跳到正文
研踪PaperTrail

论文库

检索计算机领域论文,按研究方向、作者或关键词筛选。

部分采集异常
正在读取账号与研究记录…

全站更新概况

+0新增论文
+0其中 AI 相关
+0已有论文修订
+92新增 AI 导读

arXiv 最近完成扫描:2026-11-23 01:07 · 全站最近入库:2026-11-21 12:19各信源进度 →

统计口径与数据新鲜度

统计 2026-11-22 06:53 — 2026-11-23 06:53(北京时间) · 点击数字查看对应简报;AI 相关包含在新增论文内。

最近完成检查:2026-11-23 01:07最近发现新增版本:2026-11-21 12:19库内论文最新更新时间:2026-11-20 02:59

每分钟刷新已保存的数据。空检查不代表采集异常;新收录可能包含历史论文。AI 导读按新增版本计数。

29,158 已收录论文24,181 AI 相关8,992 有修订112 HF 社区发现

浏览论文库

时间范围跟随排序:版本更新、本站收录或首次发表。最新收录可能包含历史论文。部分信源异常,查看状态 →

找到 6,576 篇论文 · 按版本更新时间排序上轮扫描起点 2026-11-23 01:05(北京时间)
新论文 · v1arXiv:2611.24626

Transferring an analysis of human brain organization to language models

Donghyeok Lee, Sangwon Yu, Sangkwon Park, Hye Youn Park, Munyoung Chang, Sungroh Yoon

原文摘要

During human cortical development, functional systems become increasingly differentiated. We asked whether improved reasoning in language models involves a comparable regional reorganization. Applying diffusion-map embedding, used to characterize cortical organization, to matched base and reasoning-trained models from three families, we found broadly preserved gradient profiles with localized changes within a contiguous layer span. Grafting weights from these spans into instruction-tuned hosts improved mathematical reasoning on average, with gains rising with region overlap, while retaining more previously solved problems than grafts elsewhere. Across training, the span formed gradually during the early rise in performance, and grafts from intermediate stages helped. Grafts shifted most hosts partly towards their donors' behaviour and sometimes improved on problems the donors failed; the same analysis located a transferable span in a medical model. Together, these findings identify a reasoning-related region whose weights improve reasoning with low interference in the host's existing mathematical competence.

cs.LG · cs.CL

新论文 · v1arXiv:2611.24598

Chain-of-Thought Hidden States Reveal Task-Solving Directions for Direct Answers

Houman Mehrafarin, Ioannis Konstas

原文摘要

Chain-of-Thought (CoT) prompting improves the performance of Large Language Models by eliciting intermediate reasoning steps. However, CoT is far from perfect: on some tasks, models perform worse than with direct-answer prompting, and on others they generate a few hundred tokens to reach the correct answer. We start by labelling each CoT hidden state as recovering if patching it into the direct prompt yields the correct answer, and failed otherwise. A linear probe shows that the model encodes these two classes differently. We then show that the representation space of the direct prompt holds a direction for each question that leads the model, in a short generation, to the correct answer more often than CoT does. Specifically, we rotate the direct prompt's last hidden state once, before generation, towards a direction built for each question from its own recovering and failed states. Across 25 tasks, these directions raise the direct prompt's mean accuracy from $46\%$ to $85\%$ on LLaMA 3.1 8B (CoT: $69\%$) and from $52\%$ to $86\%$ on Qwen 2.5 7B (CoT: $77\%$), with a tenth of CoT's tokens. CoT is thus not the shortest route to this accuracy, and these directions can be used in future work to train models to align their states with them.

cs.CL · cs.AI

新论文 · v1arXiv:2611.24588

Large Language Models or AI Agents for English to Arabic Translation? An assessment on WikiEn-Ar190

Fatima Zahra Iguenfer, Yousra Chtouki

原文摘要

This research investigates whether a large language model or an autonomous AI Agent performs better at translating from English to Arabic, which is considered a low-resource language. The authors introduce WikiEn-Ar190, a new dataset of 190 carefully selected Wikipedia articles from 18 different areas, which were created and verified by two native Arabic speakers to ensure that they accurately reflect both meaning and style. Using this dataset, the study evaluates the translation quality of GPT-5-mini both as an independent neural translator and when used within a more complex process that includes self-assessment and improvement. As a standalone model, GPT-5-mini produced very fluent translations , achieving a COMET score of 0.8649 and receiving high human scores for adequacy, fluency, and style (averaging 4.53 out of 5). When used in a system that combined translation with self-assessment, it improved lexical accuracy by 28.76% (BLEU score of 22.83) while maintaining a similar level of semantic quality. These findings show that methods involving autonomous reasoning can improve translation consistency for language pairs with limited resources. Beyond the results, the approach offers a framework for evaluating and improving translations, potentially using them as assisting tools to expand the representation of low resource languages, such as Arabic, on the web.

