OneLA: Scaling Linear-Attention Decoding to Large Beams in Generative Recommendation

By: Xiangrui Yang · Cheng Peng · Yunfeng Zhao · Liang Zeng · Ao Hu · Jiawei Yang · Shengzhe Wang · Jingshan Lv · Xiao Liang · Chen Yang · Jiaqiang Liu · Yiming Qiu

14 September 2026 at 12:00

arXiv:2609.12399v1 Announce Type: new Abstract: Generative recommendation (GR) relies on large-beam decoding to generate hundreds of candidate items, creating a new scaling challenge for recurrent linear attention. Existing linear attention serving systems either materialize a full recurrent state for every beam or repeatedly replay shared history, incurring substantial memory and traffic overhead. To address this, we present OneLA, a linear-attention decoding framework that exploits the shared prompt and short divergent suffixes of GR workloads. Specifically, OneLA represents all beam states using a single shared prompt-derived state and compact, append-only records of their divergent transitions. Using this representation, OneLA computes only the state information required at each decoding step, without reconstructing a full recurrent state for every beam. Furthermore, OneLA uses a lightweight ancestry index to track the transition records that make up each beam's history, allowing beams to be updated without moving or copying existing records. A fused GPU kernel further reuses the shared state across beams. Our analysis shows that OneLA achieves 1.54-2.46x end-to-end decode speedups while substantially reducing recurrent-state memory use and data movement.

UniPart: Towards Zero-shot Language-Grounded 3D Part Segmentation for Embodied Interaction

cs.AI, q-bio.NC updates on arXiv.org

By: Xinqiang Yu · Zekun qi · Jiawei He · Wenyao Zhang · Xuchuan Chen · Guaocai Yao · Li Yi · Zhaoxiang Zhang · He Wang

14 September 2026 at 12:00

arXiv:2609.12898v1 Announce Type: cross Abstract: Fine-grained robotic manipulation depends on understanding parts, not only whole objects. Existing 3D foundation models tend to be either generalized but object-aware, or part-aware but limited to closed-set taxonomies, which weakens zero-shot transfer. We study text-conditioned 3D part segmentation, where a free-form phrase selects a functional part on point cloud. We introduce UniPart, a feed-forward cross-modal 3D Transformer that conditions CLIP text embedding. To scale supervision, we build LangPart-1M with 160K+ Objaverse assets and 8M text to part pairs using multi-view consistent part generation. We further manually label a high-quality subset, LangPart-4K, for fine-tuning and evaluation. UniPart achieves strong zero-shot results on open-vocabulary part benchmarks and transfers to language-conditioned part grasping in real world.

The metastatic spectrum in functional and non-functional NENs: mechanistic insights from multi-omics

(Multiomics OR Omics) AND (Pancreatic)

By: Zhengjian Wang · Xuda Ji · Zhe Wang · Shunzhen Zheng · Fangfeng Liu · Jun Lu · Hong Chang

11 September 2026 at 18:00

Front Endocrinol (Lausanne). 2026 Aug 27;17:1782791. doi: 10.3389/fendo.2026.1782791. eCollection 2026.

ABSTRACT

Neuroendocrine neoplasms (NENs) are biologically heterogeneous tumors in which differentiation/grade and hormonal functionality are intersecting but non-equivalent axes. This review focuses on functional and non-functional well-differentiated neuroendocrine tumors (NETs), principally gastroenteropancreatic and pancreatic NETs, and critically evaluates how site, lineage, stage, tumor burden, genomic and epigenetic alterations, immune-stromal remodeling, metabolic adaptation, microbiome-associated signals, and treatment pressure converge on metastasis and recurrence. Apparent outcome differences by functionality are inconsistent after clinicopathological adjustment: non-functional presentation is often enriched for delayed diagnosis and adverse features, whereas functional subtypes range from typically indolent insulinomas to clinically aggressive hormone-producing tumors. We reconcile these observations through a layered model in which lineage-defining alterations and chromatin/telomere programs establish cellular state; signaling and metabolic plasticity enable stress adaptation; and hypoxia, angiogenesis, immune cells, fibroblasts, extracellular matrix, and therapy create selective niches for dissemination and relapse. We also define computational strategies for heterogeneous multi-omics integration and a staged biomarker-validation pathway. Evidence remains dominated by pancreatic NETs, and causal support is weakest for microbiome-functionality relationships and several proposed cross-omic links. A spectrum-based framework is therefore most useful when it generates testable, site- and grade-specific hypotheses rather than treating functionality as an isolated prognostic variable.

