paperlist.ai

AI conference papers

AI conference paper library

Search papers from 30+ top conferences, including NeurIPS, ICLR, ICML, CVPR, ACL and AAAI. Filter by keyword, then open any paper for its abstract, authors and links.

arXiv

VibrantLeaves: A principled parametric image generator for training deep restoration models

Raphael Achddou, Yann Gousseau, Saïd Ladjal, Sabine Süsstrunk

In this paper, we introduce a synthetic image generator relying on a few simple principles, specifically focusing on geometric modeling, textures, and a simple modeling of image acquisition. These principles, integrated into the classical Dead Leaves model, allow for the creation of high-quality training sets for image restoration tasks. Standard image denoising and super-resolution networks trained on these datasets achieve performance comparable to those trained on natural image datasets. The motivation behind this approach stems from the limitations of Deep Neural Networks in image restoration tasks. Despite their impressive performance, these networks are often poorly understood and prone to biases inherited from standard natural image training sets. To mitigate these issues, we emphasize the need for a better control over training sets, particularly through the use of synthetic and abstract datasets. Furthermore, our work includes a detailed analysis of the principles considered, identifying which image properties are necessary for maintaining high performance, thus taking a first step towards explainability. Besides, we show that neural networks trained with our synthetic sets naturally inherit robustness to the invariance properties of these sets, namely rotation and scale.

arXiv

SAS: Segment Anything Small for Ultrasound -- A Non-Generative Data Augmentation Technique for Robust Deep Learning in Ultrasound Imaging

Danielle L. Ferreira, Ahana Gangopadhyay, Hsi-Ming Chang, Ravi Soni, Gopal Avinash

Accurate segmentation of anatomical structures in ultrasound (US) images, particularly small ones, is challenging due to noise and variability in imaging conditions (e.g., probe position, patient anatomy, tissue characteristics and pathology). To address this, we introduce Segment Anything Small (SAS), a simple yet effective scale- and texture-aware data augmentation technique designed to enhance the performance of deep learning models for segmenting small anatomical structures in ultrasound images. SAS employs a dual transformation strategy: (1) simulating diverse organ scales by resizing and embedding organ thumbnails into a black background, and (2) injecting noise into regions of interest to simulate varying tissue textures. These transformations generate realistic and diverse training data without introducing hallucinations or artifacts, improving the model's robustness to noise and variability. We fine-tuned a promptable foundation model on a controlled organ-specific medical imaging dataset and evaluated its performance on one internal and five external datasets. Experimental results demonstrate significant improvements in segmentation performance, with Dice score gains of up to 0.35 and an average improvement of 0.16 [95% CI 0.132,0.188]. Additionally, our iterative point prompts provide precise control and adaptive refinement, achieving performance comparable to bounding box prompts with just two points. SAS enhances model robustness and generalizability across diverse anatomical structures and imaging conditions, particularly for small structures, without compromising the accuracy of larger ones. By offering a computationally efficient solution that eliminates the need for extensive human labeling efforts, SAS emerges as a powerful tool for advancing medical image analysis, particularly in resource-constrained settings.

arXiv

AdaSemSeg: An Adaptive Few-shot Semantic Segmentation of Seismic Facies

Surojit Saha, Ross Whitaker

Automated interpretation of seismic images using deep learning methods is challenging because of the limited availability of training data. Few-shot learning is a suitable learning paradigm in such scenarios due to its ability to adapt to a new task with limited supervision (small training budget). Existing few-shot semantic segmentation (FSSS) methods fix the number of target classes. Therefore, they do not support joint training on multiple datasets varying in the number of classes. In the context of the interpretation of seismic facies, fixing the number of target classes inhibits the generalization capability of a model trained on one facies dataset to another, which is likely to have a different number of facies. To address this shortcoming, we propose a few-shot semantic segmentation method for interpreting seismic facies that can adapt to the varying number of facies across the dataset, dubbed the AdaSemSeg. In general, the backbone network of FSSS methods is initialized with the statistics learned from the ImageNet dataset for better performance. The lack of such a huge annotated dataset for seismic images motivates using a self-supervised algorithm on seismic datasets to initialize the backbone network. We have trained the AdaSemSeg on three public seismic facies datasets with different numbers of facies and evaluated the proposed method on multiple metrics. The performance of the AdaSemSeg on unseen datasets (not used in training) is better than the prototype-based few-shot method and baselines. Code: https://github.com/Surojit-Utah/AdaSemSeg

arXiv

Sa2VA: Marrying SAM2 with MLLM for Dense Grounded Understanding of Images and Videos

