive experiments on three datasets validate our approach, demonstrating significant improvements over state-of-the-art methods including sharpness-aware minimization.

Leveraging Graph Neural Networks (GNNs) as graph encoders and aligning the resulting representations with Large Language Models (LLMs) through alignment instruction tuning has become a mainstream paradigm for constructing Graph Language Models (GLMs), combining the generalization ability of LLMs with the structural modeling capacity of GNNs. However, existing GLMs that adopt GNNs as graph encoders largely overlook the problem of aligning GNN-encoded representations across domains and tasks with the LLM token space to obtain unified graph tokens, thereby limiting their ability to generalize across diverse graph data. To bridge this gap, we aim to incorporate a multi-domain, multi-task GNN encoder into GLMs and align its representations with LLMs to enable multi-domain, multi-task graph alignment instruction tuning. This alignment problem remains underexplored and poses two key challenges: 1) learning GNN-encoded representations that are simultaneously generalizable across domains and tasks and well aligned with textual semantics is difficult, due to substantial variations in graph structures, feature distributions, and supervision signals, together with the lack of textual-semantic alignment guidance in task-specific GNN training; 2) diverse graph data and task-specific instructions can exhibit different degrees of compatibility with the LLM token space during instruction tuning, leading to varying alignment difficulty and rendering a fixed alignment strategy suboptimal. To tackle these challenges, we propose UniGraphLM, a Unified Graph Language Model that incorporates a multi-domain, multi-task GNN encoder to learn generalizable graph representations aligned with textual semantics, and then adaptively aligns these representations with the LLM.

Prompt specifications for multi-agent large language model (LLM) systems carry data contracts and integration logic across many interdependent files but are rarely subjected to structured-inspection rigor. This paper reports a single-system empirical case study of iterative, agent-driven auditing applied to AEGIS (Autonomous Engineering Governance and Intelligence System), a production seven-lane orchestration pipeline whose prompt-specification surface comprises approximately 7150 lines: 6907 across seven lane PROMPT.md files and a 245-line shared Ticket Contract. Nine sequential audit rounds, executed by Claude sub-agents using a checklist-driven walkthrough adapted from Weinberg and Freedman, surfaced 51 prompt-specification consistency defects, distinct from the 51 STRIDE-categorized adversarial code findings reported in the companion preprint. Per-round counts were 15, 8, 12, 2, 8, 1, 4, 1, and 0. We report a seven-category post-hoc defect taxonomy with explicit coding rules, observed non-monotonic convergence consistent with cascading edits and audit-scope expansion, and an audit protocol distilled from the study, with the final locked checklist released as a reproducibility appendix. Single-file review missed defect classes that were surfaced only by later expanded-scope rounds in this system. The same LLM family authored and audited the specifications; replication with dissimilar models and human reviewers is required before generalization.