This work introduces BlockRank, a method designed to make LLMs efficient and scalable for in-context ranking by aligning their internal attention and inference mechanisms with the structure of the in-context ranking task.

We propose two techniques to improve the data efficiency of LLM RL fine-tuning: difficulty-targeted online data selection and rollout replay.

We introduce the "Deadlock Attack," which uses an adversarial embedding to force LLMs into perpetual thinking loops, leading to resource exhaustion.

We propose an Interactive Text-to-3D Scene Retrieval Method to handle inherent query limitations.

We present a vision-centric token compression in LLM, inspired by human selective reading strategy.

This paper introduce the LLM-PySC2 environment, an LLM decision-making environment with complete pysc2 action space and multi-agent system.

This study introduces a behavioral game-theoretic framework to evaluate LLMs' strategic reasoning, revealing that models' reasoning depth varies by strategy, context, and demographic cues.

This paper introduces DoctorRAG, a medical AI that mimics doctor-like reasoning by combining textbook knowledge with insights from similar patient cases, using "textual gradients" to refine its answers for improved accuracy and relevance.

SemDiD generates semantically diverse LLM outputs by guiding decoding in embedding space through orthogonal direction vectors and inter-group repulsion, outperforming existing methods in Best-of-N evaluations and accelerating RLHF training.

We propose CryptoMoE, the first framework to enable private, accurate, and efficient inference for MoE-based LLMs.

Reasoning Path Compression accelerates inference of reasoning LLMs by periodically compressing KV cache of generated tokens exploiting their semantic sparsity.

For deductive formal problem-solving (D-FPS), we propose a method and a data generation pipeline.

Patent examination remains hard for NLP since evaluation needs to consider examiners’ reasoning. PANORAMA captures this with 8,143 U.S. records and full trails (applications, prior art, rejections, allowances) split into stepwise benchmarks.

We introduce a challenging benchmark to evaluate LLMs' mathematical reasoning and code-writing abilities, finding that specialized models like o1-mini outperform earlier ones but still struggle overall.

LayerNavigator is a low-overhead method that scores each LLM layer's steerability to guide multi-layer activation steering, significantly outperforming baselines while offering clear interpretability.

Training a new reasoning paradigm of LLMs explicitly contains meta-thinking in a multi-agent and multi-turn setting with RL

We propose an efficient lifelong model editing method which maintains high performance with large number of edits.

We propose Rectified Policy Optimization (RePO) to mitigate "safety compensation", which replaces the average safety metric with stricter safety constraints.

This paper demonstrates that language models possess metacognitive-like abilities to monitor and control their internal neural activations.

CoIDO is a lightweight dual-objective data selection framework that jointly optimizes importance and diversity, achieving 98.2% of full-data performance for multimodal instruction tuning using only 20% of the data.

An LLM-based framework that automates executable environment building to enable scaling up executable code data.

We introduce an RL framework that unify the training of answer generation and verification in a single model.

We analyze the effective depth of LLMs and find that they are unlikely to compose subresults, and deeper models spread out the same type of computation as the shallow ones.

A new method of uncertainty quantification for LLMs based on minimum Bayes risk framework combines model confidence with observed consistency.

Propose benchmark and sataset to evaluate LLMs extraction of state machines from network protocols standards, curated the datasets and benchmark on several open and close LLMs.

We improve the fine-tuning performance of reasoning LLMs by identifying the critical steps

This paper proposed a LLM-based plug-in, which is compatible with various RL algorithms, that enhances the efficiency of policy exploration in RL training.

This paper proposes a Multi-Model Contrastive Decoding algorithm and uses a dynamic hallucination detection mechanism in the generation phase to significantly reduce the model's hallucinations.

We propose an attention-shifting-based unlearning method for large language models, which enables precise removal of target knowledge while preserving utility and mitigating hallucinations.

We propose a taxonomy of five knowledge statuses and introduce KScope, a hierarchical testing framework to characterize LLM knowledge status.

Advancing Machine-Generated Text Detection from an Easy to Hard Supervision Perspective

Inference-time hyper-scaling uses key–value cache compression with Delayed Memory Sparsification (DMS) to boost Transformer LLM reasoning accuracy for equivalent compute or memory costs.

We propose DEAL, a continual low-rank fine-tuning framework that enables efficient and privacy-preserving adaptation of large language models.

RidgeLoRA adopts a novel series connection architecture, and enables different parameter initialization approaches for better performances.

We present MLRC-Bench, a dynamic benchmark designed to rigorously assess how well language agents address ML research challenges with objective, performance-based evaluations.

DataSIR, a multi-format sensitive information benchmark dataset, including 1,647,501 samples, designed to evaluate the performance of different models in recognizing sensitive information under evolving data leakage techniques.