PMID:42724134 | PMC:PMC13559159 | DOI:10.3389/fendo.2026.1782791

Accelerating Long-Tail Generation in Synchronous RLHF Training via Adaptive Tensor Parallelism

cs.AI, q-bio.NC updates on arXiv.org

By: Long Zhao · Qinghe Wang · Jiaan Zhu · Youhui Bai · Zewen Jin · Chaoyi Ruan · Shengnan Wang · Cheng Li

26 May 2026 at 12:00

arXiv:2605.23945v1 Announce Type: new Abstract: Reinforcement Learning from Human Feedback (RLHF) has become a key post-training paradigm for improving model quality. However, the synchronous three-stage RLHF pipeline is often bottlenecked by the generation stage, where response-length skew causes the effective batch size to shrink rapidly during decoding, leaving GPUs underutilized while a few long responses remain unfinished. Mainstream frameworks employ a static tensor parallelism (TP) configuration that cannot adapt to changing batch characteristics, leaving substantial performance headroom unexplored. We propose PAT, an adaptive TP method that dynamically reconfigures TP during the generation stage of each RLHF iteration. PAT introduces two key techniques. First, a predictor-guided online reconfiguration method decides both the reconfiguration point and the target TP configuration based on offline profiling, triggering reconfiguration only when the predicted latency benefit outweighs the reconfiguration overhead. Second, a lightweight online reconfiguration mechanism updates only the states and layouts affected by TP changes: it adapts unfinished decoding states through a cost-model-based choice between KV-cache migration and recomputation, performs in-place weight resharding, and reuses cached communication groups. We implement PAT on top of SGLang and integrate it with the VeRL framework. Evaluations on LLaMA3.1-8B and Qwen3-14B using DeepScaleR show that PAT reduces generation latency by up to 34.6% and end-to-end RLHF training iteration latency by up to 27.2% compared to the original VeRL setup.

CR1(+) tumor-associated macrophages orchestrate an immunosuppressive niche in hepatocellular carcinoma: a genetic and multi-omics dissection

(Multiomics OR Omics) AND (Pancreatic)

By: Zhengjian Wang · Zhe Wang · Xuda Ji · Liping Zhao · Kai Zheng · Wen Yu · Hanzhe Zhang · Hong Chang · Fangfeng Liu

25 May 2026 at 18:00

J Transl Med. 2026 May 25. doi: 10.1186/s12967-026-08301-z. Online ahead of print.

ABSTRACT

BACKGROUND: Hepatocellular carcinoma (HCC) remains a major global health burden and a leading cause of cancer-related mortality. Advanced disease is characterized by a profoundly immunosuppressive tumor microenvironment (TME) and limited durable responses to therapy. However, the upstream genetic determinants that drive tumor-associated macrophage (TAM) dysfunction in HCC remain poorly defined. Using an integrative genetic and multi-omics framework, we investigated complement receptor 1 (CR1) as a candidate regulator of this immunosuppressive niche.

METHODS: We combined Mendelian randomization (MR) and metabolite mediation analyses with bulk, single-cell, and spatial transcriptomics to define the role of CR1 in HCC. Public datasets included the TCGA-HCC cohort, a single-cell RNA-sequencing dataset comprising 53,474 high-quality cells from 21 samples, and two spatially profiled HCC sections. Clinical validation was performed in 30 paired HCC and adjacent liver tissues. Functional assays were conducted in THP-1-derived macrophages using CR1 gain- and loss-of-function approaches, phagocytosis assays, and macrophage-CD8⁺ T-cell co-culture experiments.

