| Total: 74
Representation Fine-tuning (ReFT), a recently proposed parameter-efficient fine-tuning (PeFT) method, significantly improves parameter efficiency by modifying the representation space alone. However, directly applying ReFT, which alters a fixed number of representations at the beginning and end positions of each layer, results in suboptimal performance for two reasons. (i) The impact of these fixed-position representations on the output is uncertain; (ii) As the sequence length increases, fine-tuning a fixed number of representations may have diminishing effects on the final results. Based on our observations that punctuation plays a crucial role in integrating representations from preceding layers and modulating those of subsequent layers, we introduce Punctuation-steered Representation Fine-tuning (PuReFT), a straightforward yet powerful approach that additionally fine-tunes punctuation representations to achieve performance improvements. Extensive evaluations on common-sense, arithmetic, and code datasets demonstrate the effectiveness and versatility of PuReFT. Furthermore, our analysis of its training speed and memory overhead confirms its greater ease of use and efficiency.
While self-consistency is known to improve performance on symbolic reasoning, its effect on the recall of encyclopedic knowledge is unclear due to a lack of targeted evaluation grounds. To address this, we establish such a knowledge recall split for the popular MMLU benchmark by applying a data-driven heuristic from prior work. We validate this split by showing that the performance patterns on the symbolic reasoning and knowledge recall subsets mirror those of GSM8K and MedMCQA, respectively. Using this solid ground, we find that self-consistency consistently improves performance across both symbolic reasoning and knowledge recall, even though its underlying CoT prompting is primarily effective for symbolic reasoning. As a result, we achieve an 89% accuracy on MMLU, the best performance to date with the use of GPT-4o.
Commonsense reasoning often requires both textual and visual knowledge, yet Large Language Models (LLMs) trained solely on text lack visual grounding (e.g., “what color is an emperor penguin’s belly?”). Visual Language Models (VLMs) perform better on visually grounded tasks but face two limitations: (i) often reduced performance on text-only commonsense reasoning compared to text-trained LLMs, and (ii) adapting newly released LLMs to vision input typically requires costly multimodal training. An alternative augments LLMs with test-time visual signals, improving visual commonsense without harming textual reasoning, but prior designs often rely on early fusion and a single image, which can be suboptimal. We propose a late multi-image fusion method: multiple images are generated from the text prompt with a lightweight parallel sampling, and their prediction probabilities are combined with those of a text-only LLM through a late-fusion layer that integrates projected visual features just before the final prediction. Across visual commonsense and NLP benchmarks, our method significantly outperforms augmented LLMs on visual reasoning, matches VLMs on vision-based tasks, and, when applied to strong LLMs such as LLaMA 3, also improves NLP performance while adding only modest test-time overhead.
While full-duplex speech agents enable natural, low-latency interaction by speaking and listening simultaneously, their consistency and task performance in multi-turn settings remain underexplored. We introduce Full-Duplex-Bench-v2 (FDB-v2), a streaming framework that integrates with an automated examiner that enforces staged goals under two pacing setups (Fast vs. Slow). FDB-v2 covers four task families—Daily, Correction, Entity Tracking, and Safety—and reports turn-taking fluency, multi-turn instruction following, and task-specific competence. The framework is extensible, supporting both commercial APIs and open-source models. When we test full-duplex systems with FDB-v2, they often get confused when people talk at the same time, struggle to handle corrections smoothly, and sometimes lose track of who or what is being talked about. Through an open-source, standardized streaming protocol and a task set, FDB-v2 makes it easy to extend to new task families, allowing the community to tailor and accelerate evaluation of multi-turn full-duplex systems.