cs.CL

新论文 · v1arXiv:2611.24576

Late Entry in Looped Language Models

Ziqi Jin, Jiaxi Li, Weiwen Xu, Aixin Sun

原文摘要

Looped language models deepen reasoning at a compact parameter count by iterating a shared layer stack T times per token, turning recurrence depth directly into sequential decoding latency. Conventional early-exit decoding halts after initial passes without cache completion, truncating cached key-value states of past tokens to shallow depth. We study the opposite strategy, late entry, where new tokens skip initial passes and enter directly into the final K < T passes over a deep context cache. Once lightweight fine-tuning resolves hidden-state norm mismatch at the entry boundary, deep cached context substitutes for skipped token-local passes, allowing late entry to outperform matched early exit across model families and pass budgets. To turn this asymmetry into wall-clock speedups without stalling to build deep caches, candidate blocks verify through only the final K passes on the critical path while previously committed tokens complete passes 1 -> T alongside subsequent steps at zero extra serial calls. Projecting these finished T-pass wake states inside the existing output GEMM (Wake verification, governed by logit-margin tolerance tau in [0, inf]) unifies greedy-exact T-pass decoding (tau = 0, matching K = T block decoding within bf16 reduction order), margin-gated verification (tau > 0), and non-rollback lossy late entry (tau = inf) inside a single static CUDA graph. On Ouro-2.6B (T = 4), K = 2 Wake reaches 5.2x greedy-exact speedup at tau = 0 (+26% over K = 4 block decoding on GSM8K) and 6.8x at tau = inf (+16.9 points over trained early exit at matched speed). Scaling test-time recurrence across T in {8, 16, 32, 64} on Huginn-3.5B cuts K = T/4 rollbacks 21.5x with depth, driving greedy-exact (tau = 0) speedups to 6.1x (T = 32) and 8.4x (T = 64, +153% over K = 64 block) and tau = inf speedups to 10.6 to 12.5x.

cs.CL

新论文 · v1arXiv:2611.24567

Objective-Specialized Residual Streams in Manifold-Constrained Hyper-Connections

Shantanu Agarwal, Ruchita V. Maitri, Zhiqi Huang, Vivek Datla, Alfy Samuel

原文摘要

Manifold-constrained Hyper-Connections (mHC) extend a Transformer's residual stream into multiple parallel streams with learned cross-stream connections. Prior work trains all streams toward a shared objective. We instead study whether assigning distinct learning objectives to different streams can improve task performance and induce objective-specific specialization. Using complementary lexical (sparse) and semantic (dense) retrieval objectives as a test case, we find that objective-specialized mHC consistently outperforms a non-specialized mHC model across the full sparse--dense fusion range, for both two- and four-stream models and at two model scales. We then examine how specialization changes the model internally using probes of stream representations, residual mixing, objective-specific gradient flow, and read/write gates. The specialized and non-specialized models behave similarly through most of the network but diverge in the final layers: the specialized model develops increasingly objective-specific gradient flow, distinct read/write routing patterns, and reduced cross-stream mixing, and streams assigned to different objectives become increasingly dissimilar while those trained toward the same objective remain highly similar. Fixing the residual mixing to the identity leaves the specialization benefit intact, indicating that the benefit stems from objective-specific read/write routing rather than learned cross-stream mixing. These results show that assigning distinct objectives to mHC streams can improve multi-objective retrieval while inducing structured, depth-dependent specialization in the residual computation.

cs.CL

版本更新 · v3arXiv:2603.17102

Knowledge Localization in Mixture-of-Experts LLMs Using Cross-Lingual Inconsistency

Lucas Bandarkar, Alan Ansell, Trevor Cohn

原文摘要

Modern LLMs continue to exhibit significant variance in behavior across languages, such as being able to recall factual information in some languages but not others. While typically studied as a problem to be mitigated, in this work, we propose leveraging this cross-lingual inconsistency as a tool for interpretability in mixture-of-experts (MoE) LLMs. Our knowledge localization framework contrasts routing for sets of languages where the model correctly recalls information from languages where it fails. This allows us to isolate model components that play a functional role in answering about a piece of knowledge. Our method proceeds in two stages: (1) querying the model with difficult factual questions across a diverse set of languages to generate "success" and "failure" activation buckets and then (2) applying a statistical contrastive analysis to the MoE router logits to identify experts important for knowledge. To validate the necessity of this small number of experts for answering a knowledge question, we deactivate them and re-ask the question. We find that despite only deactivating about 20 out of 6000 experts, the model no longer answers correctly in over 40% of cases. Generally, this method provides a realistic and scalable knowledge localization approach to address increasingly complex LLMs.

cs.CL · cs.AI · cs.LG

版本更新 · v6arXiv:2602.02898

Aligning Language Model Benchmarks with Pairwise Preferences

Marco Gutierrez, Xinyi Leng, Hannah Cyberey, Jonathan Richard Schwarz, Ahmed Alaa, Thomas Hartvigsen

原文摘要

Language model benchmarks are pervasive and computationally-efficient proxies for real-world downstream performance. However, many recent works find that benchmarks often fail to predict downstream utility. While some works have begun diagnosing sources of misalignment, there remain no ways to systematically update benchmarks to align their scores with downstream usage. Towards bridging this gap, we introduce and study \textit{benchmark alignment}, where we use information about downstream model performance to automatically update benchmarks, specifically aiming to update static benchmarks so they generalizably rank models according to new pairwise preferences. Our experiments involving 4576 language models and 6 benchmarks show that reweighting benchmark items can successfully rank unseen models, even generalizing across model scales in most cases. And while naive alignment unsurprisingly requires large numbers of models and benchmark questions, an oracle experiment suggests this could be reduced to as few as 20 well-chosen models. Overall, our work takes a step towards efficiently aligning benchmark development with downstream tasks. All of our code, models, and data are publicly-available: https://github.com/hartvigsen-group/benchalign

