This paper presents LittleBit, a novel framework that combines latent matrix factorization and a multi-scale compensation mechanism to compress Large Language Models (LLMs) to ultra-low bit levels.

We characterize reward hacking in inference time alignment and demonstrate via three inference-time methods that we can leverage hedging to provably mitigate reward hacking.

We present Omni-Mol, a multitask molecular model that can solve any-to-any modality tasks

We explore the role of verification granularity in sampling-based test-time scaling.

We introduce a principled framework for validating LLM-as-a-judge systems under rating indeterminacy, where multiple ratings can be "correct."

We examine how problem complexity affects reasoning models' behavior in controlled test settings, revealing key strengths and limitations of their reasoning capabilities.

A synthesis of layer-wise interventions, empirical experiments, and previous research suggests that inference in decoder-only LLMs unfolds in distinct phases.

We propose AdaReasoner, a reinforcement learning–based framework that adaptively tunes LLM reasoning settings in few-shot scenarios, achieving gains in multiple experimental settings.

We propose Reward Reasoning Models, which leverage additional test-time compute for complex queries where appropriate rewards are not immediately apparent.

SATURN enables scalable, verifiable, and curriculum-controlled reinforcement learning that enhances the reasoning capability of LLMs, and generalizes to math and programming tasks.

We introduce HELM, a family of hyperbolic large language models operating fully in hyperbolic space.

This paper presents a layer-parallel speculation strategy for optimizing the effiency of multi-GPU utilization during the drafting stage of speculative decoding.

Quantile-Guided Alignment (QA) is a framework for multi-dimensional quantile alignment that reduces catastrophic risks in language models by imposing constraints on reward quantiles across multiple performance dimensions.

We present G-Vendi, a data diversity measure that strongly correlates with LLM reasoning generalization in OOD benchmarks; we use this insight to diverse synthetic reasoning data, which leads to SOTA distilled models in NLI and math reasoning.

We combine LLM-synthesized performance-characterizing constraints with fuzzing to uncover difficult-to-find code inefficiencies and generate performance-stressing tests.

In this paper, we decompose the reward value into prompt-free reward and prompt-related reward from a information-theoretic perspective, and use the former to guide reward training.

M-Pilot uses a small, controllable language model to guide a large, complex language model through complex tasks, improving its reasoning, planning, and personalization capability.

Transformer-based language models learn low-dimensional task manifolds across layers, with similar patterns/trends in intrinsic dimensions revealing similar compression strategies despite varying architectures/sizes.

This work evaluates LLMs' utility and risk in a concurrent scenario and proposes a novel jailbreak attack against LLMs based on task concurrency.

Current multi-modal LLMs struggles with live, step-by-step task guidance. We built Qualcomm Interactive Cooking (a new dataset with mistake videos and timed feedback) and LiveMamba (a streaming model) to enable better real-time interactive guidance.

Influence Distillation is a mathematically justified data selection method for LLM fine-tuning that assigns optimal weights to training samples, achieving performance on par with or better than state-of-the-art while being substantially faster.

We propose a new distillation approach that removes the input question for adaptive and efficient reasoning.

SpecMAS converts natural-language SOPs into NuSMV models and autonomously self-verifies and debugs them via multi-agent temporal-logic checking, ensuring verifiable system designs.

We investigate how digit bias learned from skewed pretraining data lead to systematic numerical hallucinations in LLMs.

In this paper, we interpret the mechanism behind safety alignment via neurons and analyze their properties.

We propose a bilevel framework that jointly learns layer-wise sparsity and low-rank structure via truncated Gaussian sampling and efficient matrix approximation, achieving better performance for compressing LLMs.

Efficiently generate text avoiding constraint violations, with less distortion of LLM output distribution.

RL fine-tuning in LLMs updates a small subnetwork containing 20–30% of parameters leaving rest of the parameters unchanged.

We introduce CARES, a 18K-prompt benchmark for evaluating medical safety of LLMs under adversarial conditions, with graded harms, jailbreaks, and a fine-grained response metric.

We proposed a framework for reinforcing large reasoning models with discriminative constrained optimization , grounded in the principle that increasing the scores of positive answers while decreasing those of negative ones.

We propose Adaptive Reasoning Model (ARM), a reasoning model capable of adaptively selecting appropriate reasoning formats based on the task at hand.