RESULTS: MR analyses implicated CR1 in HCC susceptibility at both the protein and transcript levels. pQTL analysis linked genetically predicted circulating CR1 levels to HCC risk (IVW OR = 1.403, p = 0.017), and mediation analysis identified specific metabolites as candidate intermediates. Integrative multi-omics analyses showed that CR1 was preferentially enriched in TAMs, spatially co-localized with the M2 marker CD206, and associated with reduced CD8⁺ T-cell infiltration, enhanced T-cell exhaustion signatures, advanced clinicopathological features, and poorer survival. In 30 paired clinical samples, CR1-high tumors exhibited increased M2-like macrophage accumulation and reduced CD8⁺ T-cell infiltration. Functionally, CR1 overexpression drove macrophages toward an M2-like phenotype, enhanced phagocytic activity, increased PD-L1 expression, and suppressed CD8⁺ T-cell proliferation as well as IFN-gamma and granzyme B production, whereas CR1 knockdown produced the opposite phenotype.

CONCLUSIONS: Our study provides the first integrated genetic, spatial, and functional evidence that CR1⁺ TAMs constitute a clinically relevant immunoregulatory axis in HCC. These findings extend current understanding of complement-associated immunosuppression beyond canonical complement cascade activity and support CR1 as a candidate biomarker and therapeutic target for macrophage reprogramming, with potential translational relevance for combination strategies involving immune checkpoint blockade.

PMID:42185899 | DOI:10.1186/s12967-026-08301-z

CR1(+) tumor-associated macrophages orchestrate an immunosuppressive niche in hepatocellular carcinoma: a genetic and multi-omics dissection

Omics in Hepatocellular

By: Zhengjian Wang · Zhe Wang · Xuda Ji · Liping Zhao · Kai Zheng · Wen Yu · Hanzhe Zhang · Hong Chang · Fangfeng Liu

25 May 2026 at 18:00

J Transl Med. 2026 May 25. doi: 10.1186/s12967-026-08301-z. Online ahead of print.

ABSTRACT

BACKGROUND: Hepatocellular carcinoma (HCC) remains a major global health burden and a leading cause of cancer-related mortality. Advanced disease is characterized by a profoundly immunosuppressive tumor microenvironment (TME) and limited durable responses to therapy. However, the upstream genetic determinants that drive tumor-associated macrophage (TAM) dysfunction in HCC remain poorly defined. Using an integrative genetic and multi-omics framework, we investigated complement receptor 1 (CR1) as a candidate regulator of this immunosuppressive niche.

METHODS: We combined Mendelian randomization (MR) and metabolite mediation analyses with bulk, single-cell, and spatial transcriptomics to define the role of CR1 in HCC. Public datasets included the TCGA-HCC cohort, a single-cell RNA-sequencing dataset comprising 53,474 high-quality cells from 21 samples, and two spatially profiled HCC sections. Clinical validation was performed in 30 paired HCC and adjacent liver tissues. Functional assays were conducted in THP-1-derived macrophages using CR1 gain- and loss-of-function approaches, phagocytosis assays, and macrophage-CD8⁺ T-cell co-culture experiments.

RESULTS: MR analyses implicated CR1 in HCC susceptibility at both the protein and transcript levels. pQTL analysis linked genetically predicted circulating CR1 levels to HCC risk (IVW OR = 1.403, p = 0.017), and mediation analysis identified specific metabolites as candidate intermediates. Integrative multi-omics analyses showed that CR1 was preferentially enriched in TAMs, spatially co-localized with the M2 marker CD206, and associated with reduced CD8⁺ T-cell infiltration, enhanced T-cell exhaustion signatures, advanced clinicopathological features, and poorer survival. In 30 paired clinical samples, CR1-high tumors exhibited increased M2-like macrophage accumulation and reduced CD8⁺ T-cell infiltration. Functionally, CR1 overexpression drove macrophages toward an M2-like phenotype, enhanced phagocytic activity, increased PD-L1 expression, and suppressed CD8⁺ T-cell proliferation as well as IFN-gamma and granzyme B production, whereas CR1 knockdown produced the opposite phenotype.

CONCLUSIONS: Our study provides the first integrated genetic, spatial, and functional evidence that CR1⁺ TAMs constitute a clinically relevant immunoregulatory axis in HCC. These findings extend current understanding of complement-associated immunosuppression beyond canonical complement cascade activity and support CR1 as a candidate biomarker and therapeutic target for macrophage reprogramming, with potential translational relevance for combination strategies involving immune checkpoint blockade.

