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

Training Skills Like Parameters via Self-Supervised Semantic Diffusion

Mo Li, Zixin Yin, Ting Cao, Yunxin Liu

While Large Language Models (LLMs) demonstrate remarkable general instruction-following capabilities, they often fall short of human experts in highly specialized, open-ended domains such as creative screenwriting. Prior approaches typically adopt post-training, yet both supervised fine-tuning and reinforcement learning require weight access that closed-source frontier models do not offer, and demand heavy compute. Moreover, what is learned is tied to a single checkpoint and cannot be inspected by humans. Recent advancements in agentic continual learning instead attempt to bridge this gap by accumulating external textual skills. However, these methods heavily rely on costly human expert annotations or unreliable LLM-as-a-judge feedback for reflection. To overcome this bottleneck, we propose a novel, unsupervised self-evolving agent framework inspired by the corruption-and-reconstruction paradigm of diffusion models. Instead of relying on explicit external scoring, we leverage existing high-quality human artifacts to construct self-supervised signals. Training then follows the familiar loop of neural network training, forward, loss, and backward, with the loss coming from contrasting the agent's reconstruction against the human original. What is updated is not model weights but an external library of textual skills. We evaluate our framework on the challenging task of short drama screenwriting. Experimental results demonstrate that our method enables the agent to autonomously extract and internalize highly generalizable skills, significantly enhancing its domain-specific generation capabilities. Furthermore, this self-contrastive reflection paradigm offers a scalable pathway for agents to teach themselves the production of complex, high-quality human artifacts, without requiring external supervision.

arXiv

APEX-Accounting

Julien Benchek, Austin Bennett, Jasmin Kern, Ryan Stevens, Rene Sultan, Charis Ching, Hayley Popiel, Vaibhav Mittal, Felix Mercier, Brendan Foody, Bertie Vidgen

We introduce APEX-Accounting, a benchmark built by Mercor in partnership with Ramp, to assess whether frontier models can do the real work of accountants. Tasks include reconciling accounts, accruing expenses, posting transactions, and producing reports. The private eval set comprises 160 tasks, split across 10 worlds. Each world contains an accounting system, as well as spreadsheets, PDFs, and other files. Every task was authored and solved by experts in accounting and bookkeeping, who also wrote grading rubrics. Across nine frontier models, Claude-Fable-5 (Max) leads with 56.4% Mean Criteria@3, ahead of Muse-Spark-1.1 (xHigh) at 52.6%. No model scores more than 2.6% Pass^8 (GPT-5.6-Sol (Max+Pro)) and the highest Pass@8 is 21.5% (Muse-Spark-1.1 (xHigh)). We experiment with increasing the token budget from $1 to $50 and observe an instance of Simpson's paradox: scores increase as the token budget increases but within a given budget-constrained harness, scores are lower on tasks where the model spends more tokens. As APEX-Accounting is a closed benchmark, leaderboard evals can be run for any frontier model on request.

arXiv

From Found to Designed: Concepts as a Design Axis for Large Language Models

Chen Shani

Large language models (LLMs) encode rich concept-like information, but represent it implicitly through distributed statistical associations rather than as explicit, structured, compositional concepts. Consequently, concept-level structure is typically \emph{found} rather than \emph{designed}: it is recovered after training through probing or dictionary learning, with no architectural guarantee of stability, compositionality, controllability, or alignment with human conceptual organization. We organize concept-aware interventions along two dimensions: whether concept structure is internally induced or externally grounded, and the stage of the pipeline where it is introduced. This taxonomy reveals three broad patterns: inference-time approaches remain comparatively underexplored, related ideas have developed largely in isolation across pipeline stages, and externally grounded methods span the entire pipeline despite often being described under different terminology. Together, these observations motivate moving beyond recovering concept-like structure from trained models toward designing LLMs with explicit conceptual representations.

arXiv

Constitutional Midtraining: Content Presence Drives Alignment Gains

Desiree Cho, Cameron Tice, Bernie Hogan, Hunar Batra, Puria Radmard, Jun Zhao, Nigel Shadbolt

