Academic Intelligence · Curated Daily

探索全球前沿学术脉络

AcademicHub 汇聚顶级期刊与预印本平台的实时文献。定制您的专属科研雷达,利用大语言模型自动生成交叉领域文献分析简报。

01.
arXiv (CS.CL) 2026-06-12

TAB-PO: Preference Optimization with a Token-Level Adaptive Barrier for Token-Critical Structured Generation

Direct Preference Optimization (DPO) is an effective and widely adopted approach for offline alignment but is poorly matched to ontology-driven structured prediction, where preferred and rejected JSON objects often differ in only a few schema-defining tokens. In this low-edit-distance regime, sequence-level DPO spreads gradient mass across non-critical serialization tokens (gradient dilution) and can reduce likelihood on rare, under-confident preferred schema tokens (token erosion). To address these limitations, we first develop a confusion-aware preference-construction strategy that augments expert-curated ambiguity patterns with empirical structured-error modes estimated from validation-set SFT predictions, synthesizing minimally perturbed, schema-valid negatives that focus preference learning on realistic ontology-level decision errors. We then introduce Token-Adaptive Barrier Preference Optimization (TAB-PO), a post-SFT objective for token-critical structured generation. TAB-PO adds a confidence-gated token-level barrier that applies supervised anchoring to under-confident schema tokens. On the public SciERC scientific information extraction task, evaluated with Llama/Qwen models from 1.5B to 70B, TAB-PO improves ontology-critical semantic-label and relational-linking metrics over SFT by 11.59% on average, wins 100% of comparisons against the strongest token-level and sequence-level DPO variants on these metrics, and surpasses leading frontier models by 14.71%, while delivering strong gains in textual grounding.

02.
arXiv (CS.CL) 2026-06-11

ProcessThinker: Enhancing Multi-modal Large Language Models Reasoning via Rollout-based Process Reward

Visual question answering increasingly requires multi-step reasoning. Recent post-training with reinforcement learning under verifiable rewards (RLVR) and Group Relative Policy Optimization (GRPO) can improve multimodal reasoning, but most approaches rely on sparse outcome-only rewards. As a result, they struggle to tell whether an incorrect answer comes from a small mistake late in the reasoning or from an unhelpful trajectory from the start. A common solution is to train a process reward model (PRM) for step-level supervision, but this typically requires large-scale high-quality chain-of-thought annotations and additional training cost. We propose ProcessThinker, a practical post-training pipeline that provides step-level process rewards without training an explicit PRM. ProcessThinker first rewrites reasoning traces into a step-tagged format for cold-start supervised fine-tuning, then applies GRPO with a standard format reward and our rollout-based process reward. Concretely, for each intermediate step, we sample multiple continuations from that step and use the empirical success rate (final-answer verification) as the step reward. This gives dense credit assignment and encourages reasoning steps that more reliably support a correct conclusion, helping reduce inconsistent or self-contradictory progress across steps – a key issue in logical reasoning. Across four challenging video benchmarks (Video-MMMU, MMVU, VideoMathQA, and LongVideoBench), ProcessThinker consistently improves over the baseline model Qwen3-VL-8B-Instruct

03.
arXiv (CS.CV) 2026-06-15

Gefen: Optimized Stochastic Optimizer

AdamW is a default optimizer for modern deep learning, but its first and second moment states add roughly two parameter-sized buffers to training memory. We propose Gefen, a memory-efficient optimizer that automatically shares second-moment estimates across parameter blocks and quantizes the first moment using a learned codebook, thereby reducing AdamW's memory footprint by ~8x while maintaining the same performance, corresponding to a reduction of 6.5 GiB per billion parameters. The method is motivated by a theoretical result showing that large mixed Hessian entries constrain the ratio of squared gradients toward one, suggesting that Hessian-aligned parameters are natural candidates for sharing second-moment statistics. Since computing Hessians is impractical at scale, Gefen infers block structure from the initial squared gradients, requiring no architecture-specific metadata or hyperparameters beyond AdamW defaults. Gefen learns an exact histogram-based dynamic-programming quantization codebook and reuses the same blocks for first-moment scaling. Across diverse experiments, Gefen achieves the lowest peak optimizer memory among the compared AdamW-like methods while maintaining AdamW-level performance. In FSDP and DDP training, the reduced memory footprint enables larger microbatches and improves throughput significantly over AdamW, providing a practical drop-in replacement with lower memory usage that can increase throughput and enable training larger models or using larger batch sizes. We provide the complete Python implementation, including fused CUDA kernels at https://github.com/ndvbd/Gefen