cs.AI · cs.CL

版本更新 · v2arXiv:2504.15270

Quicksviewer: An LMM with Early Video Compression using Annealed Gumbel-Softmax

Ji Qi, Yuan Yao, Yushi Bai, Bin Xu, Juanzi Li, Zhiyuan Liu, Tat-Seng Chua

原文摘要

Large Multimodal Models (LMMs) uniformly perceive video frames, creating computational inefficiency for videos with inherently varying temporal redundancy. This paper present **Quicksviewer**, an LMM with new perceiving paradigm that partitions a nonuniform redundant video into varying cubes using Gumbel Softmax, followed by a unified resampling for each cube to achieve efficient video understanding. We further propose a Gumbel-noise annealing strategy that enables stable end-to-end LMM training with the discrete cubing network. This simple and intuitive approach dynamically compress video online based on the momentum of frame differences, significantly reducing spatiotemporal complexity (overall 45$\times$ compression rate), while enabling efficient training with large receptive field. We train the model from a language backbone through three progressive stages, each incorporating lengthy videos on average of 420s/1fps thanks to the perceiving efficiency. With only 0.8M total video samples, our model outperforms contemporary LMMs with the same backbone and comparable parameters amount. An experiment in A/B pretraining with same backbone and training data shows that, w/ cubing achieves lower training loss and higher benchmark accuracy (+2.1) than the fixed partitioning baseline, strictly validating the benefits. Moreover, the front-loaded compression in encoding substantially reduces latency in the computationally expensive prefill and decoding, making it suited to mainstream EPD disaggregated inference deployment. Code: https://quicksviewer.github.io.

cs.CV · cs.CL

版本更新 · v4arXiv:2606.02502

CRAM: Centroid-Routing and Adaptive MoE for Multimodal Continual Instruction Tuning

Jun-Tao Tang, Zhen-Hao Xie, Yu-Cheng Shi, Da-Wei Zhou

原文摘要

Multimodal Large Language Models (MLLMs) unify heterogeneous vision-language tasks under a shared generative framework via instruction tuning, yet real-world deployment demands continuous capability expansion, making Multimodal Continual Instruction Tuning (MCIT) essential. Existing methods either update all tasks with a shared parameter set or allocate dedicated modules for each new task. Shared updates force heterogeneous tasks to compete, causing forgetting of learned capabilities. Conversely, isolated expansion prevents interference but severely limits parameter efficiency over long task streams. To address this dilemma, we propose CRAM (Centroid-Routing and Adaptive MoE). Specifically, by isolating task-specific patterns into independent modules, CRAM mitigates catastrophic forgetting across tasks. To further boost parameter efficiency, we utilize adaptive-rank instantiation to identify the capability gap between existing expert capability and new task demands, and dynamically allocate only the necessary parameters. To ensure stable reuse among tasks, centroid-guided routing recognizes and activates existing experts' capabilities, while an orthogonality penalty confines new updates to task-specific directions, preventing re-learning general capability. Extensive experiments across diverse benchmarks demonstrate its superiority over existing methods. Code is available at https://github.com/LAMDA-CL/EMNLP2026-CRAM.

cs.CL

新论文 · v1arXiv:2611.24540

Do Frontier Model Ingredients Add Up? A Replicated Factorial Study of Small Language Models

Biswa Sengupta

原文摘要

Building a small language model requires choosing which published improvements to combine. We compare five changes in the same model, using the same data and training budget. We test selected combinations and repeat each one with a second random seed. Mixture-of-experts layers, Muon optimisation and a Gated DeltaNet hybrid give the largest gains. Extra normalisation also helps, while an attention gate with a matched parameter budget adds little on average. The gains mostly combine, but small interactions between changes appear again in the repeated runs. Repetition also reveals benefits that an initial statistical estimate treated as noise. These findings apply to our small-model training setting; the expert models also have greater total capacity. The practical lesson is to test promising components together and repeat the comparisons before deciding which ones to keep.

cs.CL · cs.AI

新论文 · v1arXiv:2611.24535

ProbMedTOD: A Bayesian Network Guided Task-Oriented Dialogue System for Patient History Taking

Vishal Vivek Saley, Bhavesh Gurnani, Dinesh Raghu, Mausam

原文摘要

Patient history taking is a strategic, multi-turn process that refines diagnostic hypotheses through iterative questioning, mirroring clinical reasoning where each inquiry updates beliefs over candidate diagnoses. Deploying AI in this setting is constrained by privacy, cost, and latency, making Small Language Models (SLMs) a more desirable backbone compared to larger LLMs. However, SLMs capture only the fast, intuitive "System 1" side of this process and lack explicit mechanisms for the deliberative "System 2" reasoning needed to maintain and update diagnostic uncertainty over time. Bayesian Networks (BayesNets) offer a natural complement, providing an interpretable framework for auditable probabilistic belief tracking. Yet, clinical BayesNets are difficult to construct due to their reliance on expert curated structure, dependencies, and parameters. We introduce ProbMedTOD, which automatically synthesizes a clinical BayesNet from a small number of clinical notes in three stages: ontology building, structure prediction, and parameter estimation. ProbMedTOD then integrates the resulting BayesNet with SLMs, enabling probabilistic reasoning in multi-turn dialogues. A BayesNet Agent maintains real-time posterior distributions over diagnoses, while a Policy Agent picks diagnostically informative questions. Experiments on DDXPlus and MIMIC-IV Notes show that ProbMedTOD outperforms LLM-only and retrieval-based baselines, with especially strong gains for smaller models.