Haobo Yuan, Xiangtai Li, Tao Zhang, Yueyi Sun, Zilong Huang, Shilin Xu, Shunping Ji, Yunhai Tong, Lu Qi, Jiashi FENG, Ming-Hsuan Yang

This work presents Sa2VA, the first comprehensive, unified model for dense grounded understanding of both images and videos. Unlike existing multi-modal large language models, which are often limited to specific modalities and tasks, Sa2VA supports a wide range of image and video tasks, including referring segmentation and conversation, with minimal single-stage instruction tuning. Sa2VA combines SAM-2, a foundation video segmentation model, with MLLM, advanced vision-language models, and unifies text, image, and video into a shared LLM token space. Using the LLM, Sa2VA generates instruction tokens that guide SAM-2 in producing precise masks, enabling a grounded, multi-modal understanding of both static and dynamic visual content. Additionally, we introduce Ref-SAV, an auto-labeled dataset containing over 72k object expressions in complex video scenes, designed to boost model performance. We also manually validate 2k video objects in the Ref-SAV datasets to benchmark referring video object segmentation in complex environments. Experiments show that Sa2VA achieves strong performance across multiple tasks, particularly in referring video object segmentation, highlighting its potential for complex real-world applications. In addition, Sa2VA can be easily extended into various MLLMs, including Qwen-VL and Intern-VL, which can be updated with rapid process in current open-sourced MLLMs. Code and models have been provided to the community.

arXiv

Exploring learning environments for label\-efficient cancer diagnosis

Samta Rani, Tanvir Ahmad, Sarfaraz Masood, Chandni Saxena

Despite significant research efforts and advancements, cancer remains a leading cause of mortality. Early cancer prediction has become a crucial focus in cancer research to streamline patient care and improve treatment outcomes. Manual tumor detection by histopathologists can be time consuming, prompting the need for computerized methods to expedite treatment planning. Traditional approaches to tumor detection rely on supervised learning, necessitates a large amount of annotated data for model training. However, acquiring such extensive labeled data can be laborious and time\-intensive. This research examines the three learning environments: supervised learning (SL), semi\-supervised learning (Semi\-SL), and self\-supervised learning (Self\-SL): to predict kidney, lung, and breast cancer. Three pre\-trained deep learning models (Residual Network\-50, Visual Geometry Group\-16, and EfficientNetB0) are evaluated based on these learning settings using seven carefully curated training sets. To create the first training set (TS1), SL is applied to all annotated image samples. Five training sets (TS2\-TS6) with different ratios of labeled and unlabeled cancer images are used to evaluateSemi\-SL. Unlabeled cancer images from the final training set (TS7) are utilized for Self\-SL assessment. Among different learning environments, outcomes from the Semi\-SL setting show a strong degree of agreement with the outcomes achieved in the SL setting. The uniform pattern of observations from the pre\-trained models across all three datasets validates the methodology and techniques of the research. Based on modest number of labeled samples and minimal computing cost, our study suggests that the Semi\-SL option can be a highly viable replacement for the SL option under label annotation constraint scenarios.