PMID:42185899 | DOI:10.1186/s12967-026-08301-z

CR1(+) tumor-associated macrophages orchestrate an immunosuppressive niche in hepatocellular carcinoma: a genetic and multi-omics dissection

(Multiomics OR Omics) AND (Lung OR gastric OR Hepatocellular)

By: Zhengjian Wang · Zhe Wang · Xuda Ji · Liping Zhao · Kai Zheng · Wen Yu · Hanzhe Zhang · Hong Chang · Fangfeng Liu

25 May 2026 at 18:00

J Transl Med. 2026 May 25. doi: 10.1186/s12967-026-08301-z. Online ahead of print.

ABSTRACT

BACKGROUND: Hepatocellular carcinoma (HCC) remains a major global health burden and a leading cause of cancer-related mortality. Advanced disease is characterized by a profoundly immunosuppressive tumor microenvironment (TME) and limited durable responses to therapy. However, the upstream genetic determinants that drive tumor-associated macrophage (TAM) dysfunction in HCC remain poorly defined. Using an integrative genetic and multi-omics framework, we investigated complement receptor 1 (CR1) as a candidate regulator of this immunosuppressive niche.

METHODS: We combined Mendelian randomization (MR) and metabolite mediation analyses with bulk, single-cell, and spatial transcriptomics to define the role of CR1 in HCC. Public datasets included the TCGA-HCC cohort, a single-cell RNA-sequencing dataset comprising 53,474 high-quality cells from 21 samples, and two spatially profiled HCC sections. Clinical validation was performed in 30 paired HCC and adjacent liver tissues. Functional assays were conducted in THP-1-derived macrophages using CR1 gain- and loss-of-function approaches, phagocytosis assays, and macrophage-CD8⁺ T-cell co-culture experiments.

RESULTS: MR analyses implicated CR1 in HCC susceptibility at both the protein and transcript levels. pQTL analysis linked genetically predicted circulating CR1 levels to HCC risk (IVW OR = 1.403, p = 0.017), and mediation analysis identified specific metabolites as candidate intermediates. Integrative multi-omics analyses showed that CR1 was preferentially enriched in TAMs, spatially co-localized with the M2 marker CD206, and associated with reduced CD8⁺ T-cell infiltration, enhanced T-cell exhaustion signatures, advanced clinicopathological features, and poorer survival. In 30 paired clinical samples, CR1-high tumors exhibited increased M2-like macrophage accumulation and reduced CD8⁺ T-cell infiltration. Functionally, CR1 overexpression drove macrophages toward an M2-like phenotype, enhanced phagocytic activity, increased PD-L1 expression, and suppressed CD8⁺ T-cell proliferation as well as IFN-gamma and granzyme B production, whereas CR1 knockdown produced the opposite phenotype.

CONCLUSIONS: Our study provides the first integrated genetic, spatial, and functional evidence that CR1⁺ TAMs constitute a clinically relevant immunoregulatory axis in HCC. These findings extend current understanding of complement-associated immunosuppression beyond canonical complement cascade activity and support CR1 as a candidate biomarker and therapeutic target for macrophage reprogramming, with potential translational relevance for combination strategies involving immune checkpoint blockade.

PMID:42185899 | DOI:10.1186/s12967-026-08301-z

TABQAWORLD: Optimizing Multimodal Reasoning for Multi-Turn Table Question Answering

cs.AI, q-bio.NC updates on arXiv.org

By: Tung Sum Thomas Kwok · Xinyu Wang · Xiaofeng Lin · Peng Lu · Chunhe Wang · Changlun Li · Hanwei Wu · Nan Tang · Elisa Kreiss · Guang Cheng