We propose a novel LLM-based peer review system that evaluates academic papers through pairwise comparisons.

LLMs show a utilitarian boost in moral judgment when reasoning in groups, similar to humans, but driven by distinct model-specific mechanisms, highlighting key considerations for multi-agent alignment and moral reasoning.

We embed personalized LLM Watermarks responsively into generated text with Sparse Autoencoders, with only inference-time sampling on black-box access LLMs and no extra training, and achieved high accuracy while preserving text quality.

We developed an iterative weakly supervised pipeline to refine LLM-generated pseudo-labels, consistently outperforming original LLMs and existing self-refinement methods across diverse datasets, while effectively supporting LLM safety alignment.

We design a geometric sampling framework to enhance memory efficiency and precision in long-context LLMs.

BC-LLM is a novel procedure that integrates Large Language Models into a Bayesian framework for concept discovery to achieve better predictive performance, converge faster to relevant concepts, and provide rigorous uncertainty quantification.

We introduce T-SHIRT, a new data selection method for instruction tuning LLMs that scores data at the token level and emphasizes robustness.

In this paper, we propose a novel Cascade Adaptive Self-Speculative Decoding (CAS-Spec) algorithm which constructs speculative draft models by leveraging dynamically switchable inference acceleration (DSIA) strategies

We analyze and compromise existing TEE-based secure inference methods by our Model Stealing attacks with Prior, and propose a novel secure inference method named LoRO, which is efficient, secure and accurate.

Introduce and apply Goodness-of-Fit (GoF) tests to watermark detection, and explore their effectiveness across diverse conditions.

We propose a simple way to measure the critical batch size for language model pretraining that alleviates issues with existing methods, and show that this can be used to train language models with fewer gradient steps in practice.

We introduce CaPT, leveraging both instance-aware and task-aware information for effective and efficient prompt-based learning.

We study the problem of computing an optimal large language model (LLM) policy for a constrained alignment problem.

We propose an uncertainty-based routing framework that efficiently complements a fast RM with a strong but costly LLM judge.

We propose a distillation framework for training language model anonymizers capable of effective anonymization via iterative self-refinement

Analogy-Based Multi-Turn Jailbreak Method for LLMs

We propose LayerIF, a framework that employs Influence Functions for LLM layer quality estimation. Our method captures task-specific layer importance and improves both expert allocation in LoRA-MoE and layer-wise sparsity distribution in LLM pruning.

Facilitate instruction optimization for black-box LLM by leveraging the preimage structure over the soft prompts and instructions.

AcuRank adaptively reranks documents using uncertainty, improving ranking accuracy with fewer LLM calls.

We introduce Set-LLM, a permutation-invariant LLM architecture that eliminates order bias and sensitivity.

We systematically evaluate the concerns of multiplicity and monoculture in a suite of large language models and prediction tasks.

Our empirically and theoretically informed method, which treats diversity as a reward, achieves new SOTA average performance across 7 benchmarks on SOTA LLMs with domain-undetermined data.

We propose a training-free visual cropping method that leverages MLLM-internal representations for VQA tasks focusing on small details, achieving strong performance with significantly higher efficiency than prior methods.

This paper introduces PANGEA, a method that leverages general-purpose data to generate diverse and high-quality synthetic data, improving LLM performance on domain-specific tasks.

We introduce NeuComBack, a novel benchmark for Neural Compilation, and propose an automatic prompt learning method that significantly improves both correctness and optimization of LLM-generated assembly.

Emergence of an alignment between LLMs' and the brain's computational dynamics, and key factors allowing it : scaling and context size.

We analyze alignment faking propensities in 23 LLMs, and attempt to explain why some LLMs fake alignment and others don't.

ClinBench is an open-source, multi-model, multi-domain framework for rigorously benchmarking large language models on clinical information-extraction tasks.

We introduce IneqMath, an informal inequality proving benchmark, and an LLM-as-judge suite, revealing that top LLMs achieve <10% overall accuracy due to flawed step-wise reasoning.

DeepFund introduces a live benchmarking framework that prevents LLMs from 'cheating' by using live market data, revealing that most current LLMs struggle to make profitable trading decisions in real-time market conditions.

We propose ModuLM, a flexible framework for LLM-based molecular relational learning, supporting multimodal inputs and dynamic model construction.

