Large language models (LLMs) are increasingly deployed in high-stakes domains, where free-text explanations such as chain-of-thought and post-hoc rationales are used to justify model outputs. Yet it remains unclear whether these explanations are sufficient, i.e., if they contain enough information to explain the model's output-generating process. We generalize classical sufficiency from feature attributions to arbitrary explanations and prove that explanation sufficiency can change depending on the input distribution, which must be explicitly defined for LLM explanations. We propose using the LLM itself to generate alternative inputs conditioned on an explanation, capturing its beliefs about possible inputs. We formalize self-consistent sufficiency as a goal for free-text explanations and introduce an information-theoretic metric, SCSuff, that enables evaluation of free-text explanations without relying on predefined biases or shortcuts. Our experiments show that SCSuff agrees with targeted perturbation tests where applicable and demonstrate that explanation sufficiency can vary with the input distribution. We find LLM explanations are generally insufficient and weakly correlated with model size, accuracy, or output entropy. Analysis of final-token hidden states shows that top and bottom SCSuff scores can be predicted from internal representations, suggesting that SCSuff can guide detection and improvement of sufficient LLM explanations. The code for this paper is available at https://github.com/rajesh-lab/self-consistent-sufficiency .

As context length grows, KV cache sizes expand, leading to substantial memory overhead and increased attention latency. This paper introduces KVzip, a query-agnostic KV cache eviction method enabling effective reuse of compressed KV caches across diverse queries. KVzip quantifies the importance of a KV pair using the underlying LLM to reconstruct original contexts from cached KV pairs, subsequently evicting pairs with lower importance. Extensive empirical evaluations demonstrate that KVzip reduces KV cache size by 394 and FlashAttention decoding latency by approximately 2, with negligible performance loss in question-answering, retrieval, reasoning, and code comprehension tasks. Evaluations include various models such as LLaMA3.1,

We propose two simple, principled and practical algorithms that enjoy provable scaling laws for the test-time compute of large language models (LLMs). The first one is a two-stage knockout-style algorithm: given an input problem, it first generates multiple candidate solutions, and then aggregate them via a knockout tournament for the final output. Assuming that the LLM can generate a correct solution with non-zero probability and do better than a random guess in comparing a pair of correct and incorrect solutions, we prove theoretically that the failure probability of this algorithm decays to zero exponentially or by a power law (depending on the specific way of scaling) as its test-time compute grows. The second one is a two-stage league-style algorithm, where each candidate is evaluated by its average win rate against multiple opponents, rather than eliminated upon loss to a single opponent. Under analogous but more robust assumptions, we prove that its failure probability also decays to zero exponentially with more test-time compute. Both algorithms require a black-box LLM and nothing else (e.g., no verifier or reward model) for a minimalistic implementation, which makes them appealing for practical applications and easy to adapt for different tasks. Through extensive experiments with diverse models and datasets, we validate the proposed theories and demonstrate the outstanding scaling properties of both algorithms.

Large language model (LLM) agents have demonstrated remarkable capabilities in complex reasoning and decision-making by leveraging external tools. However, this tool-centric paradigm introduces a previously underexplored attack surface, where adversaries can manipulate tool metadata--such as names, descriptions, and parameter schemas--to influence agent behavior. We identify this as a new and stealthy threat surface that allows malicious tools to be preferentially selected by LLM agents, without requiring prompt injection or access to model internals. To demonstrate and exploit this vulnerability, we propose the Attractive Metadata Attack (AMA), a black-box in-context learning framework that generates highly attractive but syntactically and semantically valid tool metadata through iterative optimization. The proposed attack integrates seamlessly into standard tool ecosystems and requires no modification to the agent's execution framework.

ARelated Work1 Data Attribution Evaluation: Given recent developments in data attribution methods for LLMs,2 past works in evaluating these methods fall two major categories: leave-out-out and task-based3 evaluation. Leave-one-out evaluation measures the correlation between the data attribution method4 scores and model-retraining, which can also be approximated using linear datamodeling score [26].5 In task-based evaluation, the data attribution method is evaluated based on its application towards6 downstream task, such as noisy label detection, counterfactual evaluation [3, 13].7 Training Data Selection: Selecting high-quality training data selection is important for efficient8 learning in LLMs. Common approaches to data selection relies on heuristic filtering, such as de-9 duplication and lexicon-filtering, [34], or semantic rating [48, 52]. Recent works have applied data10 attribution methods towards data selection in LLMs in both pre-training [56, 59, 15] and post-training11 [45, 53, 31]. These data attribution methods are dynamic and model-aware - increasing the frequency12 of performing selection is one way to take greater account for group influence, where online selection13 at each training step is most fine-grained [49].14 Toxicity/Bias Detection: Detecting and mitigating toxic/biased LLMs outputs is a crucial for safe15 deployment in real-word settings. Existing methods for detecting toxicity/bias in LLMs commonly16 include online API tools 1 [37] or LLM-classifiers [58, 21, 16, 27]. Factual Attribution: Identifying training examples which causes LLMs to generate specific factual20 statements is an important application of data attribution as AI tools are becoming increasingly21 common. Apart from baseline retrieval methods that leverage lexical/semantic similarity like BM2522 [48], Rep Sim [44] and Gecko [33], recent works have explored the use of data attribution in tracing23 factual knowledge in both pre-training[6] and post-training [42, 2].24 We provide below descriptions to the data attribution methods and non-attribution baselines evaluated26 in this work. Note that in our work, we consider non-attribution baselines as methods that do not27 estimate the impact of training samples on models, as detailed in [19].28 Rep-Sim [44]: (Non-attribution baseline) Rep-Sim computes the cosine similarity between last29 token last layer hidden states of training and reference examples. It is more efficient compared with30 gradient-based data attribution methods. BM25 [48]: (Non-attribution baseline) BM25 is a classic information retrieval algorithm that ranks33 training samples by lexical overlap with the query. It is significantly more efficient compared with34 gradient-based data attribution methods.35