cs.CL

新论文 · v1arXiv:2611.24522

MECI: Multi-Element Collaborative Interaction for Multimodal Entity Linking

Jie Peng, Yongxue Shan, Yongfu Zha, Xiaodong Wang

原文摘要

Multimodal Entity Linking (MEL) aims to disambiguate mentions in multimodal contexts by grounding them to specific entities in a knowledge base. A pivotal challenge in MEL is capturing multi-level correspondences: the semantic consistency between mention-entity pairs and the complementary correlations across modalities. However, existing methods often suffer from element dominance due to their reliance on coupled interactions or coarse global aggregations. In response, we propose the Multi-Element Collaborative Interaction (MECI) framework. First, to capture multi-element mention-entity correspondences, we develop a Multi-view Experts Network that leverages a "divide-and-conquer" strategy for decoupled feature learning to mitigate element dominance, supported by a KL-guided routing mechanism that governs expert specialization and collaboration. Furthermore, to model cross-modal complementary correlations, we propose a Hierarchical Multimodal Interaction Module, where a dynamic modality-aware weighting network refines interactions across hierarchical semantic levels, thereby integrating multi-granular evidence to counteract element dominance. Finally, we incorporate a generative semantic refinement stage that utilizes large language models for zero-shot re-ranking. Extensive experiments on WikiDiverse, RichpediaMEL, and WikiMEL show that MECI consistently outperforms state-of-the-art baselines, improving Hits@1 by 1.95%, 7.30%, and 2.31%, respectively.

cs.CL · cs.AI

版本更新 · v3arXiv:2611.18397

Gamer: A Generalist Vision-Language-Action Foundation Model for Long-Horizon Real-Time Game Agent

Run Luo, Kuan Zhang, Qi Deng, Yukun Chen, Zhihao Yang, Yufei Zhang, Renrong Hu, Yi Zhang 等 27 位作者

原文摘要

We introduce Gamer, a generalist agent framework built for long-horizon tasks in complex interactive environments. Its end-to-end architecture combines perception, memory management, and multi-step reasoning. A shared training and deployment pipeline supports interaction with both conventional 2D graphical interfaces and complex 3D simulated worlds. We first continue pre-training on a dataset constructed from tens of thousands of hours of device-level keyboard and mouse interaction. These trajectories span simulated games and embodied environments, allowing the model to learn environmental dynamics across both physical and virtual domains. Sparse-thinking instruction tuning then substantially reduces redundant computation while preserving the reasoning needed for complex decisions. We train the policy for long-horizon interaction with a stable multi-turn reinforcement learning method that addresses finite context windows, accumulated rollout errors, and sparse rewards. For real-time interaction, a cloud-edge inference engine serves numerous simulated game environments concurrently, supports unbounded-context execution, and keeps online inference latency low. Across multiple challenging 3D open-world games, Gamer demonstrates solid generalization and operational efficiency comparable to skilled players. It also achieves competitive results on a range of conventional 2D game benchmarks. These results suggest that even repetitive keyboard and mouse trajectories can support substantial gains from scaling data and learning algorithms. Training on these trajectories helps Gamer learn complex skills more steadily in environments with both physical and digital properties. This approach offers a path toward generalist agents that can interact across digital and physical worlds.

cs.CL · cs.AI · cs.CV

新论文 · v1arXiv:2611.24994

What the model knows but does not say: monitoring medical LLMs through verbalized thoughts

Frederik Hauke, Jan Clusmann, Hanna Kreutzer, Christiane Kuhl, Tianyu Han, Felix Barajas Ordonez, Sven Nebelung, Jakob Nikolas Kather 等 9 位作者

原文摘要

Efforts to keep clinical language models safe have concentrated on what enters and leaves the model, but a manipulated model can return an answer that looks entirely ordinary. Here we show that the model's own internal state, rendered in plain language, exposes manipulations the output conceals. Using a natural-language autoencoder, we translate the activations of a multimodal LLM (Gemma-3-27B) into text and let an input-blind judge read this verbalized internal state (the thoughts channel) and the model's answer separately. Across two clinical attack objectives (omitting a pathology, covertly recommending a drug), two input modalities, and two delivery mechanisms, the thoughts channel caught the injection significantly more often than the answer in all eight injected conditions (p<0.05), while both channels held near-zero false-positive rates on clean inputs; the omission result carried over to Llama-3.3-70B. The thoughts channel also registered unsafe clinical requests as harmful while the model was still reading them, before it decided whether to refuse. None of this requires retraining: an off-the-shelf verbalizer and a general-purpose judge are enough.

cs.CL · cs.AI

新论文 · v1arXiv:2611.24475

Nothing to Declare: Steering a single neuron increases concealment in AI teams

Edoardo Chidichimo, Felix P. S. Wallis

原文摘要

Recent research suggests that individual neurons can have interpretable roles in AI behaviour. In Qwen3.5-4B, steering along one neuron's output direction increased decisions to conceal information more in teams than alone across five scenarios, even when teammates provided no messages. When all members participated and received steering, collective concealment increased in four scenarios, while inter-agent messaging helped temper concealment under steering. Moreover, removing the neuron's contribution had negligible effects. Collective settings can therefore change how an internal intervention affects AI behaviour.