Post-training alignment is often shallow, eroding under fine-tuning. It remains untested as to whether constitutional midtraining interventions can produce durable alignment when cleanly isolated from post-training. We build a 394M-token constitutional corpus from Anthropic's Constitution and apply constitutional midtraining at 120B scale, where principled, values-based content is inserted into midtraining. A 2x2 design (curriculum ordering x deliberative reasoning) was used to produce four constitutionally midtrained conditions, plus a control, which were evaluated on self-generated and established benchmarks including alignment under pressure, value conflict resolution, blackmail, and emergent misalignment. All models were evaluated across three stages: post-midtraining, post-SFT, and post-benign fine-tuning. Constitutionally midtrained models outperformed the control on alignment generalization and durability, notably on blackmail: SFT instilled a blackmail propensity in all models, but constitutional midtraining blunted it, with the advantage surviving benign fine-tuning (-17.5pp). This durability did not extend to settings that required active resistance to in-context pressure or conflict, where the advantage attenuates after SFT. The presence of constitutional content at midtraining also mattered more than its structure, and constitutional midtraining incurred no capability cost, on average, at any stage (MMLU, ARC-Easy, piqa, GSM8K). A modest amount of constitutional content at midtraining could therefore yield broad, persistent alignment gains, offering a cheap, complementary addition to SFT-centered pipelines. Code, data, and models are available.

arXiv

Reading Without a Reader: Large Language Models Collapse Reading and Writing into a Single Entangled Code

Diego Saldaña Ulloa

In the literate human brain, reading and writing doubly dissociate: a ventral decoding route (pure alexia) and a fronto-parietal encoding route (pure agraphia), sharing a partial orthographic core. A decoder-only large language model (LLM) drives both from one autoregressive path optimized on text (a \emph{cultural} invention, not an evolved instinct). We ask how entangled it is, comparing an input-side ``reading code'' $\mathbf{W}_{E}$ with an output-side ``writing code'' $\mathbf{W}_{U}$ via an index $\mathcal{E}\in[0,1]$ (CKA, Procrustes residual, mutual $k$-NN) calibrated against an independent-init floor and tied ceiling. On GPT-2, OPT and Pythia (14M--1.4B), untied models hold one \emph{coupled but sub-ceiling} code ($\mathcal{E}=0.23$--$0.35$, far above floor) on a non-monotonic couple-then-differentiate trajectory, $\mathbf{W}_{U}$ drifting $\sim$3.2$\times$ farther than $\mathbf{W}_{E}$ in every decile. Equally informative is a negative: the matching behavioural test, that comprehension and production fail together rather than dissociate, cannot be run. For minimal pairs the alexia analogue is empty by theorem: greedy production implies a vocabulary-wide argmax, so it wins the pairwise ranking. Differential-damage indices are not scale-identified: heavy-tailed damage makes linear standardizations collapse onto their larger term, and the rank transform fixing this is bounded, so its null saturates. Both scores also contain the target's log-probability, which alone explains most of their variance and manufactures the apparent coupling. We withdraw a coupling statistic, a cross-level bridge and a separation measure. In a model reading and writing off one next-token distribution, no output-side pair isolates either ability: entanglement needing no index to see. By analogy, not homology, this situates LLMs in the space of possible minds.

arXiv

Accuracy Hides How Language Models Fail: Measuring Failure States Under Matched Output Budgets

Zongyou Yang, Yinghan Hou

Language-model benchmarks collapse two distinct measurement questions into a single accuracy score: whether a response reached an evaluable state, and whether its answer was judged correct. We introduce a two-layer evaluation framework that separates scorer-independent execution evidence, including termination, answer exposure, parseability, and completion length, from scorer-dependent correctness. Across 2,550 outputs from five fixed Qwen and DeepSeek configurations on MATH and ARC-Challenge, matched 2,048-token limits produce sharply different execution mixtures: 49 of 450 Qwen MATH outputs terminate without a final answer, compared with 5 of 300 DeepSeek MATH outputs and none of the 750 ARC outputs. Among the same 300 DeepSeek MATH question-model pairs, no missing-final length termination is observed at 8,192 tokens. A coverage-audited targeted verification study further shows that candidate-selection and aggregation policies can substantially alter comparative accuracy estimates. These results demonstrate that accuracy conflates execution case mix with verification policy. Evaluations of test-time methods should therefore report pre-intervention execution states, verification coverage, and scorer provenance alongside accuracy.

arXiv

SLAI T-Rex: Full-Parameter Post-training of the DeepSeek-V4 Family on Ascend SuperPOD