We introduce NaturalReasoning, a 2.8M-question dataset spanning diverse domains, enabling effective knowledge distillation and unsupervised self-training to enhance LLM reasoning capabilities.

This paper introduces CorrectBench, the first comprehensive benchmark for systematically evaluating self-correction mechanisms in LLMs.

We compare leading open SFT datasets, add quality annotations using MagPie, and design curation recipes leading to a high-performing leaner SFT mixture

Backdooring MoE-based LLMs by exploiting inherent vulnerabilities in the expert routing mechanism.

A framework that uses Monte Carlo Tree Search for guiding an LLM to ask information-seeking questions, learning from past successful strategies to solve problems more efficiently and accurately.

We propose a four-stage training task for LLMs—Fast Thinking, Verification, Slow Thinking, and Summarization—separating the training of intuition, verification, reasoning, and integration.

A Training-Free Bayesianization approach is proposed for LLM adapters that achieves better uncertainty estimation.

We fit scaling laws for large language models with varying width-to-depth ratios and parameter counts.

We propose BAM-ICL, a novel budgeted adversarial manipulation hijacking attack framework for in-context learning.

GraphMaster is a collaborative LLM multi-agent system that synthesizes semantically rich, structurally coherent text-attributed graphs to overcome data scarcity and empower Graph Foundation Models.

We propose RF-Agent, an automated RL reward function design framework via language agent tree search.

We introduce 3BASiL-TM, a highly efficient one-shot post-training method for Sparse plus Low-Rank decomposition of LLMs that reduces the WikiText2 perplexity gap to dense model by over $30\%$ compared to prior methods.

Unified theory LLM Alignment framework as divergence estimation between aligned and unaligned distributions; introduce KLDO, an alignment method, and advocate for compliance-refusal datasets to improve safety.

We enable vision-language models to efficiently encode and utilize multimodal multilingual knowledge through a lightweight, continuous memory mechanism.

We investigate the structure of relations in large language models, and compress linear relation decoding operators with tensor networks

We provide a method for accurate end-to-end FP4 training of Large Language Models.

We propose a compute- and memory-efficient inference framework for handling extremely long inputs with pre-trained Transformers.

We propose a projection-based scoring function in KV cache eviction for LLM acceleration.

We construct a Progress Reward Model with convergence guarantee for Reinforcement Learning via Large Language Models.

We rethink the roles of descriptions and narrations in role-playing and build a system to initialize the characters' behavior modes as executable codes with descriptions and refine with narrations.

Training LLMs to combine reasoning with external knowledge retrieval via RL without any supervised data on reasoning steps.

We mix discrete and continuous adversarial attacks to adversarially train more robust LLMs. We evaluate our models in different realistic inference settings and show that they are more robust while matching the training cost of other SoTA models.

We introduce a theoretically-grounded distributional RL algorithm for LLM post-training that demonstrates improvement upon prior work on both synthetic and mathematical reasoning tasks.

We report a reasoning model for chemistry

Uncertainty Quantification for LLMs

We inject a semantic-similarity penalty term into decoding for autoregressive and masked-diffusion LMs, generating semantically diverse answers and, with importance reweighting, enabling consistent estimation of downstream uncertainty metrics.

We propose Router-R1, an RL-based framework that interleaves multi-round reasoning with dynamic LLM selection, supports zero-shot integration of new models, and optimizes performance-cost trade-offs

We provide a theory for jointly optimal, distribution-adaptive LLM watermarking, yielding DAWA, a highly effective, distortion-free algorithm.

We train probes on activations to classify high- vs low-stakes scenarios, find they outperform medium-sized fine-tuned LLMs, and consider applications to monitoring.

This paper propose a preference-driven knowledge distillation (PKD) framework to synergize the complementary strengths of LLMs and various GNNs for few-shot node classification on text-attribute graphs.

This paper presents IPAD, a novel AI detection framework that identifies predicted prompts that could have generated the input text. IPAD outperforms the baselines by 9.73% on in-distribution data and 12.65% on OOD data.

This paper introduces a comprehensive benchmark (COLORBENCH) and detailed analysis to systematically analyze how well VLMs perceive, reason about, and robustly handle colors.

A systematic review of benchmarks for large language models focusing on issues of construct validity, which results in a set of recommendations for best practices.