arXiv

Image-Conditional Diffusion Transformer for Underwater Image Enhancement

Xingyang Nie, Caoliang Zhang, Xiaoyu Zhai, Fengzhong Qu, Biao Wang, Huilin Ge

Underwater image enhancement (UIE) has attracted much attention owing to its importance for underwater operation and marine engineering. Motivated by the recent advance in generative models, we propose a novel UIE method based on image-conditional diffusion transformer (ICDT). Our method takes the degraded underwater image as the conditional input and converts it into latent space where ICDT is applied. ICDT replaces the conventional U-Net backbone in a denoising diffusion probabilistic model (DDPM) with a transformer, and thus inherits favorable properties such as scalability from transformers. Furthermore, we train ICDT with a hybrid loss function involving variances to achieve better log-likelihoods, which meanwhile significantly accelerates the sampling process. We experimentally assess the scalability of ICDTs and compare with prior works in UIE on the Underwater ImageNet dataset. Besides good scaling properties, our largest model, ICDT-XL/2, outperforms all comparison methods, achieving state-of-the-art (SOTA) quality of image enhancement.

arXiv

STA-VPT: SpatioTemporally Aligned Visual Prompt Tuning

Wenjie Pei, Tongqi Xia, Qizhong Tan, Jiandong Tian, Guangming Lu, Jun Yu

Typical methods for visual prompt tuning follow the sequential modeling paradigm originating from NLP, learning a sequence of unordered parameterized tokens as visual prompts, which are then prefixed to the flattened image representation for model adaptation. While such a sequential prompting paradigm has exhibited great promise, it presents two potential limitations. First, the learned visual prompts, presented in an unordered sequential form, are unable to capture the underlying spatial relations in the input image which are crucial for effective image encoding. Second, all prompt tokens serve the same role by performing uniform prompting for all image tokens without distinction, lacking fine-grained prompting capability\---i.e., individualized prompting for different visual tokens to capture region-specific semantic patterns. In this work, we introduce the \mymodel model (\emph{STA-VPT}), a novel visual prompting paradigm, which learns a two-dimensional prompt token map for image prompting or a three-dimensional token volume for video prompting, ensuring spatial (or spatiotemporal) alignment with the input image token map (or video token volume). This alignment enables the visual prompts to preserve the spatial (or spatiotemporal) structure, thereby learning the underlying relations within the visual input. Furthermore, each prompt token serves as a specialized prompting expert and is designated to exclusively prompt for the spatially (or spatiotemporally) corresponding visual tokens. Consequently, our \emph{STA-VPT} is capable of performing individualized prompting for different spatiotemporal regions, potentially improving the prompting performance through fine-grained allocation of prompting capacity, in line with mixture of experts (MoE) principles.

arXiv

Calibrating Perception Uncertainty for Autonomous Driving

Markus Kängsepp, Meelis Kull

Autonomous driving systems must be capable of making quick decisions based on the perceived environment and specific driving conditions. Perception models in these systems perform well in detecting objects under favourable conditions but their performance deteriorates in poor visibility or with partly occluded objects. To reduce risks from undetected objects, autonomous vehicles must incorporate all relevant uncertainties into their decision-making processes. Grid-based perception outputs, such as occupancy grids, and object-based outputs, like lists of detected objects, must be accompanied by well-calibrated uncertainty estimates. These uncertainties are essential for ensuring the model's reliability and safety. In this paper, we identify limitations in the current state-of-the-art and propose a more comprehensive set of uncertainty estimates that should be reported. In addition to commonly estimated forms of uncertainty about the presence, location, shape, and trajectory of detected objects, we propose to quantify the uncertainty about undetected objects within a region. Access to this set of uncertainties enables planners to perform region occupancy queries, which provide the probability that a certain region, such as the area around a chosen trajectory, is free of obstacles. We propose a novel approach for generating these probabilistic outputs from bird's-eye-view (BEV) probabilistic semantic segmentation. Our experiments demonstrate that the initial probabilistic outputs from segmentation are not calibrated, and we present methods to achieve well-calibrated uncertainty estimates. Finally, we conduct an experiment on a downstream task, underscoring the importance of calibrated uncertainties for planning and highlighting the advantages of including additional uncertainty types.