Retrieval-augmented generation (RAG) over long documents typically involves splitting the text into smaller chunks, which serve as the basic units for retrieval. However, due to dependencies across the original document, contextual information is often essential for accurately interpreting each chunk. To address this, prior work has explored encoding longer context windows to produce embeddings for longer chunks, yet their gains in retrieval and downstream tasks remain limited. This is because (1) longer chunks strain the capacity of embedding models due to the increased amount of information they must encode, and (2) many real-world applications still require returning localized evidence due to constraints on model or human bandwidth. To this end, we propose an alternative approach to this challenge by representing short chunks in a way that is conditioned on a broader context window to enhance retrieval performance – i.e., situating a chunk’s meaning within its context. We further show that existing embedding models are not well-equipped to encode such situated context effectively, and thus introduce a new training paradigm and develop the first situated embedding model. To evaluate our method, we curate a book-plot retrieval dataset specifically designed to assess situated retrieval capabilities. On this benchmark, our 1B-parameter model substantially outperforms state-of-the-art embedding models, including several with up to 7B parameters.
The world is in the grip of ecological, meaning, and language crises that are converging into a metacrisis. Big AI is accelerating them all. LLM engineering sits at the core. Despite the public good motives of language engineers and the promise of LLMs, this work is being leveraged to create unprecedented wealth and power for a handful of individuals and corporations while causing existential harm to life on earth. As a profession, we urgently need to come together to explore alternatives and to design a life-affirming future for our field of natural language processing that is centered on human flourishing on a living planet.
Current research on Event Factuality Prediction (EFP) predominantly treats LLMs as passive classifiers, where high aggregate metrics often mask shortcut learning and unreliable reasoning. In this position paper, we argue for a focus shift from event factuality to meta-factivity. We introduce the Meta-Factivity Framework (MFF), a theoretical roadmap that moves evaluation beyond surface recognition to belief trajectory reasoning and epistemic regulation. By framing hallucination as a failure of meta-cognitive control, we advocate for a transition from measuring black-box accuracy to evaluating white-box cognition, laying the groundwork for a more rigorous benchmark for explainable self-governance.
Transformers often display an attention sink: probability mass concentrates on a fixed, content-agnostic position. Are sinks a byproduct of the optimization/training regime? Or are they sometimes functionally necessary in softmax Transformers? We prove that, in some settings, it is the latter: computing a simple trigger-conditional behavior necessarily induces a sink in softmax self-attention models. Our results formalize a familiar intuition: normalization over a probability simplex must force attention to collapse onto a stable anchor to realize a default state (e.g., when the model needs to ignore the input). We instantiate this with a concrete task: when a designated trigger token appears, the model must return the average of all preceding token representations, and otherwise output zero, a task which mirrors the functionality of attention heads in the wild (Barbero et al., 2025; Guo et al., 2024). We also prove that non-normalized ReLU attention can solve the same task without any sink, confirming that the normalization constraint is the fundamental driver of sink behavior. Experiments validate our predictions and demonstrate they extend beyond the theoretically analyzed setting: softmax models develop strong sinks while ReLU attention eliminates them in both single-head and multi-head variants.
Transformers commonly exhibit an attention sink: disproportionately high attention to the first position. We study this behavior in GPT-2–style models with learned query biases and absolute positional embeddings. Combining structural analysis with causal interventions, validated across natural-language, mathematical, and code inputs, we find that the sink arises from the interaction among (i) a learned query bias, (ii) the first-layer MLP transformation of the positional encoding, and (iii) structure in the key projection. Crucially, each component we identify is individually dispensable: architectures omitting each of them robustly exhibit sinks. This indicates that attention sinks may arise through distinct circuits across architectures. These findings inform mitigation of sinks, and motivate broader investigation into why sinks emerges.
Classifier-based Quality Filtering has recently emerged as a fundamental technique in constructing pre-training corpora. The ability to deploy a single model that can replace or supplement a set of heuristics has proven effective across numerous Large Language Models. In this work, we expose a critical vulnerability in this approach by demonstrating how a straightforward Wikipedia-style reformatting operation can substantially alter a model’s quality assessment and enable low-quality content to surpass filtering thresholds. Our analysis reveals that the FineWeb-Edu CQF model would reverse its filtering decision for approximately 7% of evaluated documents, thereby admitting content into the pre-training corpus that would otherwise have been excluded.