7 April 2026 at 12:00

arXiv:2604.03393v1 Announce Type: new Abstract: Multimodal reasoning has emerged as a powerful framework for enhancing reasoning capabilities of reasoning models. While multi-turn table reasoning methods have improved reasoning accuracy through tool use and reward modeling, they rely on fixed text serialization for table state readouts. This introduces representation errors in table encoding that significantly accumulate over multiple turns. Such accumulation is alleviated by tabular grounding methods in the expense of inference compute and cost, rendering real world deployment impractical. To address this, we introduce TABQAWORLD, a table reasoning framework that jointly optimizes tabular action through representation and estimation. For representation, TABQAWORLD employs an action-conditioned multimodal selection policy, which dynamically switches between visual and textual representations to maximize table state readout reliability. For estimation, TABQAWORLD optimizes stepwise reasoning trajectory through table metadata including dimension, data types and key values, safely planning trajectory and compressing low-complexity actions to reduce conversation turns and latency. Designed as a training-free framework, empirical evaluations show that TABQAWORLD achieves state-of-the-art performance with 4.87% accuracy improvements over baselines, with 5.42% accuracy gain and 33.35% inference latency reduction over static settings, establishing a new standard for reliable and efficient table reasoning.

Four Generations of Quantum Biomedical Sensors

cs.AI, q-bio.NC updates on arXiv.org

By: Xin Jin · Priyam Srivastava · Ronghe Wang · Yuqing Li · Jonathan Beaumariage · Tom Purdy · M. V. Gurudev Dutt · Kang Kim · Kaushik Seshadreesan · Junyu Liu

3 April 2026 at 12:00

arXiv:2603.29944v2 Announce Type: replace-cross Abstract: Quantum sensing technologies offer transformative potential for ultra-sensitive biomedical sensing, yet their clinical translation remains constrained by classical noise limits and a reliance on macroscopic ensembles. We propose a unifying generational framework to organize the evolving landscape of quantum biosensors based on their utilization of quantum resources. First-generation devices utilize discrete energy levels for signal transduction but follow classical scaling laws. Second-generation sensors exploit quantum coherence to reach the standard quantum limit, while third-generation architectures leverage entanglement and spin squeezing to approach Heisenberg-limited precision. We further define an emerging fourth generation characterized by the end-to-end integration of quantum sensing with quantum learning and variational circuits, enabling adaptive inference directly within the quantum domain. By analyzing critical parameters such as bandwidth matching and sensor-tissue proximity, we identify key technological bottlenecks and propose a roadmap for transitioning from measuring physical observables to extracting structured biological information with quantum-enhanced intelligence.

Pyruvate is a natural suppressor of interferon signaling by inducing STAT1 protein pyruvylation

Cell

By: Yibo Zuo · Qin Wang · Wanying Tian · Xinhe Wang · Zhijin Zheng · Wei He · Renxia Zhang · Qian Zhao · Ying Miao · Yukang Yuan · Tingting Zhang · Qun Cui · Yuerong Zhang · Chunyan Liu · Haiyan Zhou · Hui Zheng

27 February 2026 at 08:00

Yibo et al. identify protein pyruvylation as a post-translational modification that can modulate immune signaling and host antiviral response.

Four Generations of Quantum Biomedical Sensors

cs.AI, q-bio.NC updates on arXiv.org

By: Xin Jin · Priyam Srivastava · Ronghe Wang · Yuqing Li · Jonathan Beaumariage · Tom Purdy · M. V. Gurudev Dutt · Kang Kim · Kaushik Seshadreesan · Junyu Liu

1 April 2026 at 12:00

arXiv:2603.29944v1 Announce Type: cross Abstract: Quantum sensing technologies offer transformative potential for ultra-sensitive biomedical sensing, yet their clinical translation remains constrained by classical noise limits and a reliance on macroscopic ensembles. We propose a unifying generational framework to organize the evolving landscape of quantum biosensors based on their utilization of quantum resources. First-generation devices utilize discrete energy levels for signal transduction but follow classical scaling laws. Second-generation sensors exploit quantum coherence to reach the standard quantum limit, while third-generation architectures leverage entanglement and spin squeezing to approach Heisenberg-limited precision. We further define an emerging fourth generation characterized by the end-to-end integration of quantum sensing with quantum learning and variational circuits, enabling adaptive inference directly within the quantum domain. By analyzing critical parameters such as bandwidth matching and sensor-tissue proximity, we identify key technological bottlenecks and propose a roadmap for transitioning from measuring physical observables to extracting structured biological information with quantum-enhanced intelligence.