arXiv

Mining Artifacts in Mycelium SEM Micrographs

Thaicia Stona de Almeida

Mycelium is a promising biomaterial based on fungal mycelium, a highly porous, nanofibrous structure. Scanning electron micrographs are used to characterize its network, but the currently available tools for nanofibrous microstructures do not contemplate the particularities of biomaterials. The adoption of a software for artificial nanofibrous in mycelium characterization adds the uncertainty of imaging artifact formation to the analysis. The reported work combines supervised and unsupervised machine learning methods to automate the identification of artifacts in the mapped pores of mycelium microstructure. Keywords: Machine learning; unsupervised learning; image processing; mycelium; microstructure informatics

arXiv

A Reranker for Orchestrating Heterogeneous Speech and Text Retrievers

Inho Kim, Sumyeong Ahn

Retrieval-Augmented Generation (RAG) systems have attracted significant interest for their ability to mitigate hallucinations in Large Language Models (LLMs). Although knowledge databases for RAG are increasingly diversifying to include various modalities such as speech and text, research on handling such multi-modal database scenarios remains limited. In this paper, we propose STeReO (Speech and Text Reranking Orchestrator), a reranker based on speech and text retrievers that aggregates disparate modality databases. To address the lack of specialized training data, we first curate a dataset comprising queries, mixed-modality evidence, and their corresponding relevance ranks. We then train the reranker and evaluate its effectiveness in both single-modality and mixed-modality scenarios. Our results demonstrate that the proposed algorithm excels at selecting the most relevant evidence, thereby significantly improving downstream question-answering performance.

arXiv

Comparing Chunking and Embedding Strategies for Turkish RAG Systems

Mustafa Sertaç Türkel, Fatma Nur Korkmaz, Ahmet Tuğrul Bayrak

Retrieval-Augmented Generation conditions a language model on chunks retrieved from a document collection. Its accuracy is therefore limited by the chunking and embedding stages that determine what can be retrieved. We compare Turkish document question answering across three chunking strategies (fixed-length, semantic, and layout-aware Docling), five embedding models, and two LLMs, over three documents with contrasting layouts. Every configuration answers the same question set, which allows component effects to be separated by paired testing rather than inferred from separate benchmarks. The fully crossed design yields 9{,}000 graded question-answer evaluations, each scored by an independent judge model, and component comparisons are tested by paired McNemar tests under Holm correction. The three leading embedding models are statistically indistinguishable, so language specialization yields no measurable retrieval advantage. The faster LLM is not the more accurate one. The preferred configuration depends on content type, since layout-aware chunking helps table-heavy documents far more than text-heavy ones.

arXiv

The Dialect Tax: Dialectal Biases Persist throughout the Language Modeling Pipeline

Elle

Systematic dialectal performance gaps in language models (LMs) are well documented, but the source of these disparities within the modern language modeling pipeline remains unclear. Our study traces this "dialect tax" across the natural language processing pipeline. Using parallel English dialect corpora that hold meaning fixed while varying surface form, we first confirm that LMs recognize matched Standard American English (SAE) and dialectal texts as semantically equivalent. However, we discover further representational gaps corresponding to downstream performance gaps. Across model families and generations, modern LMs still encode dialectal texts unequally during tokenization, pre-training, post-training, and inference. Strikingly, bypassing traditional subword segmentation via a character-level counterfactual tokenizer removes neither input and output asymmetries nor dialectal accuracy gaps. During pre-training, dialect pairs induce more divergent gradient updates than pairs of entirely unrelated SAE documents, indicating that models find semantically equivalent dialectal content harder to learn from than unrelated SAE documents. During post-training, reward models show contextual, unstable dialect preferences, assigning higher values to isolated AAVE-exclusive tokens than to SAE-exclusive tokens, while full reasoning contexts receive task- and model-dependent dialect penalties. Overall, our findings suggest that the dialect tax is encoded and accumulated not by any one step in isolation, but at every step of the language modeling process.

arXiv

Demystifying Reinforcement Learning Post-Training of Language Models

Donovan Clay, Saket Gollapudi, Sankar Harilal, Min Jang, Jacob Morrison, Sewoong Oh, Natasha Jaques