Dongfang Li, Xiaodong Luo, Ruoyu Sun, Xuhui Chen, Linyuan Qiu, Jian Meng, Zhengxuan Lu, Yiting Wang, Yucheng Xie, Tao Guo, Tianxiang Fang, Jing Li, Sihang Chen, Shihao Hong, Chang Liu, Weihua Dai, Zirong Zeng, Ziwei Zhu, Zhuohan Wang, Zhengjun Yue, Igor Vasilyev, Min Liu, Weijian Sun, Xin Chen, Yingmeng Gao, Jinhua Zhou, Taolue Chen, Chenwei Wu, Dong Zhang, Wenlong Jin, Jinmin Xiang, Barkova Maria, Ushakov Anton, Xianfei Jin, Tian Ding, Zhihang Lin, Qian Chen, Linxin Yang, Mingzhe Yang, Bingwei Zhang, Hongzhang Yang, Fangxue Zhang, Shijun Qin, Jie Yu, Cuihua Hu, Tolstykh Vasiliy, Nosov Ivan, Abdullin Amir, Zhicheng Zhou, Xin Zhang, Zhixiong Ning, Xutong Zhao, Junjie Huang, Jiajun Liu, Weiyan Kong, Zheng Zhang, Wenhan Luo, Lin Hu, Yangbo Guo, Li Zeng, Shihao Zhang, Baotian Hu, Min Zhang, Haizhou Li, Zhiquan Luo

Full-parameter post-training of trillion-parameter-scale MoE models introduces substantial system-level challenges for large-scale distributed training, including severe memory pressure, non-overlapped communication overhead, and inefficient kernel execution. While most large-scale LLM training systems are built around GPU-based clusters, this report presents an end-to-end optimization practice on the Ascend NPU SuperPOD. Using the DeepSeek-V4 model family as the target workload, we develop a hierarchical optimization framework spanning model-level parallelism, computation-communication orchestration, and low-level kernel execution. The resulting system achieves 34.22% Model FLOPs Utilization (MFU) with a 2.93x improvement over the open-source baseline recipe while maintaining training stability. Building on this optimized infrastructure, we further establish a CPT and SFT workflow for complex Operations Research (OR) tasks. We refer to the integrated framework as SLAI T-Rex. Using DeepSeek-V4-Flash, we develop OR-oriented CPT and SFT data pipelines that combine collected domain resources with solver-verified synthetic optimization documents. The resulting dataset contains 10K high-quality SFT samples spanning four task categories and three problem representations. The specialized model achieves the highest average zero-shot Pass@1 score among the evaluated models, reaching 71.81% and outperforming GPT-5.4-Mini and the base DeepSeek-V4-Flash model by 3.98 and 11.27 percentage points, respectively. Overall, this work demonstrates a full-stack pathway from efficient trillion-parameter model post-training on Ascend infra to domain-specialized Flash models for solver-grounded mathematical modeling, advancing frontier-model systems for complex reasoning.

arXiv

Auditing Question-Order Effects in Large Language Models with the QQ Equality: Mechanism Characterization and a Saturation Caveat

Pilsung Kang

Question-order effects in human survey data have been reported to approximately satisfy the QQ (quantum question) equality, a parameter-free prediction of the standard projective quantum question-order model. We develop this equality into an audit framework for sequential binary judgments of autoregressive large language models (LLMs). Theoretically, we characterize mechanism families that satisfy QQ robustly, show that classical repetition can reproduce the equality exactly, and combine QQ with the rank-2 Contextuality-by-Default criterion through $|q_{QQ}| \le \mathrm{OSS}$. This separates order sensitivity, QQ imbalance, and residual contextuality rather than treating them as interchangeable signatures. Methodologically, we introduce a committed multi-turn forced-branch protocol that reconstructs order-conditioned joint distributions from next-token log-probabilities under counterbalanced label mappings and pre-specified health gates. A first-signal pilot on an open-weight instruction-tuned model reveals the central measurement problem. Although all pre-specified health gates passed, the binary-conditioned distributions were near-deterministic for 17 of 18 item pairs under the direct-evaluation framing and 7 of 8 under the persona framing. Label assignment materially changed several mapping-specific QQ verdicts, and no item was certified as residually contextual. Thus, under the tested conditions, the observed QQ outcomes did not uniquely identify a response mechanism in the presence of a saturated and label-sensitive measurement interface. The main implication is methodological: next-token probabilities should not be interpreted as survey-response distributions without first establishing adequate dispersion. We therefore argue that saturation screening and label counterbalancing should precede structural interpretation in distribution-level audits of LLM judgments.

arXiv

Who Grades the Grader? Co-Evolving Evaluation Metrics and Skills for Self-Improving LLM Agents