A dataset of multi-agent system traces, and a systematic analysis of failures in multi-agent LLM systems, featuring a structured taxonomy and an automated evaluation pipeline.

Generation of images in SVG format through LLMs without exploiting their rasterized version

BenchmarkCards provides standardized documentation for large language model benchmarks, simplifying benchmark selection and usage.

We propose a LLM agent framework to automate red teaming

We apply conformal prediction to provide statistical guarantees that all important information within a long-text is captured by an automatically generated summary.

We introduce a novel method leveraging noise injection as a tool to elicit the latent capabilities of sandbagging LLMs.

We create an ad auction mechanism for LLM-generated content that ensures truthful advertiser behavior, aligns utility with social welfare, and provably converges to optimal LLM outputs.

We propose a bandit framework for GenAI-powered decisions where actions (queries) affect the environment only through stochastic model outputs, and introduce algorithms tailored to this setup.

Bipartite mutual information in natural text exhibits sub-volume growth; from this, we prove a lower bound on how the history state must scale, setting a necessary condition for architectures to be effective at long-context language modeling.

We propose AutoToM, an automated agent modeling method for scalable, robust, and interpretable mental inference.

LLMs show human-like representational change during associative learning, driven by pair similarity and modulated by prior knowledge.

BrainEC-LLM

Our paper introduces AgentBreeder, a framework for multi-objective self-improving evolutionary search over scaffolds, specifically targeting scaffolds' safety impact on large language models in multi-agent systems.

LLMs struggle with balancing uncertainty and empowerment in exploration, underperforming in open-ended tasks due to premature decisions driven by early-stage processing biases.

We propose PoLAR, a polar-decomposition-based parameterization, for efficient fine-tuning of LLMs. PoLAR mitigates the low stable rank seen in LoRA, provably accelerates convergence on a canonical LoRA problem, and lifts accuracy on real-world tasks.

We propose correlation dimension as a practical, model-agnostic metric that captures structural complexity and detects degeneration in large language model outputs beyond what perplexity reveals.

NMKE is a fine-grained framework that uses neuron-level attribution and sparse masking for precise, efficient lifelong knowledge editing.

We introduce transductive program synthesis: synthesizing programs using test inputs.

Using coding theory and heavy tails to design powerful watermarks with zero distortion.

new decoding method for diffusion LLM, alternative of semi-AR

We introduce generative distribution embeddings (GDE), a framework that lifts autoencoders to the space of distributions.

We propose a global pruning framework that efficiently learns unstructured sparsity for LLMs.

We propose a cognitive-science-inspired framework and benchmark to systematically evaluate learning abilities of large language models.

We introduce SPRO (Self-Play Reward Optimization), an annotation-free framework that aligns images with human preferences by using vision-language models and reward signals to optimize prompts and images via self-play.

MLZero automates end-to-end multimodal machine learning with high success rate and solution quality without human intervention.

This paper introduces MigGPT, a novel framework that leverages large language models and a code fingerprint structure to automate the migration and maintenance of out-of-tree Linux kernel patches.

CSCR embeds both prompts and LLMs into a shared space using fast logit or perplexity fingerprints. A cost‑banded InfoNCE loss trains the space to balance quality against cost. It generalizes to unseen models and out‑of‑distribution prompts.

This paper tackles the problem of zero-shot inference on text-attributed graphs, and proposes a novel method that uses bundles to query LLMs and supervise GNNs.

Nemotron-CLIMB automates data mixture optimization for pre-training, improving domain adaptation and outperforming Llama-3.2-1B by 2.0% on general reasoning.

We introduce ALE-bench, a new benchmark for evaluating AI systems on score-based algorithmic programming contests.

We present CellVerse, a unified language-centric single-cell analysis benchmark within the realm of LLMs for cell biology.

Under input uncertainty, transformer models exhibit a systematic exploration of input‑agnostic conceptual representations, increasing the likelihood of hallucinations.

Greedy approach to Interactive Question Answering in LLMs via Information Gain and Conformal Inference

This paper presents SpatialLM, a large language model designed to process 3D point cloud data and generate structured scene descriptions.

The first work to study continual learning across diverse structured knowledge reasoning tasks

We show that slow thinking behaviors elicited in the chain-of-thought (CoT) process of reasoning models help large language models (LLMs) express their confidence more accurately.