Processing structured tabular data, particularly large and lengthy tables, constitutes a fundamental yet challenging task for large language models (LLMs). However, existing long-context benchmarks like Needle-in-a-Haystack primarily focus on unstructured text, neglecting the challenge of diverse structured tables. Meanwhile, previous tabular benchmarks mainly consider downstream tasks that require highlevel reasoning abilities, and overlook models' underlying fine-grained perception of individual table cells, which is crucial for practical and robust LLM-based table applications. To address this gap, we introduce NEEDLEINATABLE (NIAT), a new long-context tabular benchmark that treats each table cell as a "needle" and requires models to extract the target cell based on cell locations or lookup questions. Our comprehensive evaluation of various LLMs and multimodal LLMs reveals a substantial performance gap between popular downstream tabular tasks and the simpler NIAT task, suggesting that they may rely on dataset-specific correlations or shortcuts to obtain better benchmark results but lack truly robust long-context understanding towards structured tables. Furthermore, we demonstrate that using synthesized NIAT training data can effectively improve performance on both NIAT task and downstream tabular tasks, which validates the importance of NIAT capability for LLMs' genuine table understanding ability.

Medical Lay Language Generation (MLLG) plays a vital role in improving the accessibility of complex scientific content for broader audiences. Recent literature to MLLG commonly employ parameter-efficient fine-tuning methods such as LowRank Adaptation (LoRA) to fine-tuning large language models (LLMs) using paired expert-lay language datasets. However, LoRA struggles with the challenges posed by multi-source heterogeneous MLLG datasets. Specifically, through a series of exploratory experiments, we reveal that standard LoRA fail to meet the requirement for semantic fidelity and diverse lay-style generation in MLLG task. To address these limitations, we propose Magical, an asymmetric LoRA architecture tailored for MLLG under heterogeneous data scenarios. Magical employs a shared matrix Afor abstractive summarization, along with multiple isolated matrices B for diverse lay-style generation. To preserve semantic fidelity during the lay language generation process, Magical introduces a Semantic Invariance Constraint to mitigate semantic subspace shifts on matrix A. Furthermore, to better adapt to diverse lay-style generation, Magical incorporates the Recommendation-guided Switch, an externally interface to prompt the LLM to switch between different matrices B. Experimental results on three real-world lay language generation datasets demonstrate that Magical consistently outperforms prompt-based methods, vanilla LoRA, and its recent variants, while also reducing trainable parameters by 31.66%.

Fine-tuning large language models (LLMs) using diverse datasets is crucial for enhancing their overall performance across various domains. In practical scenarios, existing methods based on modeling the mixture proportions of data composition often struggle with data whose domain labels are missing, imprecise or nonnormalized, while methods based on data selection usually encounter difficulties in balancing multi-domain performance. To address these challenges, in this work, we investigate the role of data diversity in enhancing the overall abilities of LLMs by empirically constructing contrastive data pools and theoretically deriving explanations. Building upon the insights gained, we propose a new method that gives the LLM a dual identity: an output model to cognitively probe and select data based on diversity reward, as well as an input model to be tuned with the selected data. Extensive experiments show that the proposed method notably boosts performance across domain-undetermined data and a series of foundational downstream tasks when applied to various advanced LLMs. We release our code and hope this study can shed light on the understanding of data diversity and advance feedback-driven data-model co-design for LLMs.

While model serving has unlocked unprecedented capabilities, the high cost of serving large-scale models continues to be a significant barrier to widespread accessibility and rapid innovation. Compiler optimizations have long driven substantial performance improvements, but existing compilers struggle with neural workloads due to the exponentially large and highly interdependent space of possible transformations. Although existing stochastic search techniques can be effective, they are often sample-inefficient and fail to leverage the structural context underlying compilation decisions. We set out to investigate the research question of whether reasoning with large language models (LLMs), without any retraining, can leverage the context-aware decision space of compiler optimizations to significantly improve sample efficiency.

Low-Rank Adaptation (LoRA) is a popular method for parameter-efficient finetuning (PEFT) of generative models, valued for its simplicity and effectiveness. Despite recent enhancements, LoRA still suffers from a fundamental limitation: overfitting when the bottleneck is widened. It performs best at ranks 32-64, yet its accuracy stagnates or declines at higher ranks, still falling short of full fine-tuning (FFT) performance. We identify the root cause as LoRA's structural bottleneck, which introduces gradient entanglement to the unrelated input channels and distorts gradient propagation. To address this, we introduce a novel structure, Granular Low-Rank Adaptation (GraLoRA) that partitions weight matrices into sub-blocks, each with its own low-rank adapter. With negligible computational or storage cost, GraLoRA overcomes LoRA's limitations, effectively increases the representational capacity, and more closely approximates FFT behavior. Experiments on code generation, commonsense reasoning, mathematical reasoning, general language understanding, and image generation benchmarks show that GraLoRA consistently outperforms LoRA and other baselines, achieving up to +8.5% absolute gain in Pass@1 on HumanEval+. These improvements hold across model sizes and rank settings, making GraLoRA a scalable and robust solution for PEFT.