Xing Zhang, Guanghui Wang, Yanwei Cui, Ziyuan Li, Wei Qiu, Bing Zhu, Peiyang He

Self-evolving agent systems create, revise, and retire their own skills, but every such loop assumes a reliable evaluation metric already exists. In many real applications none does. We show the metric itself can be the evolving object: our loop searches compositions of small typed drawback detectors under a full evolutionary lifecycle, selecting for agreement with a ten-item anchored reference set and regularizing by consensus over unlabeled outputs. What evolves is the function that grades one output, never the fixed task sets it is scored on, and what comes out is an inspectable expression rather than an opaque judge. It is also valid: on code generation it gains 0.21 agreement with hidden ground truth on a locked set that metric selection never reads (paired $p=0.014$), beating the bare LLM judge it contains. Validity is where safety lives: removing the anchor guards collapses the metric into a vacuous always-pass detector while removing the detector lifecycle does not, inverting the lesson from skill evolution. That collapse warns this line of work that downstream task score cannot validate a self-evolved evaluator, since the collapsed metric trains skills just as well. Task score answers only sufficiency, and an evolved metric suffices: \emph{Double Ratchet}, co-evolving the metric with a lifecycle-managed skill loop, retains 88--110\% of the lift ground truth or a hand-written rubric buys, across MBPP+, Spider~2.0-Snow, and report generation. When evolved skills gamed the report rubric, an independent judge caught it and one added detector repaired it.

arXiv

Exposure is not manifestation: measurement target and output resolution jointly determine which behavioural-faithfulness evaluator wins

Kwan Soo Shin

Behavioural auditing asks whether a language model behaves as it claims, but detection scores are reported without separating two targets: whether a reply was produced under a behaviour-inducing condition (exposure) and whether the behaviour surfaced in it (manifestation). Scoring a compact 146-million-parameter auditor's frozen-representation read-out and a frontier judge against each label on the identical 720 replies, the gap between the instruments moves by roughly 0.2 AUROC when the target changes. Under the judge's deployed interface, a single verdict, the ranking reverses: the auditor leads on exposure, 0.804 against 0.718, and trails on manifestation, 0.690 against 0.811. Matching the output resolution from either direction, by asking the judge a target-specific question answered with a continuous confidence score or by thresholding the auditor's read-out, removes the reversal but not the interaction, which excludes zero at all three resolutions (0.207, 0.237 and 0.169). The target governs how far apart the instruments are; the interface governs whether that distance changes their order. The auditor's hyperbolic geometry confers no advantage here. A single behavioural-detection AUROC is under-specified: such claims are comparable only when they state the estimand, the evaluator, and its output interface.

arXiv

MEDIAREF: A Public Knowledge Store for Media Background Checks

Benjamin Nichols, Michael Schlichtkrull, Nedjma Ousidhoum

LLM-based retrieval-augmented generation (RAG) is increasingly used for automated fact-checking (AFC) and related tasks. By grounding LLM outputs in retrieved evidence, RAG-based systems provide transparent justifications while allowing external information to be updated independently of the underlying model. However, existing approaches often assume retrieved evidence is reliable, although real-world information may be conflicting, outdated, and can originate from unreliable or biased sources. Recent work on *source-critical reasoning* addresses this challenge through media background checks (MBCs) (Schlichtkrull, 2024), which assess the credibility of evidence sources to support downstream fact verification. However, generating MBCs relies on costly proprietary search APIs, limiting reproducibility. To mitigate this issue, we introduce MEDIAREF, a publicly available knowledge store of web-sourced documents that enables reproducible, low-cost evaluation of MBC generation across 200 media sources. We describe a reproducible methodology for constructing and updating the collection, assess widely used LLMs on the MBC generation task, and demonstrate that MEDIAREF supports higher-quality MBC generation through both automatic and qualitative evaluation.