04.
arXiv (CS.AI) 2026-06-17

LongWebBench: Evaluating Structural and Functional Webpage Generation in Long-Horizon Settings

arXiv:2606.17727v1 Announce Type: new Abstract: Recent vision-language models (VLMs) have shown promising progress in generating webpages from visual inputs, yet existing evaluations mainly focus on short, single-screen, and largely static webpages. We introduce LongWebBench, a benchmark for evaluating long-horizon webpage generation from both structural and functional perspectives. LongWebBench contains 490 real-world long webpages for structural fidelity evaluation and 507 goal-oriented interaction tasks over 129 webpages for functional evaluation. It employs two complementary protocols: a multi-dimensional VLM-based metric for assessing long-range structural coherence, and a DOM-augmented agent-based pipeline for end-to-end functional verification. We further examine the automatic evaluation protocols through human agreement analysis. Experiments with state-of-the-art open-source and proprietary VLMs under single-image and multi-image settings reveal that structural fidelity degrades as webpage length increases, while visually plausible generations often fail to support executable multi-step interactions. These results highlight the need to evaluate long webpage generation beyond visual similarity, with executable interaction as a core criterion. Our code and data are available at https://github.com/zheny2751-dotcom/LongWebBench.

05.
arXiv (CS.LG) 2026-06-19

Weibull Weight-Scale Parameter Evolution under AdamW Training Dynamics

作者:

arXiv:2606.19367v1 Announce Type: new Abstract: Building on a two-parameter Weibull framework for diagnosing transformer weight distributions, we study why the Weibull weight-scale parameter $\lambda$ grows, overshoots, and then relaxes during AdamW training. We derive a leading-order three-force decomposition of the squared weight norm from the AdamW update: an alignment force measuring the correlation between weights and the adaptive update direction, an injection force from adaptive step magnitude, and a decay force from decoupled weight decay. On self-trained Pythia-70M models with ground-truth optimizer moments, alignment dominates the rise phase, contributing 88-94% of the absolute force budget across four random seeds and remaining robust to super-weight removal. Near saturation, alignment and decay approach balance, explaining the transition from weight-scale growth to relaxation. These force dynamics directly govern the squared-norm component underlying $\lambda(t)$; the remaining RMS-to-Weibull reconstruction offset is measurable and decomposes into bridge and integration components, totaling approximately 5-6% in densely sampled regions. To extend the analysis to real models where optimizer moments are unavailable, we introduce a spline displacement method that recovers the alignment force from sparse checkpoints with approximately 92-94% accuracy, about twice the naive two-point baseline. We further observe that the peak value of $\lambda(t)$ varies with training-data coherence in our experiments, suggesting a data-dependent component of weight-scale growth that we leave to a controlled follow-up study. Code and data are available at https://github.com/tiexinding/NPM-Weibull-public.

06.
arXiv (CS.CL) 2026-06-17

Unintended Effects of Geographic Conditioning in Large Language Models

Modern conversational AI systems frequently rely on user metadata to localize responses, yet the unintended regional biases introduced by this hidden context remain poorly understood. In this work, we evaluate location leakage: the phenomenon where a model generates geographic references despite receiving a geographically neutral user prompt. Across both creative writing and open-ended Q&A prompts, even state-of-the-art LLMs systematically favor region-specific outputs when exposed to location metadata, with leakage spiking by up to 793 times above baseline (e.g., from 0.04% to 31.7% for Llama 3.1-8B, and 21.3% and 8.8% for Qwen3-8B and Claude Sonnet 4.6, respectively). Our analysis further shows a novel structural conditioning effect: replacing the injected location with the placeholder "Unknown" still elevates leakage by up to 72 times above baseline, demonstrating that the user profile frame itself, independent of any geographic content, acts as a generative conditioning signal.