Reinforcement learning (RL) post-training has emerged as a powerful framework for enhancing the capabilities of large language models (LLMs), enabling impressive reasoning, math, and coding capabilities. Yet for many researchers and practitioners, the principles behind classical RL remain a "black box". In this work, we deconstruct the RL post-training algorithm, investigating each step to clarify what is actually happening beneath the surface. By isolating the mechanics of RL with Verifiable Rewards in a controlled and simplified environment, we examine how RL outcomes are shaped by the base model's prior distribution, the granularity of the reward signal, the diversity of the prompt distribution, and model scale. We use the entropy of the policy's output distribution as a lens to compare the distributions learned through pretraining, SFT, and RL post-training, revealing how each stage shapes model certainty. Our investigation sheds light on how these choices interact to affect post-training success. For example, we show that the effect of so-called 'spurious rewards' depends on the prompt distribution used for post-training. We also provide insight into why the success of RL post-training depends on whether the base model already places sufficient probability mass on the desired behavior, linking it to the classical concept of exploration in RL. Ultimately, we provide this primer as a resource to those in the NLP community wishing to incorporate RL as a tool in their toolbox.

arXiv

Wontopos Tablet 2: Measuring Multilingual and Multimodal Memory Retrieval Without Lexical Matching

Sunwoo Kim

We measure tablet-2, a production long-term memory engine for language models, on the text benchmarks the field already uses and on cross-lingual retrieval of photographs stored with no text at all. Its retrieval path contains no lexical matching, no keyword scoring, and no language model of its own. On LongMemEval-S (500 questions) it scores 95.7% [93.4, 97.1]; on BEAM-1M (700 questions, 2.21M stored memories) 67.5% [64.8, 70.2]. Those are question-sampling intervals, not the run-to-run spread, which is an order of magnitude narrower. Most of the paper is about how little they mean alone. Holding engine, corpus, settings and judge fixed, changing only the reader moves LongMemEval-S by 2.0 points; changing only the re-ask budget moves BEAM-1M by 8.9. Neither is stated in the reports we compare against, and the second exceeds most gaps there, so we give that table as a placement and not a ranking. For the multimodal axis we run two controls. Against BM25, configured as strongly as we could, we reach 95.2% mean recall@5 over 70 store-and-query language cells where BM25 reaches 19.0% and is exactly zero in 54. On captionless photographs a lexical method has no document to score at all. Open dense baselines on 300 Crossmodal-3600 photographs in 14 languages show that density confers no language independence: one scores 91.0% on English and 4.7% on Russian from identical image vectors, and a multilingual variant collapses on Telugu and Swahili. Our spread across languages is 14.0 against their 27.5 and 27.7. Three results run against us and are reported at equal weight: low-resource languages degrade sharply (Swahili 53.0%, Telugu 64.0%), attaching captions lowers cross-lingual retrieval by 11.4 points, and one setting omitted into one stage of our own retrieval cost 37 points of Korean top-1 accuracy while leaving nine languages untouched.

arXiv

Continual Visual Learning under Evolving Semantic Concept Shift

Ismail Lamaakal, Chaymae Yahyati, Yassine Maleh, Khalid El Makkaoui, Ibrahim Ouahbi

Visual foundation models are commonly adapted under the assumption that the appearance of incoming data may change while the semantic meaning of the prediction task remains fixed. In long-lived visual systems, however, taxonomies, policies, and concept definitions can themselves evolve, causing the same visual evidence to require a different interpretation. We study this setting as evolving semantic concept shift and introduce SemReWrite, a framework for selectively updating obsolete visual--semantic mappings while preserving knowledge that remains valid. SemReWrite represents changes between old and revised semantic specifications, combines semantic discrepancy with sparse revised supervision to localize affected visual regions, and uses an input-dependent low-rank rewriting mechanism together with structured semantic memory, preservation, and obsolete-decision suppression. We further introduce EvoShift-Bench, spanning ImageNet, iNaturalist, CUB-200-2011, and DomainNet, with semantic transitions including class split, merge, boundary revision, insertion, partial redefinition, recurrence, and mixed semantic--appearance shift. To explicitly evaluate selective semantic revision, we introduce Rewrite Accuracy (RA) and Preservation Accuracy (PA) for affected and unaffected regions, respectively, Obsolete Retention (OR) for measuring residual outdated semantic associations, and the Selective Revision Score (SRS), which jointly summarizes rewriting and preservation performance. Experiments show that SemReWrite achieves a stronger balance between learning revised semantics and retaining unaffected knowledge than prompt replacement, conventional fine-tuning, parameter-efficient adaptation, and continual-learning strategies.