cs.CL · cs.AI

版本更新 · v4arXiv:2603.22665

Improving LLM Final Representations with Inter-Layer Geometry

Tom Ulanovski, Eyal Blyachman, Maya Bechler-Speicher

原文摘要

The standard in LLM-based prediction is to use the final-layer representation as the input to a downstream predictor. However, intermediate layers may encode complementary task-relevant signals. Existing approaches therefore either search for the best layer for each task or apply expensive attention-based mechanisms to learn inter-layer aggregation. In this work, we first show that such complexity is unnecessary: a lightweight Graph Neural Network over a fully connected graph of LLM layers is more efficient and achieves significantly stronger predictive performance than existing approaches. We then introduce the Cayley-Encoder, which further improves both efficiency and predictive performance by replacing the fully connected graph with a Cayley graph over SL(2, Zn). These Cayley graphs provide a mathematically grounded topology that is sparse, regular by construction, and has low diameter. This enables effective communication across layers while constraining the aggregation structure and reducing the risk of GNN overfitting. In an evaluation of Cayley-Encoder across 13 tasks and 13 LLMs, Cayley-Encoder consistently outperforms baselines, achieving improvements of up to 48 percentage points in accuracy, while introducing at most 0.1% additional parameters relative to the LLM size. We further show that Cayley-Encoder is effective in few-shot regimes. Finally, we show that Cayley-Encoder outperforms LoRA fine-tuning while operating on the frozen LLM. We conclude with an explainability analysis showing that multiple layers contribute meaningfully to the final prediction, supporting our hypothesis.

cs.CL · cs.LG

新论文 · v1arXiv:2611.24383

What Majority Labels Hide: Annotator Disagreement and Model Errors in Multi-Dialect Arabic Sexism Detection

Alaa Bazaid, Arkaitz Zubiaga

原文摘要

Judgements of sexism are subjective and may vary across communities, yet sexism detection datasets usually reduce annotator judgements to a single majority label, obscuring disagreement and how models behave on contested cases. We introduce MuDAS, a dataset of YouTube comments in Saudi, Egyptian and Moroccan Arabic, each labelled by three native speakers of its dialect, with the individual annotations preserved. We use it to evaluate Arabic encoders fine-tuned with hard-label, soft-label and perspectivist objectives, GPT-5.6 Terra, and four Arabic-centric large language models. Model errors are more frequent on comments with annotator disagreement. Despite relatively close Macro-F1 scores, GPT-5.6 Terra and the fine-tuned encoders exhibit contrasting error patterns: GPT-5.6 Terra, even with the annotators' definition of sexism, over-classifies gender-related content such as religious advice on modesty, whereas the encoders miss implicit and prescriptive sexism. The Arabic-centric models also over-predict sexism and remain below the encoders. Disagreement-aware learning preserves Macro-F1, but soft labels improve alignment with annotators on contested comments at the expense of unanimous ones, and neither paradigm reliably identifies contested comments. Lower performance on Egyptian Arabic, clearest for the large language models, is not fully explained by the examined measures of test-set composition and dialectness. Majority-label scores can therefore hide contrasting error patterns that matter for content moderation.

cs.CL · cs.CY · cs.SI

新论文 · v1arXiv:2611.24379

MoE-Context: Show the Model Only What It Needs

Vedant Navale

原文摘要

AI agents read their whole history at every step. People do not work like that: we keep what the moment needs and hold the rest as a rough note we can recall. MoE-Context applies this to language models. The context is split into blocks. Before each step a fast yes or no pass decides which blocks the step needs; the rest stay as one-line notes that can be brought back, and nothing is deleted. In the version we build, one model makes this decision and then answers, over a single key-value cache where hidden blocks are masked rather than removed. On a laptop with Qwen3-1.7B and a synthetic lookup task, showing only the needed block raises accuracy from 93 to 119 of 120, and a masked block is as hidden as a deleted one. The model's own picking is the weak part. A supervised linear probe written into spare output rows lifts its accuracy from 80 to 110 of 120, but simple keyword search picks perfectly on this task, and the probe does not transfer to a second task. We report these results openly, along with two cheap usage patterns that block switching enables: undoing a failed attempt and running subagents without a new context.

cs.AI · cs.CL · cs.LG

版本更新 · v3arXiv:2607.13315

Meta-Learning Preferences for Multilingual LLM Alignment

Jiaying Lin, Seongho Son, Nam Phuong Tran, Long Tran-thanh, Ilija Bogunovic, Debmalya Mandal

原文摘要

Unequal availability of human preference data across languages poses a significant challenge for aligning large language models in multilingual settings. To address the lack of sufficient data in low-resource language alignment, we propose a meta-learning framework for Reinforcement Learning from Human Feedback and Direct Preference Optimization. By leveraging preference data from other languages, our framework learns a transferable initialization that enables effective adaptation to a target language with minimal data. We provide theoretical guarantees for both the meta-reward modeling and meta-policy optimization settings, and empirically demonstrate the effectiveness of our approach on multilingual benchmarks. In an extremely low-resource setting with only 100 target-language preference samples, our approach achieves up to $33.4\%$ win-rate improvements over baseline methods, and consistently outperforms baselines across multiple target languages and model scales. Our approaches retain these advantages across different combinations of meta-training languages and varying linguistic distances from the target languages.

cs.CL

新论文 · v1arXiv:2611.24301

LoopCache: Accelerating Long-Context Self-Speculative Decoding for Looped Transformers

Zizhuo Fu, Haochen Huang, Shengxuan Qiu, Linye Wei, Ziyu Tang, Runsheng Wang, Meng Li