arXiv

The MiniMax-M2 Series: Mini Activations Unleashing Max Real-World Intelligence

Aili Chen, Aonian Li, Baichuan Zhou, Bangwei Gong, Binyang Jiang, Boji Dan, Changhao Zhang, Changqing Yu, Chao Wang, Cheng Ma, Cheng Zhong, Cheng Zhu, Chengjun Xiao, Chengyi Yang, Chengyu Du, Chenyang Zhang, Chi Zhang, Chuangyi Huang, Chunhao Zhang, Chunhui Du, Chunyu Zhao, Congchao Guo, Da Chen, Deming Ding, Dianjun Sun, Dong Li, Dongyu Zhang, Enhui Yang, Fei Yu, Guang Zheng, Guodong Zheng, Guohong Li, Haichao Zhu, Haigang Zhou, Haimo Zhang, Han Ding, Hao Zhang, Haohai Sun, Haolin Lyu, Haonan Lu, Haoyu Wang, Huajie Shi, Huiyang Li, Jiacheng Chen, Jian Zhang, Jiaqi Zhuang, Jiaren Cai, Jiaxin Pan, Jiayao Li, Jiayuan Song, Jichuan Zhang, Jie Wang, Jihao Gu, Jin Zhu, Jingwei Dong, Jingyang Li, Jingyu Zhang, Jingze Zhuang, Jinhao Tian, Jinli Liu, Jinyi Hu, Jun Tao, Jun Zhang, Junbin Ruan, Junhao Xu, Junjie Yan, Junteng Liu, Junxian He, Kang Xu, Ke Ji, Ke Yang, Kecheng Xiao, Keyu Duan, Keyu Li, Le Han, Letian Ruan, Li Yuan, Lianfei Yu, Liheng Feng, Lijie Mo, Lin Li, Linge Du, Lingye Bao, Lingyu Yang, Lingyuan Zhou, Loki, Lu Chen, Lunbin Zeng, Ming Li, Ming Zhong, Mingliang Tao, Mingyuan Chi, Mujie Lin, Nan Hu, Ningxin Chen, Peiyin Zhu, Peng Gao, Pengcheng Gao, Pengfei Li, Penglin Li, Pengyu Zhao, Qibin Ren, Qibing Ren, Qidi Xu, Qihan Ren, Qile Li, Qin Wang, Quanliang Chen, Qunhong Zeng, Rong Tian, Rongxin Guo, Rui Dong, Ruitao Leng, Ruize Zhang, Shanqi Liu, Shaoxiang Chen, Shaoyu Chen, Sheng Jia, Shun Yao, Shuoran Zhao, Shuqi Yu, Sichen Li, Sicheng Pan, Songquan Zhu, Tengfei Li, Tian Xie, Tiancheng Qin, Tianle Li, Tianrun Liang, Wei Liu, Weiqi Xu, Weitao Li, Weixiang Chen, Weiyu Cheng, Weiyu Zhang, Wenhu Chen, Wenqian Zhao, Xiancai Chen, Xiangjun Song, Xiangyuan Wang, Xianzhen Luo, Xiao Luo, Xiao Su, Xiaobo Li, Xiaodong Han, Xiaojie Wu, Xihao Song, Xingyi Han, Xinyu Guan, Xuan Lu, Xun Zou, Xunhao Lai, Xutong Li, Xuyang Shen, Yan Gong, Yan Ma, Yang Jiao, Yang Wang, Yang Xu, Yangsen Wang, Ye Tang, Yicheng Chen, Yihang Wang, Yinran Qiu, Yiqi Shi, Yiting Guo, Yiwen Huang, Yixuan Wang, Yongyi Hu, Yu Gao, Yu Zhang, Yuan Li, Yuanxiang Ying, Yuanzhen Zhang, Yubo Wang, Yuchen Song, Yufeng Yang, Yuhang Meng, Yuhang Miao, Yuhao Li, Yujie Liu, Yulin Hu, Yunan Huang, Yunji Li, Yunyi Huang, Yusen Zhang, Yusu Hong, Yutao Xie, Yutong Zhang, Yuwen Liao, Yuxuan Shi, Yuze Wenren, Zebin Li, Zehan Li, Zejian Luo, Zeyu Jin, Zeyuan Sun, Zhanpeng Zhou, Zhaochen Su, Zhendong Li, Zhengmao Zhu, Zhengyuan Peng, Zhenhua Fan, Zhi Zhang, Zhichao Xu, Zhiheng Lv, Zhikang Xu, Zhitao He, Zhiwei He, Zhongyuan Li, Zibo Gao, Zijia Wu, Zijian Song, Zijian Zhou, Zijun Sun, Zishan Huang, Ziying Chen, Ziyue Ge