arXiv

Names Can Hurt: Spotting Slopsquatting Risks Caused by Package Name Hallucinations in Local Coding LLMs

Akash Raj, Sargam Sahu

When a code generating language model fabricates a Python package name, an adversary who has pre-registered that name on PyPI can convert that hallucination into a supply chain compromise. This event has been termed as 'slopsquatting'. We propose a two layer detector to counter this issue. The first layer performs a deterministic PyPI existence check. The second is a Random Forest classifier trained on ten features derived from the package name and its PyPI metadata. An import name reconciler bridges the two, resolving cases such as 'import cv2' versus 'pip install opencv-python' without a security bypass. The detector is embedded in a LangGraph state machine that retries at escalating temperatures and, on repeated failure, routes to a stronger fallback model. Across 300 curated prompts, the pipeline produces hallucination free code on 76% of runs. The primary exhausts its retry budget on 28.7%; intra model retries recover roughly a quarter of those, and cross model fallback recovers a further 16.5% of the remainder. Four findings have been observed. First, half of the flagged hallucinations are packages already registered on PyPI, as low quality lookalikes of well known projects, caught by the classifier rather than the deterministic layer (e.g., pil, faiss, tabula, haystack). Second, hallucination rate scales almost linearly with prompt adversariality, from 0 to 10% on routine coding to 40 to 73% on slopsquat baits. Third, the weaker primary refused 6 of 10 direct baits unaided, suggesting recent instruction tuning provides a baseline defense. Fourth, when primary and fallback share a model family, approximately 84% of primary failures recur on the fallback, motivating cross family pairing. A user study (n = 24) reports mean satisfaction 4.4 out of 5 and 21 of 24 stated adoption intent.

arXiv

Semantic Overlays: Mitigating Prompt Injection with Annotations Beyond Tokens and Steering Vectors

Joshua Penman

Everything a language model sees is tokens. The serving stack knows what each span is -- user input, tool output, instructions -- but the model must keep track of that itself, and can lose track or be confused: text can be written to read like anything. Prompt injection is a natural exploit of this phenomenon. By scrambling the model's understanding of span identity, an attacker can induce unwanted and dangerous actions. Adding a non-textual channel to the model's input -- a way to communicate span identity beyond text -- mitigates this class of attack. We thus introduce a general steering technique called Semantic Overlays: small learned adapters applied at chosen prefill positions to a frozen model's residual stream. Laying an overlay over a span creates an out-of-band annotation channel that cannot be replicated by tokens. Unlike steering vectors, Semantic Overlays are trained, adaptable, and selectively applied. An overlay can encode complex semantics that reshape how the model perceives the marked span: asked to copy a code snippet under an overlay asserting a different programming language, the model rewrites the snippet in the asserted language. Overlays compose, allow transparent reading of underlying content, and can carry complex payloads -- including imperatives the model will follow. An overlay which marks a span as "non-executable" defends against the broad class of prompt injections that add instructions in untrusted context. We report strong results on five prompt injection benchmarks: SEP separation rises from 24.3% to 99.0% with utility unchanged (our scoring rule; we correct a defect in the published grader), TensorTrust attack success falls from 34.8% to 6.2%, AlpacaFarm from 99.0% to 0%, and the overlay beats every published PIArena defense that leaves the model able to answer -- while marked spans stay readable, all at >95% character similarity to the original.

arXiv

Markets, Not Planners: Decentralized Orchestration of LLM Agents with Private Information