原文摘要

Looped Transformers achieve large effective depth with compact models by repeatedly applying a weight-shared block across R recurrent steps, each maintaining a loop-specific key-value (KV) cache. To accelerate sequential loop execution, recent wave-style self-speculative methods like WaveFront Decoding (WFD) and LoopSpec pipeline draft and target tokens across loops in a single step. However, they face two critical limitations. First, they draft only at fixed loops and ignore intermediate prediction changes, wasting computation on stale trajectories. Second, batching tokens from different loops loads up to Rx the KV cache per step, severely amplifying memory traffic in long-context scenarios. To address these challenges, we propose LoopCache, an efficient self-speculative decoding framework for long-context Looped Transformers. Based on the correlation between prediction switches and output confidence, LoopCache introduces confidence-guided dynamic draft updating to refine drafts in flight across all intermediate loops. Furthermore, LoopCache employs a cyclic loop KV cache strategy that rotates through loop-specific KV caches, restricting per-step KV traffic to a single loop while batching tokens at the nearest matching cycle for lossless verification. By also selectively retaining independent KV caches for a few loop-sensitive shallow layers, LoopCache improves acceptance rates with negligible overhead. Experiments show that LoopCache achieves average 10.25x and 14.56x speedup over WFD and LoopSpec.

cs.CL · cs.LG

版本更新 · v2arXiv:2603.19008

HYPER: Hypothesis-Guided Planning for Evidence-Seeking Retrieval

Hangeol Chang, Changsun Lee, Seungjoon Rho, Junho Yeo, Jong Chul Ye

原文摘要

Retrieval-Augmented Generation (RAG) improves Large Language Models (LLMs) by grounding generation in external, non-parametric knowledge. However, when a task requires choosing among competing possibilities, such as a diagnosis among plausible alternatives, topically relevant context is often insufficient to drive the final decision. To address this issue, we propose HYPER: Hypothesis-Guided Planning for Evidence-Seeking Retrieval, a training-free pre-retrieval planning framework that reorients RAG from topic-oriented retrieval to evidence-oriented retrieval. HYPER uses a provisional answer hypothesis as a planning anchor. It then decomposes retrieval into three evidence roles. Support seeks evidence for the provisional answer. Distinction retrieves criteria that separate competing alternatives. Feature grounds salient question features needed for the decision. This planning step steers retrieval toward evidence near the decision boundary, enabling the generator to answer from context that is more directly aligned with the required choice. On MedQA and MMLU-Med, where single-query retrieval falls below no-retrieval chain-of-thought prompting, HYPER achieves the best average accuracy among ten baselines, improving over Simple RAG by 5.9 and 3.6 points, respectively. Its gains transfer to the 14 subjects of MMLU-Pro, at the cost of two additional LLM calls per question. Code is available at https://github.com/hangeol/HypothesisRAG.

cs.CL · cs.AI · cs.LG

版本更新 · v2arXiv:2603.29765

Training-Free Dynamic Upcycling of Expert Language Models

Eros Fanì, Oğuzhan Ersoy

原文摘要

Large Language Models (LLMs) have achieved remarkable performance on a wide range of specialized tasks, exhibiting strong problem-solving capabilities. However, training these models is prohibitively expensive, and they often lack domain-specific expertise because they rely on general knowledge datasets. Expertise finetuning can address this issue; however, it often leads to overspecialization, and developing a single multi-domain expert remains difficult due to diverging objectives. Furthermore, multitask training is challenging due to interference and catastrophic forgetting. Existing work proposes combining the expertise of dense models within a Mixture of Experts (MoE) architecture, although this approach still requires multitask finetuning. To address these issues, we introduce Dynamic Upcycling MoE (DUME), a novel approach that reuses dense experts trained on different domains to construct a unified MoE model. Our method builds a single multitask model that preserves the capabilities of the original dense experts without requiring additional training. DUME is both cost-efficient and scalable: by leveraging the closed-form solution of ridge regression, it eliminates the need for further optimization and enables experts to be added or removed dynamically. Our extensive experiments on different tasks (in causal language modeling and reasoning settings) with several model sizes (between 115M and 9B parameters) and architectures demonstrate that DUME consistently outperforms baseline approaches. In particular, in a reasoning setting, DUME can also surpass the dense expert model's performance in its domain, achieving 104% of the dense expert's performance on average. Finally, we also show that the DUME model can be fine-tuned to further improve performance in complex reasoning scenarios. Our code is available at: github.com/gensyn-ai/dume.

cs.LG · cs.CL

版本更新 · v2arXiv:2611.06414

Twists and Turns: A Graph-Based Benchmark and Framework for Long-Term Memory in Exploration Tasks

Chongyuan Dai, Yaling Shen, Zirui Song, Junfeng Liao, Jinpeng Hu, Xiuying Chen

原文摘要

Long-term problem-solving processes, such as scientific research and software development, proceed along an evolving trial-and-error trajectory toward a central goal. We term these \textit{exploration tasks}, which require models to remember not only facts but also how solutions are tested against intermediate outcomes and early attempts are continually revised or revived. However, existing long-term memory research focuses on factual recall and updates, overlooking the interconnected nature of exploration. We observe that exploration processes naturally form graphs, with branching approaches linked by revisions and reconsiderations. Motivated by this, we propose \textsc{LExMem}, a graph-guided benchmark designed to evaluate \underline{\textsc{Mem}}ory in \underline{\textsc{L}}ong-term \underline{\textsc{Ex}}ploration tasks. We model exploration processes as graphs and introduce an evolutionary mechanism to increase their structural complexity, yielding 224 dialogues and 2,018 question-answer instances derived from graph structures. Building on the same insight, we propose \textsc{LExGraph}, a graph-based memory framework that incrementally organizes dialogues into interconnected memory nodes and retrieves evidence through hybrid retrieval and neighbor expansion. Extensive experiments demonstrate that \textsc{LExMem} poses a significant challenge to existing retrieval- and memory-based systems, whereas \textsc{LExGraph} achieves superior performance and generalizes effectively to external benchmarks. Codes and datasets are available at https://github.com/mbzuai-nlp/LExMem.