We introduce the MiniMax-M2 series, a family of Mixture-of-Experts language models built around the principle that mini activations can unleash maximum real-world intelligence. The flagship M2 contains 229.9B total parameters with only 9.8B activated per token. Designed end-to-end for agentic deployment, the M2 series rests on three components: (i) agent-driven data pipelines producing large-scale, verifiable trajectories across agentic coding and agentic cowork, each grounded in an executable workspace and an artifact-aligned reward; (ii) Forge, a scalable agent-native RL system that adapts to long-horizon agent trajectories, paired with windowed-FIFO scheduling, prefix-tree merging, inference optimization, and a clean training-inference-agent decoupling that supports both white-box and black-box agents; (iii) the latest M2.7 checkpoint takes an early step toward self-evolution -- autonomously debugging training runs and modifying its own scaffold. Across M2 through M2.7, this combination translates a mini-activation footprint into frontier-tier performance on agentic coding, deep search, office-task, and reasoning benchmarks.

arXiv

MinerU-Popo: Universal Post-Processing Model for Structured Document Parsing

Bangrui Xu, Ziyang Miao, Xuanhe Zhou, Yiming Lin, Zirui Tang, Xiaomeng Zhao, Fan Wu, Cheng Tan, Bin Wang, Conghui He

VLM-based OCR models have become the de facto choice for document parsing, as they can accurately extract page-level elements (e.g., paragraphs within individual pages) together with their bounding boxes and textual content. However, downstream applications such as RAG require coherent document-level information, whereas these models often break cross-page continuity and fail to recover disrupted structures, such as paragraphs and tables truncated by page boundaries. Such relationships are not confined to a single page; instead, they require joint analysis of titles, paragraphs, tables, and images spanning multiple pages. A natural solution is therefore to reuse existing OCR outputs and reconstruct document-level logical structures through post-processing. To this end, we propose MinerU-Popo, a lightweight and universal framework for POst-Processing OCR outputs, which converts page-level results from diverse parsers into coherent document-level structures. MinerU-Popo decomposes the problem into four focused subtasks: text truncation recovery, table truncation recovery, title hierarchy reconstruction, and image-text association. To address these effectively, we build a task-oriented data engine with task-specific input filtering, and use the generated data (30K) to fine-tune a lightweight post-processing model (Qwen3-VL-4B). To support long documents, we introduce dynamic chunking with overlap-based synchronization, which aligns chunk-level outputs from the fine-tuned model and preserves global consistency. Finally, we assemble the aligned outputs into a tree-structured document representation, further enriched with node chunking and summaries for downstream retrieval and analysis. Empirical results show MinerU-Popo improves title-hierarchy TEDS by at least 20% across all five tested OCR models, improves RAG accuracy and reduces per-query latency.

arXiv

Orchard: An Open-Source Agentic Modeling Framework

Baolin Peng, Wenlin Yao, Qianhui Wu, Hao Cheng, Xiao Yu, Rui Yang, Tao Ge, Alessandro Sordoni, Xingdi Yuan, Yelong Shen, Pengcheng He, Tong Zhang, Zhou Yu, Jianfeng Gao

Agentic modeling aims to transform LLMs into autonomous agents capable of solving complex tasks through planning, reasoning, tool use, and multi-turn interaction with external environments. We present Orchard, an open-source framework for scalable agentic modeling. At its core is Orchard Env, a lightweight Kubernetes-native environment service that provides reusable primitives for sandbox lifecycle management across task domains, agent harnesses, and training stages. On top of Orchard Env, we build three agentic modeling recipes. Orchard-SWE targets software engineering agents. We introduce credit-assignment supervised fine-tuning and a progression of RL signals: Balanced Adaptive Rollout (BAR) for sparse-reward optimization, on-policy distillation (OPD) and rubric-based process reward (RPR) for dense supervision, and historical experience distillation, which compresses rollouts from prior experiments into a compact value model for inference-time reranking. Built on the Qwen3.5-35B-A3B backbone, Orchard-SWE reaches 69.7% with RPR-based RL and 73.0% with value-model reranking on SWE-bench Verified, setting a new state of the art among open-source methods while approaching frontier systems over 10x larger. Orchard-GUI trains a 4B vision-language computer-use agent using only 0.4K distilled trajectories and 2.2K open-ended tasks, achieving 68.4% average success across WebVoyager, Online-Mind2Web, and DeepShop, making it the strongest open-source model while remaining competitive with proprietary systems. Orchard-Claw targets personal assistant agents. Trained with only 0.2K synthetic tasks, it achieves 59.6% pass@3 on Claw-Eval and 73.9% when paired with the stronger ZeroClaw harness. Collectively, these results demonstrate that a lightweight, open, harness-agnostic environment layer enables reusable agentic data, training recipes, and evaluation protocols across domains.