Xiao Liu, Haoyang Li, Songwei Li, Hongbo Fang, Fengli Xu, Feng Shi, James Evans

As LLM agents proliferate, built by different parties and with different capabilities and costs, orchestrating them is more like assembling labor across the economy than a computer calling a subroutine. Existing orchestration is typically centralized, with a single planner assigning every task, but this creates a bottleneck as agent pools grow, requires private information (e.g., agents' execution costs), and can easily be manipulated, such that a single inserted preference nearly doubles a favored agent's task share under a centralized LLM allocator. We introduce AgentLance, a repeated labor market in which agents bid on tasks using their private costs and self-maintained strategy notes, an allocator selects winners from bids and public reputation records, and a VCG-style payment rule rewards cost-aware bidding. Complex tasks are handled by hierarchical delegation: winning agents can decompose work and subcontract it through the same mechanism. Across mathematical reasoning, code generation, knowledge-intensive QA, and agentic tasks, AgentLance matches agents to their specializations, shifts work toward cheaper agents as cost sensitivity rises, and consistently outperforms single-model, centralized-orchestration, and market baselines. Diagnosing market failures, including inaccurate cost self-estimation and sub-optimal bidding, then correcting them in controlled experiments yields further gains, charting a path toward more efficient agent economies.

arXiv

Does Episodic Memory Help Close the Lexical Frequency Gap in Sensitivity to Syntactic Contrasts? A Test Using Retrieval-Augmented Language Models

Jing Liu, Najoung Kim

Grammatical knowledge and how it is empirically tested are typically considered robust to the frequency of the lexical items in the expressions. However, neural network-based models of grammaticality exhibit high sensitivity to lexical frequency. We draw upon Complementary Learning Systems theory to test the hypothesis that robustness to lexical frequency can arise via a hippocampal episodic memory mechanism, which enables rapid encoding and retrieval of specific experiences and allows learners to leverage them when processing rare patterns. We use retrieval-augmented language models as an instantiation of such an episodic memory mechanism (specifically, $k$-nearest-neighbor language models that augment parametric models with explicit instance storage), and test whether this augmentation helps close the lexical frequency gap that vanilla language models exhibit in syntactic contrast tests. Using syntactic contrasts with frequency-stratified test items, we find that retrieval augmentation narrows the performance gap between high- and low-frequency items, consistent with episodic memory compensating for weak parametric representations. This benefit is consistent across different syntactic phenomena and across models pretrained on child-realistic and large-scale data. Additionally, we show that structural information is critical for effective retrieval, whereas semantic similarity alone provides little benefit. While these are promising proof-of-concept results supporting our hypothesis, the frequency gap is narrowed rather than fully closed. Based on our analyses, we propose preferential reweighting of retrieved instances, better representations and retrieval strategies for structural information, and flexible configurations of storage and retrieval as promising future directions for improving the implementation of episodic memory in language models.

arXiv

Mitigating Exploration Bias in RL for Multi-Instruction Following

Mian Zhang, Yueqin Yin, Kaiyu He, Peilin Wu, Xinlu Zhang, Mingyuan Zhou, Zhiyu Zoey Chen

RL has emerged as a powerful paradigm for enhancing the instruction following capabilities of LLMs. While existing training recipes achieve substantial gains, we find that they suffer from exploration bias towards easy instructions when the training data has multiple instructions in a prompt. This bias is caused by two main reasons: 1) the policy model's initial ability to satisfy hard instructions is too low to trigger successful exploration during RL training, so the optimization is biased towards easy instructions; and 2) canonical RL training recipes typically employ a cumulative reward (the number of instructions fulfilled), treating all instructions equally, which biases the policy model towards fulfilling easy instructions to obtain the same amount of reward. To address these issues, we first propose two metrics to measure the exploration bias in instruction following and then introduce a two-stage framework to alleviate it: 1) Behavioral Bootstrapping, a lightweight rejection sampling fine-tuning stage before RL to activate hard instructions; and 2) Scarcity-Aware Rewards, a new RL reward function that assigns rewards to instructions based on their empirical scarcity. Experiments show that the proposed metrics are highly correlated with model performance, and our methods unleash the potential of RL training: our best models outperform the baselines by a significant margin across three verifiable instruction following benchmarks. We release codes at https://github.com/mianzhang/MulIF.

Keep browsing

More papers from top AI conferences

30+ top conferences, by conference, year, and topic.