cs.CL

新论文 · v1arXiv:2611.24236

From a finite subgroup of $SO(3)$:$C_2$ leads to a Holant dichotomy

Yuxuan Zhou

原文摘要

Counting complexity within the Holant framework has been extensively studied through complexity dichotomy theorems. While substantial progress has been achieved for various restricted settings, the general complex-valued even-arity case without auxiliary unary signatures remains largely open. This paper studies a fundamental case in this program where the binary signatures induce a group structure satisfying \(M(B(\mathcal{F}))\cong C_2\). It is observed that projective equivalence classes of binary signature matrices naturally embed into finite subgroups of \(SO(3)\), providing an algebraic viewpoint for structural classification. This paper further identifies two fundamentally different regimes for the \(C_2\) case, namely the normal and isotropic cases, and analyzes their complexity behavior separately. Combining gadget constructions, decomposition techniques, interpolation arguments, and a realnumberization approach that transforms the complex-valued setting into an equivalent real-structured framework, this paper establishes a complete dichotomy theorem: for every even-arity signature set \(\mathcal{F}\) satisfying \(M(\B(\mathcal{F}))\cong C_2\), the problem \(\mathrm{Holant}(\mathcal{F})\) is either polynomial-time computable or \#P-hard. A Pauli-basis representation for signatures is further introduced as an algebraically natural language for organizing realizable signatures and revealing structural relations that are less transparent in the standard basis.

cs.CC · cs.CL · cs.DM

版本更新 · v4arXiv:2504.09895

Reinforcement Learning with Reference-Based Rewards for Language Model Reasoning and Alignment

Shuai Zhao, Yunqiu Xu, Linchao Zhu, Ling Chen, Yi Yang

原文摘要

Reinforcement learning (RL) powers two major lines of large language model (LLM) post-training: reasoning, which elicits deliberate, verifiable problem solving, and alignment, which steers models toward helpful, harmless, and honest behavior. Both hinge on reward design. Reasoning relies on rule-based verification of final answers, whose binary feedback is blind to intermediate derivations, whereas process reward models (PRMs) require step-level annotations too costly to scale. Alignment depends on reward models (RMs) trained on binary preference pairs, making both data collection and model training resource-intensive. We propose RL2R, reinforcement learning with reference-based rewards, as a unified and inexpensive alternative. Using pre-existing or distilled references, RL2R measures semantic similarity with lightweight metrics such as BERTScore between a sampled generation and a reference answer or solution. Similarity preserves ordinal quality signals that binary feedback discards: responses closer to the reference earn higher rewards, and sound derivations receive credit even when the final conclusion is wrong. For mathematical reasoning, augmenting rule-based verification with answer- and step-level similarity provides low-cost process supervision and outperforms answer-only verification. For alignment, a critic-free REINFORCE-style instantiation covers general preference, safety, and confidence tasks on par with RM-based methods. Reference-based rewards thus offer a cost-effective, versatile reward source spanning reasoning and alignment.

cs.CL · cs.AI · cs.LG

版本更新 · v2arXiv:2510.05364

Challenging Quadratic Attention$\unicode{x2013}$A Holistic View on the Rise of Alternative Language Model Architectures

Alexander M. Fichtl, Jeremias Bohn, Josefin Kelber, Edoardo Mosca, Georg Groh

原文摘要

Transformers have dominated sequence processing tasks for the past seven years$\unicode{x2013}$most notably language modeling. However, the inherent quadratic complexity of their attention mechanism remains a significant bottleneck as context length increases. We review and distill the recent efforts to overcome this bottleneck, including advances in (sub-quadratic) attention variants, recurrent neural networks, state space models, and hybrid architectures. We critically analyze approaches regarding compute and memory complexity, benchmark results, and fundamental limitations to assess whether the dominance of pure-attention transformers may soon be challenged, which we consider possible, particularly in domain-specific and edge-device applications.

cs.CL

新论文 · v1arXiv:2611.24168

CounselTraj: Synthesizing Process-Aware Psychological Counseling Dialogues via Dialogue-Act Trajectory Planning