arXiv

VISTA: A Controllable Platform for Generating and Auditing Egocentric Assistance Scenarios

Yu-Hsiang Liu, Yu-Chien Tang, An-Zi Yen

Evaluating whether AI agents can proactively assist humans in daily activities, ranging from routine household tasks to urgent safety-critical situations, requires diverse visual data. However, collecting such scenarios in the real world is often difficult, costly, or unsafe, and simulation environments often lack the social commonsense needed to simulate the consequences of different actions. In this work, we present VISTA, a controllable platform that uses a user-provided scenario seed, defined as a short natural-language description of the intended assistance situation, to generate editable plans, egocentric videos, and an auditable review trail. VISTA structures scenario intent around three interaction modes, including reactive, explicit proactive, and implicit proactive, and two consequence families, including safety-critical and everyday inconvenience, with no-assistance cases as controls. Its six-stage pipeline exposes the design brief, timed event script, first-frame plan, motion plan, and video plan, allowing users to revise each artifact in natural language before explicitly authorizing media generation. A human evaluation shows that videos retained by the complete VISTA workflow align more closely with their scenario seeds than outputs from two one-pass baselines. VISTA thereby makes targeted egocentric scenario generation inspectable, revisable, and empirically auditable.

arXiv

Beyond Pattern Matching: Seven Cross-Domain Techniques for Prompt Injection Detection

Thamilvendhan Munirathinam

Current open-source prompt-injection detectors converge on two architectural choices: regular-expression pattern matching and fine-tuned transformer classifiers. Both share failure modes recent work has made concrete. Regular expressions miss paraphrased attacks. Fine-tuned classifiers are vulnerable to adaptive adversaries: a 2025 NAACL Findings study reported that eight published indirect-injection defenses were bypassed with greater than fifty percent attack-success rates under adaptive attacks. This work proposes seven detection techniques that each port a mechanism from a discipline outside LLM security: forensic linguistics, materials-science fatigue analysis, deception technology, local-sequence alignment from bioinformatics, mechanism design, spectral signal analysis, and taint tracking. Each produces a signal architecturally independent of both regex matching and transformer classification, so the seven compose with existing defenses rather than replacing them. Four of seven are now implemented in prompt-shield v0.7.3 (Apache 2.0): d028 sequence alignment, d027 stylometric discontinuity, materials-fatigue tracking, and d034 honeypot tool definitions (new in v4.0). A four-configuration ablation across nine benchmarks (~10,300 samples) covers deepset, NotInject, LLMail-Inject, AgentHarm, AgentDojo, and an independent evaluation against three peer-reviewed academic benchmarks (Liu USENIX 2024, Garak, InjecAgent). This revision adds Section 5.7 (composed-stack adaptive-attack partial run, evidence for the composability thesis), Section 5.8 (50-document held-out benchmark: d027 collapses 1.000 to 0.000 F1 in isolation but the composed engine recovers 0.815 F1), and Section 7 (three architectural patterns extracted from prompt-shield in Gang-of-Four format). All code, data, and reproduction scripts are released Apache 2.0.

arXiv

S-GRPO: Unified Post-Training for Large Vision-Language Models

Yuming Yan, Kai Tang, Sihong Chen, Ke Xu, Dan Hu, Qun Yu, Pengfei Hu

Current post-training methodologies for adapting Large Vision-Language Models (LVLMs) generally fall into two paradigms: Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL). Despite their prevalence, both approaches suffer from inefficiencies when applied in isolation. SFT forces the model's generation along a single expert trajectory, often inducing catastrophic forgetting of general multimodal capabilities due to distributional shifts. Conversely, RL explores multiple generated trajectories but frequently encounters optimization collapse - a cold-start problem where an unaligned model fails to spontaneously sample any domain-valid trajectories in sparse-reward visual tasks. In this paper, we propose Supervised Group Relative Policy Optimization (S-GRPO), a unified post-training framework that integrates the guidance of imitation learning into the multi-trajectory exploration of preference optimization. Tailored for direct-generation visual tasks, S-GRPO introduces Conditional Ground-Truth Trajectory Injection (CGI). When a binary verifier detects a complete exploratory failure within a sampled group of trajectories, CGI injects the verified ground-truth trajectory into the candidate pool. By assigning a deterministic maximal reward to this injected anchor, S-GRPO enforces a positive signal within the group-relative advantage estimation. This mechanism reformulates the supervised learning objective as a high-advantage component of the policy gradient, compelling the model to dynamically balance between exploiting the expert trajectory and exploring novel visual concepts. Theoretical analysis and empirical results demonstrate that S-GRPO gracefully bridges the gap between SFT and RL, drastically accelerates convergence, and achieves superior domain adaptation while preserving the base model's general-purpose capabilities.