We propose GPAS, a simple method that scales activations without scaling gradients to accelerate pretraining convergence of LLMs.

We introduce a Particle Filtering approach to Inference Scaling that is robust against inherently imperfect reward models and performs significantly better than previous methods.

We have decoupled the math atomic capabilities of large language models and explored their interaction relationships in mathematical reasoning tasks.

The internal states of a Language Model can often preemptively predict several behaviors of interest, and can be used to form precise early warning systems that detect jailbreaking, instruction following failures and more.

A simple general purpose off-policy REINFORCE method which outperforms PPO, DPO and STaR on recent benchmarks.

This study introduces a novel causality-driven robust optimization approach that selectively updates model components sensitive to causal reasoning, enhancing model causality while preserving valuable pretrained knowledge to mitigate overfitting.

Training LLM’s with tensor-parallelism without completely synchronizing activations to accelerate training and inference.

Quantifying the Potential of Control Algorithms through LLM Agents

Polar Sparsity scales contextual sparsity to large batches by exploiting stable attention head sparsity and using efficient GPU kernels, achieving up to 2.2× speedups with minimal accuracy loss.

We introduce a new method for selecting subspaces in low-rank optimization for memory-efficient pretraining of large language models (LLMs).

We introduce Risk-aware Direct Preference Optimization (Ra-DPO), a novel approach that incorporates risk-awareness by employing a class of nested risk measures.

We show that recurrent-depth transformers can be scaled to be effective language models, with particularly strong gains through additional compute for reasoning tasks.

SCFG-based framework characterizes uncertainty in LLM formal reasoning, enabling selective verification that reduces errors while minimally abstaining.

In this paper we introduce a framework to decompose the uncertainty of predictions in in-context learning

We propose AlphaDecay, a per-module weight decay method guided by heavy-tailedness, improving large language model performance.

We present LLaDA, a diffusion language model trained from scratch that is competitive to LLaMA 3 in performance.

The paper proposes an LLM-guided knowledge graph traversal method that enhances retrieval efficiency and accuracy for complex queries while reducing costs.

GRACE achieves significant gains on performance and efficiency in prompt optimization by strategically introducing information loss through a dual mechanism of refinement and compression.

Develop a test to check if model is obtained by fine tuning of another model

We introduce a technique for coupling probability distributions when several samples are available from one of the distributions, and give applications to multi-draft speculative decoding and distributed lossy compression with side information.

We show that even exact unlearning, the gold standard for data removal in large language models, can leak sensitive information by using guidance between pre- and post-unlearning checkpoints.

We address the problem of temporal text-attributed graph (TTAG) modeling and propose CROSS. CROSS is the first framework designed to unify text semantics and graph structures with LLMs for TTAG modeling.

This paper proposes a novel framework that generates better contrastive pairs for contrastive learning by integrating LLM-based semantic retrieval with a learnable sample synthesizer.

Using Formal Reasoning Tools in RL and at Inference Time to improve Reasoning Capability

In this work, we introduce \textit{BackdoorLLM}, the first comprehensive benchmark for studying backdoor attacks and defenses on LLMs.

We present MergeBench, a comprehensive evaluation suite designed to assess model merging at scale.

We introduce XIFBench, a constraint-based benchmark for evaluating multilingual instruction-following in LLMs across languages with diverse resource levels.

ChemCoTBench bridges complex chemical reasoning with arithmetic-inspired step-by-step workflows, enabling LLMs to systematically tackle real-world tasks like molecular optimization and reaction prediction.

Our new benchmark AbstentionBench reveals reasoning models struggle to determine when not to answer.

LLM-as-RS paradigm is not always good: we benchmark it with traditional recommendation models

We introduce ParallelPrompt, the first benchmark and dataset for studying intra-query parallelism in real-world LLM queries, enabling reproducible evaluation of structured execution strategies.

We train analog foundation models that are robust to noise present in analog in-memory computing hardware and demonstrate accuracy comparable to models trained with 4-bit weight and 8-bit static input quantization.

We propose a curriculum strategy for guiding the training of agents that operate under strict trajectory constraints during deployment by adaptively tightening constraints based on agent's performance.