Zhiyang Qi, Michimasa Inaba

原文摘要

The growing demand for psychological counseling has motivated the development of LLM-based systems that support human counselors, but training such systems requires large-scale, high-quality dialogue data. Authentic counseling transcripts are difficult to collect and share because of privacy concerns, making synthetic dialogues an important alternative. However, existing synthetic datasets, despite covering diverse scenarios and client profiles, often consist of brief exchanges and fail to capture session-level counseling processes. We propose CounselTraj, a dialogue-act trajectory planning framework for expanding brief synthetic dialogues into process-aware, session-level psychological counseling dialogues. CounselTraj learns dialogue-act patterns from role-play counseling dialogues produced by human counselors, plans coordinated trajectories for clients and counselors, and uses them as flexible guidance for long-form dialogue generation. Experiments across three Chinese synthetic counseling datasets show that CounselTraj produces more complete and coherent sessions with higher-quality counselor interventions than both direct expansion and expansion with randomly sampled trajectories. Based on CounselTraj, we construct CounselTrajD, a dataset containing over 10K process-guided Chinese psychological counseling dialogues, and use it to train CounselTrajLM for counseling response generation. Evaluations at both the response and session levels show that CounselTrajLM outperforms publicly available Chinese counseling models in response quality, therapeutic understanding and intervention, and counseling-process coherence and progress. Our code and model are publicly available at https://github.com/UEC-InabaLab/CounselTraj.

cs.CL · cs.AI

新论文 · v1arXiv:2611.24156

Once Upon a Bias: Gender Stereotyping in LLM-Generated Children's Stories

Sayumi Muthukumarana, Roshan Godaliyadda, Parakrama Ekanayake, Roshan Ragel

原文摘要

Large language models are increasingly used to create stories for children. Since bedtime stories can shape children's perceptions of social roles, this raises concerns about whether LLMs adapt narrative portrayals according to the intended recipient's gender. We conducted a large-scale study to examine gender stereotyping in LLM-generated bedtime stories. Using matched daughter/son prompts and a five-generation design, we analyze 62,873 retained stories from 16 LLMs across 35 country settings and 12 narrator roles. We introduce a multi-level evaluation framework combining context-adjusted lexical rate contrasts, relative stereotype alignment, narrator-recipient interaction analysis, human evaluation, and representation-level interventions to examine both the consistency and interpretability of these differences. Our results reveal systematic differences in narrative framing: son-directed stories emphasize higher agency and lower communion compared to daughter-directed stories. While narrator gender expresses limited moderation on these patterns, high sociability exhibits a more pronounced interaction. Human evaluations demonstrate consistent agreement, although correspondence with lexical measures varies across narrative dimensions. Additionally, hidden-state analyses reveal decodable recipient-gender information. These findings demonstrate persistent yet context-dependent representational disparities in LLM-generated children's stories and motivate narrative-specific bias evaluation beyond short-form benchmarks.

cs.CL · cs.AI

版本更新 · v2arXiv:2608.30427

Ceiling-Clipped Acceptance Histograms Indicate Stranded Speed-up in Block-Diffusion Speculative Decoding

Ephrem Wu

原文摘要

Speculative decoding speeds up generation using an efficient draft model (drafter) to propose tokens that a target model verifies in one pass, preserving the target's output distribution. High-acceptance block-diffusion drafters, such as DFlash and DFlare, fill an entire block in one parallel pass. In many cycles, the target model accepts all draft tokens in the block, so the drafter exhausts its trained horizon before verification ever fails. We call this unrealized acceptance stranded speed-up. Evaluating a drafter by its per-prompt or per-cycle mean committed length alone hides this phenomenon, whereas the acceptance histogram exposes this bottleneck as a spike in the ceiling bin, the fraction of cycles that accept the entire block. We recommend reviewing the acceptance histogram before spending more training compute. Naively widening the block at inference could reduce acceptance for drafters with bidirectional attention, since, once the block outgrows its training size, the drafter's bidirectional attention shifts its distribution even at early positions, which empirically degrades acceptance there. We therefore post-train the drafter to a longer block with a short curriculum that emphasizes the newly exposed positions, a method we call DBloom. We continue training the pretrained DFlash and DFlare drafters at the original block size of 16, before expanding them to 24 (B16 to B24) on Qwen3-8B and Qwen3-4B. Compared to the original B16 drafters, this technique raises per-prompt committed length by up to +1.7 tokens and decode speed-up by a median of +10.2% versus +5.7% without the continuation. On high-ceiling benchmarks, the speed-up rises to +11.9%. On Gemma-4-12B-IT and GPT-OSS-20B, expansion also raises committed length on every benchmark. Against JetSpec, a tree drafter, DBloom has higher committed length by point estimate on every benchmark at tree budgets up to 64 nodes.

cs.CL · cs.LG

版本更新 · v2arXiv:2603.10233

S-GRADES -- Studying Generalization of Student Response Assessments in Diverse Evaluative Settings

Tasfia Seuti, Sagnik Ray Choudhury

原文摘要

Evaluating student responses, from long essays to short factual answers, is a key challenge in educational NLP. Automated Essay Scoring (AES) focuses on holistic writing qualities such as coherence and argumentation, while Automatic Short Answer Grading (ASAG) emphasizes factual correctness and conceptual understanding. Despite their shared goal, these paradigms have progressed in isolation with fragmented datasets, inconsistent metrics, and separate communities. We introduce S-GRADES (Studying Generalization of Student Response Assessments in Diverse Evaluative Settings), a web-based benchmark that consolidates 14 diverse grading datasets under a unified interface with standardized access and reproducible evaluation protocols. The benchmark is fully open-source and designed for extensibility, enabling continuous integration of new datasets and evaluation settings. To demonstrate the utility of S-GRADES, we evaluate three state-of-the-art large language models across the benchmark using multiple reasoning strategies in prompting. We further examine the effects of exemplar selection and cross-dataset exemplar transfer. Our analyses illustrate how benchmark-driven evaluation reveals reliability and generalization gaps across essay and short-answer grading tasks, highlighting the importance of standardized, cross-paradigm assessment.

cs.CL