AVA-Bench: Atomic Visual Ability Benchmark for Vision Foundation Models

cs.AI, q-bio.NC updates on arXiv.org

By: Zheda Mai · Arpita Chowdhury · Zihe Wang · Sooyoung Jeon · Lemeng Wang · Jiacheng Hou · Wei-Lun Chao

1 April 2026 at 12:00

arXiv:2506.09082v4 Announce Type: replace-cross Abstract: The rise of vision foundation models (VFMs) calls for systematic evaluation. A common approach pairs VFMs with large language models (LLMs) as general-purpose heads, followed by evaluation on broad Visual Question Answering (VQA) benchmarks. However, this protocol has two key blind spots: (i) the instruction tuning data may not align with VQA test distributions, meaning a wrong prediction can stem from such data mismatch rather than a VFM' visual shortcomings; (ii) VQA benchmarks often require multiple visual abilities, making it hard to tell whether errors stem from lacking all required abilities or just a single critical one. To address these gaps, we introduce AVA-Bench, the first benchmark that explicitly disentangles 14 Atomic Visual Abilities (AVAs) -- foundational skills like localization, depth estimation, and spatial understanding that collectively support complex visual reasoning tasks. By decoupling AVAs and matching training and test distributions within each, AVA-Bench pinpoints exactly where a VFM excels or falters. Applying AVA-Bench to leading VFMs thus reveals distinctive "ability fingerprints," turning VFM selection from educated guesswork into principled engineering. Notably, we find that a 0.5B LLM yields similar VFM rankings as a 7B LLM while cutting GPU hours by 8x, enabling more efficient evaluation. By offering a comprehensive and transparent benchmark, we hope AVA-Bench lays the foundation for the next generation of VFMs.

ToolTree: Efficient LLM Agent Tool Planning via Dual-Feedback Monte Carlo Tree Search and Bidirectional Pruning

cs.AI, q-bio.NC updates on arXiv.org

By: Shuo Yang · Soyeon Caren Han · Yihao Ding · Shuhe Wang · Eduard Hoy

16 March 2026 at 12:00

arXiv:2603.12740v1 Announce Type: new Abstract: Large Language Model (LLM) agents are increasingly applied to complex, multi-step tasks that require interaction with diverse external tools across various domains. However, current LLM agent tool planning methods typically rely on greedy, reactive tool selection strategies that lack foresight and fail to account for inter-tool dependencies. In this paper, we present ToolTree, a novel Monte Carlo tree search-inspired planning paradigm for tool planning. ToolTree explores possible tool usage trajectories using a dual-stage LLM evaluation and bidirectional pruning mechanism that enables the agent to make informed, adaptive decisions over extended tool-use sequences while pruning less promising branches before and after the tool execution. Empirical evaluations across both open-set and closed-set tool planning tasks on 4 benchmarks demonstrate that ToolTree consistently improves performance while keeping the highest efficiency, achieving an average gain of around 10\% compared to the state-of-the-art planning paradigm.

Asynchronous Verified Semantic Caching for Tiered LLM Architectures

cs.AI, q-bio.NC updates on arXiv.org

By: Asmit Kumar Singh · Haozhe Wang · Laxmi Naga Santosh Attaluri · Tak Chiam · Weihua Zhu

16 March 2026 at 12:00

arXiv:2602.13165v2 Announce Type: replace-cross Abstract: Large language models (LLMs) now sit in the critical path of search, assistance, and agentic workflows, making semantic caching essential for reducing inference cost and latency. Production deployments typically use a tiered static-dynamic design: a static cache of curated, offline vetted responses mined from logs, backed by a dynamic cache populated online. In practice, both tiers are commonly governed by a single embedding similarity threshold, which induces a hard tradeoff: conservative thresholds miss safe reuse opportunities, while aggressive thresholds risk serving semantically incorrect responses. We introduce Krites, an asynchronous, LLM-judged caching policy that expands static coverage without changing serving decisions. On the critical path, Krites behaves exactly like a standard static threshold policy. When the nearest static neighbor of the prompt falls just below the static threshold, Krites asynchronously invokes an LLM judge to verify whether the static response is acceptable for the new prompt. Approved matches are promoted into the dynamic cache, allowing future repeats and paraphrases to reuse curated static answers and expanding static reach over time. In trace-driven simulations on conversational and search workloads, Krites increases the fraction of requests served with curated static answers (direct static hits plus verified promotions) by up to3.9 times for conversational traffic and search-style queries relative to tuned baselines, with unchanged critical path latency.