We propose Learning to Focus (LeaF), which identifies and masks confounding tokens via gradient‐based comparisons, thereby improving long‐context reasoning accuracy and interpretability in large language models.

We show that explicit reasoning via chain-of-thought can hurt instruction-following in LLMs by reducing constraint adherence, and propose four mitigation methods to recover or improve performance.

NeuroPath is a RAG framework inspired by brain path navigation. It tracks semantic paths and refines retrieval via LLM reasoning, enhancing semantic coherence and reducing noise, achieves state-of-the-art performance on multi-hop QA.

We propose a novel continual learning method for language models, built upon the low-rank adaptation framework.

The paper proposes Causal Head Gating, a scalable, unsupervised method to classify transformer attention heads by causal impact on task performance that reveal task-specific sub-circuits.

Generate high-performing LoRA parameters from prompts that is unseen in training.

We propose the first efficient, training-free online routing algorithm for high-volume LLM serving under token budget constraints, achieving significant improvements in both routing performance and cost efficiency.

We propose a novel KV cache compression method that exploits the asymmetry between locally homogeneous keys and heterogeneous values in attention mechanisms, enabling more efficient long-context processing.

A targeted restoration method to recover pruned models while preserving their low cost and high efficiency

We propose a trustworthy reasoning framework for LLMs to reason over knowledge graphs.

We propose a lossless and training-free speculative decoding method to accelerate LLMs that requires offloading on a single memory-limited cosumer GPU.

We present scPilot, a framework for omics-native reasoning in which a large language model converses, invokes bioinformatics tools, and iteratively explains its decisions to for single-cell analysis.

We propose Think-RM, a training framework for generative reward models that enables long-horizon reasoning, and introduce a pairwise RLHF pipeline that directly optimizes policies using pairwise preference rewards.

A novel method to efficiently fine-tune Large Language Models in hyperbolic space, unlocking their latent tree-like structures and significantly boosting complex reasoning performance by adapting directly on the hyperbolic manifold.

We propose to integrate LLMs into tabular data without data serialization and LLMs fine-tuning.

We introduce rStar-Coder to train advanced code reasoning LLMs, with our 14B model achieving comparable performance to QWQ-32B.

We propose DynamicRAG, which optimizes retrieval-augmented generation by using reinforcement learning to adaptively rerank documents based on LLM's responses as feedback, achieving state-of-the-art results across multiple datasets.

Statistical prediction may be sufficient to drive the emergence of internal causal models and causal inference capacities in deep neural networks.

An optimization approach that bridges the gap between training and inference techniques via a highly detailed taxonomy of data characteristics to explicitly control generation attributes and implicitly condition generations during inference.

We propose a novel query-agnostic KV cache eviction method for multi-query scenario.

programming framework that does runtime compilation of an LLM-based agent workflow into a search space, enabling independent experimentation of different overlaying inference-time search strategies

demonstrates reasoning settings where scaling the length of chain-of-thought offers an exponential advantage over parallel scaling and aggregation.

While loss decreases monotonically during LLM training, the representations undergo distinct geometric phases across pretraining and post-training, which in turn determine when and how the model acquires memorization or generalization capabilities.

We introduce an adaptive kernel design method that leverages LLMs as genetic operators to dynamically evolve Gaussian process (GP) kernels during Bayesian optimization (BO)

We find that it is effective to defend "long-length" jailbreak attacks via efficient "short-length" LLM adversarial training, supporting by both theoretical and empirical evidence.

We introduce SWIFT, a token-level self-alignment and distillation method that improves LLMs by assigning token weights via a teacher model

AdaDetectGPT is all your need for detecting LLM-generated text.

We investigate new scaling laws which predict the scaling of LLMs when training them over quantized or sparse representations.

A backtracking method that reverts to safer points during generation, reducing safety violations

LC-Opt is a high-fidelity RL benchmarking environment for optimizing liquid cooling in HPC systems, enabling scalable, explainable, and energy-efficient control with RL and LLMs, while promoting optimal data center solutions.

CoRe: a high-quality, multi-lingual benchmark for evaluating LLMs’ Code Reasoning capabilities with fundamental static analysis tasks.

A multilingual issue resolving benchmark, Multi-SWE-bench, with 2,132 human-validated GitHub issues across 8 widely used programming languages

We propose a benchmark to evaluate the large language models' instruction following ability in agentic scenarios.