Group-based reinforcement learning methods, like Group Relative Policy Optimization (GRPO), are widely used nowadays to post-train large language models. Despite their empirical success, they exhibit structural mismatches between reward optimization and the underlying training objective. In this paper, we present a theoretical analysis of GRPO style methods by studying them within a unified surrogate formulation. This perspective reveals recurring properties that affect all the methods under analysis: (i) non-uniform group weighting induces systematic gradient biases on shared prefix tokens; (ii) interactions with the AdamW optimizer make training dynamics largely insensitive to reward scaling; and (iii) optimizer momentum can push policy updates beyond the intended clipping region under repeated optimization steps. We believe that these findings highlight fundamental limitations of current approaches and provide principled guidance for the design of future formulations.
Long-term conversational agents need memory systems that capture relationships between events, not merely isolated facts, to support temporal reasoning and multi-hop question answering. Current approaches face a fundamental trade-off: flat memory is efficient but fails to model relational structure, while graph-based memory enables structured reasoning at the cost of expensive and fragile construction. To address these issues, we propose StructMem, a structure-enriched hierarchical memory framework that preserves event-level bindings and induces cross-event connections. By temporally anchoring dual perspectives and performing periodic semantic consolidation, StructMem improves temporal reasoning and multi-hop performance on LoCoMo, while substantially reducing token usage, API calls, and runtime compared to prior memory systems.
3D visual grounding (3DVG) aims to localize objects in a 3D scene based on natural language queries. In this work, we explore zero-shot 3DVG from multi-view images alone, without requiring any geometric supervision or object priors. We introduce Z3D, a universal grounding pipeline that flexibly operates on multi-view images while optionally incorporating camera poses and depth maps. We identify key bottlenecks in prior zero-shot methods causing significant performance degradation and address them with (i) a state-of-the-art zero-shot 3D instance segmentation method to generate high-quality 3D bounding box proposals and (ii) advanced reasoning via prompt-based segmentation, which utilizes full capabilities of modern VLMs. Extensive experiments on the ScanRefer and Nr3D benchmarks demonstrate that our approach achieves state-of-the-art performance among zero-shot methods.
Retrieval-Augmented Generation (RAG) improves the factual accuracy of large language model (LLM) outputs by grounding generation in external knowledge. Recent agentic RAG systems extend this paradigm by introducing critic agents that evaluate model responses and iteratively refine outputs. However, most prior work implicitly assumes reliable critic feedback and focuses on planning strategies, while paying limited attention to the robustness of the error correction process itself, which is often hindered by misaligned error categories and ineffective or incorrect corrections. We hypothesize that RAG performance can be improved without explicit error categorization. To this end, we propose RePAIR, a response–action learning paradigm that directly maps flawed RAG outputs to error-mitigating action plans without relying on fine-grained error taxonomies or explicit critic supervision. Across multiple benchmarks, RePAIR consistently improves agentic RAG performance.
Agentic LLM inference with long contexts is increasingly limited by memory bandwidth rather than compute. In this setting, SwiGLU MLP blocks, whose large weights exceed cache capacity, become a major yet under-optimized bottleneck in the Transformer architecture. We propose DeepFusionKernel, a deeply fused kernel that cuts HBM traffic and boosts cache reuse, delivering up to 13.2% speedup on H100 and 9.7% on A100 over SGLang. Integrated with SGLang and paired with a kernel scheduler, DeepFusionKernel ensures consistent accelerations across generation lengths, while remaining adaptable to diverse models, inference configurations, and hardware platforms.
Anchoring bias causes large language models (LLMs) to shift quantitative judgments in response to irrelevant numerical primes. We analyze this bias as a function of model confidence and accuracy in base, instruction-tuned, and distilled variants of Llama and Qwen models. We find that anchoring susceptibility is negatively correlated with model confidence without regard to accuracy: confidently incorrect models resist anchoring as effectively as accurate ones, provided their internal priors are sufficiently strong. We further show that post-training impacts the strength of this relationship, and that models are more susceptible to high anchors than to low anchors. Our findings suggest anchoring resistance is a structural property of distributional concentration (certainty) rather than knowledge correctness (factual accuracy), with implications for deploying LLMs in numerical reasoning tasks.