The Choice of Divergence: A Neglected Key to Mitigating Diversity Collapse in Reinforcement Learning with Verifiable Reward

cs.AI, q-bio.NC updates on arXiv.org

By: Long Li · Zhijian Zhou · Jiaran Hao · Jason Klein Liu · Yanting Miao · Wei Pang · Xiaoyu Tan · Wei Chu · Zhe Wang · Shirui Pan · Chao Qu · Yuan Qi

4 March 2026 at 13:00

arXiv:2509.07430v4 Announce Type: replace-cross Abstract: A central paradox in fine-tuning Large Language Models (LLMs) with Reinforcement Learning with Verifiable Reward (RLVR) is the frequent degradation of multi-attempt performance (Pass@k) despite improvements in single-attempt accuracy (Pass@1). This is often accompanied by catastrophic forgetting, where models lose previously acquired skills. While various methods have been proposed, the choice and function of the divergence term have been surprisingly unexamined as a proactive solution. We argue that standard RLVR objectives -- both those using the mode-seeking reverse KL-divergence and those forgoing a divergence term entirely -- lack a crucial mechanism for knowledge retention. The reverse-KL actively accelerates this decay by narrowing the policy, while its absence provides no safeguard against the model drifting from its diverse knowledge base. We propose a fundamental shift in perspective: using the divergence term itself as the solution. Our framework, Diversity-Preserving Hybrid RL (DPH-RL), leverages mass-covering f-divergences (like forward-KL and JS-divergence) to function as a rehearsal mechanism. By continuously referencing the initial policy, this approach forces the model to maintain broad solution coverage. Extensive experiments on math and SQL generation demonstrate that DPH-RL not only resolves the Pass@k degradation but improves both Pass@1 and Pass@k in- and out-of-domain. Additionally, DPH-RL is more training-efficient because it computes f-divergence using generator functions, requiring only sampling from the initial policy and no online reference model. Our work highlights a crucial, overlooked axis for improving RLVR, demonstrating that the proper selection of a divergence measure is a powerful tool for building more general and diverse reasoning models.

CausalFlip: A Benchmark for LLM Causal Judgment Beyond Semantic Matching

cs.AI, q-bio.NC updates on arXiv.org

By: Yuzhe Wang · Yaochen Zhu · Jundong Li

24 February 2026 at 13:00

arXiv:2602.20094v1 Announce Type: new Abstract: As large language models (LLMs) witness increasing deployment in complex, high-stakes decision-making scenarios, it becomes imperative to ground their reasoning in causality rather than spurious correlations. However, strong performance on traditional reasoning benchmarks does not guarantee true causal reasoning ability of LLMs, as high accuracy may still arise from memorizing semantic patterns instead of analyzing the underlying true causal structures. To bridge this critical gap, we propose a new causal reasoning benchmark, CausalFlip, designed to encourage the development of new LLM paradigm or training algorithms that ground LLM reasoning in causality rather than semantic correlation. CausalFlip consists of causal judgment questions built over event triples that could form different confounder, chain, and collider relations. Based on this, for each event triple, we construct pairs of semantically similar questions that reuse the same events but yield opposite causal answers, where models that rely heavily on semantic matching are systematically driven toward incorrect predictions. To further probe models' reliance on semantic patterns, we introduce a noisy-prefix evaluation that prepends causally irrelevant text before intermediate causal reasoning steps without altering the underlying causal relations or the logic of the reasoning process. We evaluate LLMs under multiple training paradigms, including answer-only training, explicit Chain-of-Thought (CoT) supervision, and a proposed internalized causal reasoning approach that aims to mitigate explicit reliance on correlation in the reasoning process. Our results show that explicit CoT can still be misled by spurious semantic correlations, where internalizing reasoning steps yields substantially improved causal grounding, suggesting that it is promising to better elicit the latent causal reasoning capabilities of base LLMs.

Reading view