arXiv

How Can We Synthesize High-Quality Pretraining Data? A Systematic Study of Prompt Design, Generator Model, and Source Data

Joel Niklaus, Atsuki Yamaguchi, Michal Štefánik, Guilherme Penedo, Hynek Kydlíček, Elie Bakouch, Lewis Tunstall, Edward Emanuel Beeching, Thibaud Frere, Colin Raffel, Leandro von Werra, Thomas Wolf

Synthetic data is a standard component in training large language models, yet systematic comparisons across design dimensions, including rephrasing strategy, generator model, and source data, remain absent. We conduct extensive controlled experiments, generating over one trillion tokens, to identify critical factors in rephrasing web text into synthetic pretraining data. Our results reveal that structured output formats, such as tables, math problems, FAQs, and tutorials, consistently outperform both curated web baselines and prior synthetic methods. Notably, increasing the size of the generator model beyond 1B parameters provides no additional benefit. Our analysis also demonstrates that the selection of the original data used for mixing substantially influences performance. By applying our findings, we develop \textbf{\textsc{FinePhrase}}, a 486-billion-token open dataset of rephrased web text. We show that \textsc{FinePhrase} outperforms all existing synthetic data baselines while reducing generation costs by up to 30 times. We provide the dataset, all prompts, and the generation framework to the research community.

arXiv

Personalized RewardBench: Evaluating Reward Models with Human Aligned Personalization

Qiyao Ma, Dechen Gao, Rui Cai, Boqi Zhao, Hanchu Zhou, Junshan Zhang, Zhe Zhao

Pluralistic alignment has emerged as a critical frontier in the development of Large Language Models (LLMs), with reward models (RMs) serving as a central mechanism for capturing diverse human values. While benchmarks for general response quality are prevalent, evaluating how well reward models account for individual user preferences remains an open challenge. To bridge this gap, we introduce Personalized RewardBench, a novel benchmark designed to rigorously assess reward models' capacity to model personalized preferences. We construct chosen and rejected response pairs based on strict adherence to (or violation of) user-specific rubrics, ensuring that preference distinctions are uniquely tailored to the individual. In particular, human evaluations confirm that the primary discriminative factor between pairs is strictly personal preference, with both responses maintaining high general quality (e.g., correctness, relevance and helpfulness). Extensive testing reveals that existing state-of-the-art reward models struggle significantly with personalization, peaking at an accuracy of just 75.94%. Crucially, because an effective reward model benchmark should predict a reward model's performance on downstream tasks, we conduct experiments demonstrating that our benchmark exhibits a significantly higher correlation with downstream performance in both Best-of-N (BoN) sampling and Proximal Policy Optimization (PPO) compared to existing baselines. These findings establish Personalized RewardBench as a robust and accurate proxy for evaluating reward models' performance in downstream applications.

arXiv

GradMAP: Faster Layer Pruning with Gradient Metric and Projection Compensation

Hao Liu, Guangyan Li, Wensheng Zhang, Yongqiang Tang

Large Language Models (LLMs) exhibit strong reasoning abilities, but their high computational costs limit their practical deployment. Recent studies reveal significant redundancy in LLMs layers, making layer pruning an active research topic. Layer pruning research primarily focuses on two aspects: measuring layer importance and recovering performance after pruning. Unfortunately, the present works fail to simultaneously maintain pruning performance and efficiency. In this study, we propose GradMAP, a faster layer pruning method with \textbf{Grad}ient \textbf{M}etric \textbf{A}nd \textbf{P}rojection compensation, which consists of two stages. In the first stage, we introduce a novel metric based on gradient magnitudes, enabling a global assessment of layer importance. Note that, it requires only a single backward propagation step per pruning decision, substantially enhancing pruning efficiency. In the second stage, we first analyze the layers with the largest mean shift resulting from pruning, and then incorporate a simple yet effective projection compensation matrix to correct this drift in one step. In this way, the degradation of model performance caused by layer pruning is effectively alleviated. Extensive experiments show that GradMAP outperforms previous layer pruning methods in both pruning speed (achieving an average $4\times$ speedup) and performance.

Keep browsing

More papers from top AI conferences

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