Relation extraction represents a fundamental component in the process of creating knowledge graphs, among other applications. Large language models (LLMs) have been adopted as a promising tool for relation extraction, both in supervised and in-context learning settings. However, in this work we show that their performance still lags behind much smaller architectures when the linguistic graph underlying a text has great complexity. To demonstrate this, we evaluate four LLMs against a graph-based parser on six relation extraction datasets with sentence graphs of varying sizes and complexities. Our results show that the graph-based parser increasingly outperforms the LLMs, as the number of relations in the input documents increases. This makes the much lighter graph-based parser a superior choice in the presence of complex linguistic graphs.
One of biggest missing capabilities in state-of-the-art AI systems is the ability to learn continually after deployment. However, implementing an inference-time learning system has several challenges including the large memory requirement of gradient-based algorithms that are used to train state-of-the-art LLMs. Evolutionary Strategies (ES) have recently re-emerged as a gradient-free alternative to traditional learning algorithms and have shown encouraging performance on specific tasks in LLMs. In this paper, we perform a more comprehensive analysis of ES and specifically evaluate its forgetting curves when training for a larger number of update steps. We find that although ES is able to reach performance numbers closer to GRPO for math and reasoning tasks, it is accompanied by significant forgetting of prior abilities. We also show that the updates made using ES are much less sparse and have a larger l2 norm compared to corresponding GRPO updates, explaining the contrasting forgetting curves between the two algorithms. With this study, we aim to specifically highlight the issue of forgetting in gradient-free algorithms like ES and hope to inspire future work to mitigate these issues.
Reinforcement Learning (RL) enhances LLM reasoning, yet a paradox emerges as models scale: strong base models saturate standard benchmarks (e.g., MATH), yielding correct but homogeneous solutions. In such environments, the lack of failure cases causes the advantage signal in group-relative algorithms (e.g., GRPO) to vanish, driving policies into mode collapse. To address this, we propose Constrained Uniform Top-K Sampling (CUTS), a parameter-free decoding strategy enforcing structure-preserving exploration. Unlike standard sampling that follows model biases, CUTS flattens the local optimization landscape by sampling uniformly from constrained high-confidence candidates. We integrate this into Mixed-CUTS, a training framework synergizing exploitative and exploratory rollouts to amplify intra-group advantage variance. Experiments on Qwen3 models demonstrate that our approach prevents policy degeneration and significantly boosts out-of-domain generalization. Notably, Mixed-CUTS improves Pass@1 accuracy on the challenging AIME25 benchmark by up to 15.1% over standard GRPO, validating that maintaining diversity within the semantic manifold is critical for rigorous reasoning.
Recent large language models (LLMs) perform strongly on mathematical benchmarks yet often misapply lemmas, importing conclusions without validating assumptions. We formalize lemma-judging as a structured prediction task: given a statement and a candidate lemma, the model must output a precondition check and a conclusion-utility check, from which a usefulness decision is derived. We present RULES, which encodes this specification via a two-section output and trains with reinforcement learning plus section-aware loss masking to assign penalty to the section responsible for errors. Training and evaluation draw on diverse natural-language and formal proof corpora; robustness is assessed with a held-out perturbation suite; and end-to-end evaluation spans competition-style, perturbation-aligned, and theorem-based problems across various LLMs. Results show consistent in-domain gains over both a vanilla model and a single-label RL baseline, larger improvements on applicability-breaking perturbations, and parity or modest gains on end-to-end tasks; ablations indicate that the two-section outputs and section-aware reinforcement are both necessary for robustness.
Analog in-memory computing (AIMC) offers substantial efficiency gains for transformer inference but introduces hardware-induced noise that can distort attention behavior. Prior studies primarily focus on AIMC evaluations for vision tasks and CNN-based models. They largely overlook how hardware-induced noise perturbs internal attention dynamics in NLP models. In this work, we present the first fine-grained analysis of analog vulnerability in pretrained transformers, examining projection submodules, attention heads, and layer-wise dynamics across multiple NLP tasks. Results show that query (Q), key (K), and value (V) projections are the most sensitive components, while feed-forward layers remain comparatively robust. Also, analog noise yields depth-dependent degradation in higher layers, leading to scattered attention and disrupted token routing. This pre-deployment analysis mitigates potential resource misuse before physical deployment and offers practical guidance for designing noise-resilient analog NLP transformers.