We present a fairshare training data pricing framework for large language models where data pricing depends on the data valuation (which estimates data contribution towards model learning), and benefits both sellers and buyers.

We propose a lesson-based framework for multiple LLM agents to collaboratively solve coding problems by learning from each other.

We propose Magical, an asymmetric LoRA architecture to meet the requirement for semantic fidelity and diverse lay-style generation in medical lay language generation task.

We investigate catastrophic forgetting in LLM medical knowledge injection and propose an internal knowledge augmentation learning method to mitigate it.

A runtime verification framework of LLMs with domain knowledge.

We introduced Multi-agent KTO, a method that trains LLM to play Werewolf through direct gameplay. Our approach outperforms GPT-4o and RL+LLM methods, achieving human-competitive performance.

Efficient Long Context Fine-tuning through Dynamic Data Scheduling

Based on systematic empirical analysis on post-training compression, we propose a calibration data curation framework to help pruning and quantization methods better preserve critical LLM capabilities.

We propose a test-time detoxification framework that models toxicity transitions within the latent representation space to enable stable and precise representation editing guidance.

We propose a Belief-Calibrated Consensus Seeking (BCCS) framework to facilitate stable consensus in multi-agent system via selecting optimal collaborators and calibrating the consensus judgment by system-internal beliefs.

We propose Fira, a plug-and-play memory-efficient training framework of LLMs to enable full-rank training consistently under the low-rank constraint.

This study makes a key contribution by introducing a novel systematic framework to interpret the translation mechanisms of LLMs from a computational components perspective, an area previously unexplored.

We find applying a query-dependent head-specific sigmoid gate after the Scaled Dot-Product Attention (SDPA) consistently improves performance, improves scaling properties and mitigates the `massive activation' and `attention sink'.

This paper introduces a novel method for generating benchmarks to evaluate semantic similarity methods for LLM outputs, achieving cross-domain scalability and not replying on human judgment

MedAgentBoard finds multi-agent useful for some medical tasks, but single LLMs or conventional methods are often better, especially for VQA/EHR. Task-specific AI choice is key.

We propose Causal Intervention TableQA (CIT), which is based on a structural causal graph and applies front-door adjustment to eliminate bias caused by token co-occurrence

We propose an efficient, training-free prompt compression method that retains key information within long inputs using the evaluator heads we identified in transformer-based LLMs.

We propose DORA, a provably optimal resource allocation strategy for efficient test-time scaling.

We empirically analyze, predict, and mitigate forgetting in upstream data during LLM fine-tuning.

We introduce ReCAP, a recursive context-aware framework that enables LLMs to perform long-horizon reasoning and planning through task decomposition, execution, and backtracking, significantly improving performance in complex, multi-step environments.

A pioneering multimodal LLM framework for table modality.

We introduce a lightweight, post-hoc routing framework, with provable guarantees, that safely delegates between language models with competing objectives.

NeSyPr: Neurosymbolic Proceduralization For Efficient Embodied Reasoning

QRPO is a SoTA alignment algorithm that can fit the KL-regularized RL objective without relying on preferences.

We propose a novel code watermarking framework that learn from Abstract Syntax Trees of code to identify where and how to insert watermarks.

PartEvo introduces feature-assisted niche construction in abstract language spaces, allowing the integration of niche-based search strategies into LLM-assisted evolutionary search, ultimately offering more efficient automated algorithm discovery.

We mechanistically analyze how large language models solve synthetic propositional logic problems.

DynaPipe dynamically redistributes layers and uses asynchronous coordination to balance computation during LLM inference, significantly reducing latency and outperforming existing pipeline parallelism systems.

We propose a novel benchmark for pattern recognition for many-shot in-context learning for large language models and conduct extensive empirical analysis with many insights.

Optimize KV cache eviction by adaptively allocating budgets across different attention heads for efficient LLM inference

We propose ExSearch, an agentic search framework, where the LLM learns to retrieve useful information as the reasoning unfolds through a self-incentivized process.

We analyze what kind of LLMs have large improvement in RL finetuning and propose behavior injection augmentation to prepare the LLMs for RL.

We show that limiting a model's confidence during training can improve test-time scaling in mathematical reasoning.

We propose a method which exploit KV cache sparsity efficiently and dynamically through Top-P sampling.

Towards Reliable Code-as-Policies: A Neuro-Symbolic Framework for Embodied Task Planning

We introduce the first theoretical framework for analyzing LLM reasoning errors, and bridge two typical sampling-based test-time scaling methods to achieve both low error and fast convergence.

RetrievalAttention improves decoding speed and reduces GPU memory usage in Transformer-based LLMs by using pre-built, attention-aware KV vector indexes stored in CPU memory, achieving significant efficiency gains without compromising accuracy.

We propose HP-KG, which enhances LLMs' planning capabilities while reducing model scale requirements for robotic manipulation tasks.