Evaluation of modern large language model (LLM) systems increasingly relies on multi-stage pipelines such as retrieval-augmented generation, tool-using agents, and prompt chains. Reproducing reported evaluation results for these systems often requires evaluation-specific artifacts beyond model weights and datasets, including prompts, judge configurations, retrieval snapshots, and intermediate traces, yet their availability has not been systematically examined.We introduce ReproEvalCard, a lightweight reporting standard that specifies the minimal artifacts required to reproduce and validate evaluations of LLM pipelines. To motivate this standard, we audit 55 pipeline-based LLM papers published between 2022 and 2025 and quantify the availability of reproducibility-critical evaluation artifacts. We find that randomness controls are missing in 75% of papers and intermediate execution traces in 61%, substantially limiting evaluation reproducibility. We further demonstrate ReproEvalCard through a worked example and provide a concise checklist for authors and reviewers, aiming to improve reproducibility and comparability in LLM evaluation.
Large language models (LLMs) now give reasoning before answering, excelling in tasks like multiple-choice question answering (MCQA). Yet, a concern is that LLMs do not solve MCQs as intended, as work finds LLMs sans reasoning succeed in MCQA without using the question, i.e., choices-only. Such partial-input success is often deemed problematic, but reasoning traces could reveal if these strategies are truly shallow in choices-only settings. To study these strategies, reasoning LLMs solve MCQs in full and choices-only inputs; test-time reasoning often boosts accuracy on full and in choices-only half the time. While possibly due to shallow shortcuts, choices-only success is barely affected by the length of reasoning traces, and after finding traces pass faithfulness tests, we show they use less problematic strategies like inferring missing questions. In all, we challenge claims that partial-input success is always a flaw, so we discuss how reasoning traces could separate problematic data from less problematic reasoning.
Automated 3D radiology report generation often suffers from clinical hallucinations and a lack of the iterative verification found in human practice. While recent Vision-Language Models (VLMs) have advanced the field, they typically operate as monolithic “black-box” systems without the collaborative oversight characteristic of clinical workflows. To address these challenges, we propose MARCH (Multi-Agent Radiology Clinical Hierarchy), a multi-agent framework that emulates the professional hierarchy of radiology departments and assigns specialized roles to distinct agents. MARCH utilizes a Resident Agent for initial drafting with multi-scale CT feature extraction, multiple Fellow Agents for retrieval-augmented revision, and an Attending Agent that orchestrates an iterative, stance-based consensus discourse to resolve diagnostic discrepancies. On the RadGenome-ChestCT dataset, MARCH significantly outperforms state-of-the-art baselines in both clinical fidelity and linguistic accuracy. Our work demonstrates that modeling human-like organizational structures enhances the reliability of AI in high-stakes medical domains.
Prefix parsing asks whether an input prefix can be extended to a complete string generated by a given grammar. In the weighted setting, it also provides prefix probabilities, which are central to context-free language modeling, psycholinguistic analysis, and syntactically constrained generation from large language models. We introduce the prefix grammar transformation, an efficient reduction of prefix parsing to ordinary parsing. Given a grammar, our method constructs another grammar that generates exactly the prefixes of its original strings. Prefix parsing is then solved by applying any ordinary parsing algorithm on the transformed grammar without modification. The reduction is both elegant and practical: the transformed grammar is only a small factor larger than the input, and any optimized implementation can be used directly, eliminating the need for bespoke prefix-parsing algorithms. We also present a strategy—based on algorithmic differentiation—for computing the next-token weight vector, i.e., the prefix weights of all one-token extensions, enabling efficient prediction of the next token. Together, these contributions yield a simple, general, and efficient framework for prefix parsing.