07.
arXiv (CS.CL) 2026-06-16

Understanding Scam Trends and Rail Paths from Reddit Self-Disclosure Narratives

Online scam behavior is inherently multi-stage, and the lifecycle includes temporally ordered rails and events rather than isolated signals. Existing works analyze characteristics of scam types and rails, but they do not track scam trends across years. Moreover, the work on the relations between rails is hampered due to the lack of open-source datasets with annotations and coverage of different scam types. To address these gaps, we build a dataset to analyze the yearly trend of scam characteristics and rail paths using Reddit self-disclosure narratives from 2023 to 2025. We collect 21,304 posts from scam-related subreddits with at least one rail among identity, communication, platform, and payment for trend analysis by heuristic annotation. Then, we label 1,800 posts containing explicit or recoverable scam chains by an LLM-assisted method for scam path analysis. The method is evaluated with human annotation. Lastly, we run a topic model on the comments of the posts to analyze the community support behavior. The results reveal that scam processes are predominantly multi-rail. Across years, different scam types and rail components dominate. Different scam types vary systematically in path complexity. Reddit support behaviors have become more detailed over time. This work supports synthetic scam chain data simulation and AI-related scam risk assessment, though findings may not generalise to other platforms.

08.
arXiv (CS.CL) 2026-06-15

Be My Tutor: On-Policy Co-Distillation for Mutual LLM Improvement via Peer Feedback

We study multi-domain LLM training in which two models, each stronger in a different domain, co-evolve by tutoring each other through on-policy feedback. Unlike one-way distillation or single-model fine-tuning, our goal is mutual Pareto improvement: each model improves across domains without losing its original strength. To this end, we propose On-Policy Co-Distillation (OPCoD), where each student's self-distillation is conditioned on its own correct rollout and feedback from its peer. To make feedback exchange effective, OPCoD uses cognizance-based gating to decide when to give feedback and feedback anchoring to ground feedback in the problem. On Science Q\&A tasks, OPCoD consistently outperforms baselines and achieves Pareto improvement across all evaluated domain pairs and students.

09.
arXiv (CS.CV) 2026-06-18

Confidence is Not Reliability: Rethinking MC Dropout in Brain Tumour Segmentation

Glioma segmentation in multiparametric MRI is a critical component of treatment planning. A segmentation model that fails silently on treatment-critical sub-regions represents a patient safety risk that overlap-based metrics such as Dice scores cannot expose. We ask whether voxel-level uncertainty estimation via Monte Carlo (MC) Dropout can reliably identify segmentation errors in clinically critical sub-regions, and whether calibration failure modes are detectable from standard reporting metrics alone. In an empirical two-model case study on 126 BraTS21 patients, we evaluate a high-performance pretrained SegResNet and a locally trained UNet with residual units (UNet-Res). MC dropout preserved segmentation accuracy ($|\Delta Dice|$ $

10.
arXiv (CS.LG) 2026-06-15

LoMC: Localized Multidirectional Correction for Refusal Suppression in Routed Foundation Models

arXiv:2606.13709v1 Announce Type: cross Abstract: We study controlled post-training refusal suppression in routed MoE and hybrid-MoE foundation models, aiming to increase non-refusal target-response behavior while preserving general capability under a compact intervention footprint. Existing broad direction-based edits can perturb general-purpose computation, whereas support-only expert edits often lack sufficient capacity to correct heterogeneous refusal representations. To address this limitation, we introduce Localized Multidirectional Correction (LoMC), a support-gated intervention framework that follows a support-then-correction execution order: it first identifies a compact edit support, then aggregates prototype correction directions into layer-wise correction directions, and finally applies rank-one layer-wise correction only within the selected support. By using the edit support as a structural gating constraint, LoMC increases correction capacity without expanding the intervention scope. Experiments on text-only and multimodal safety benchmarks across four routed backbones show that LoMC substantially improves non-refusal target-response behavior while maintaining general capability under a compact intervention footprint.

11.
arXiv (CS.CV) 2026-06-17

Response-Aware Multimodal Learning for Post-Treatment Visual Acuity Forecasting

Long-term visual acuity (VA) forecasting after anti-VEGF therapy is important for counseling and follow-up planning in diabetic macular edema (DME), yet remains challenging when only early post-treatment findings are available. While prior OCT-based methods mainly focus on short-term response or single-endpoint prediction, multi-horizon VA forecasting from early longitudinal data remains insufficiently under-explored. In this study, we assembled a real-world cohort of 188 anti-VEGF–treated DME patients with paired baseline and month-1 OCT scans, along with tabular OCT-derived biomarkers and non-imaging clinical variables. Using only these early data, we formulate a multi-horizon VA forecasting problem aimed at predicting visual outcomes at 3, 6, 12, 18, and 24 months, reflecting clinically meaningful follow-up intervals. We propose ReVA, a response-aware multimodal framework that combines baseline and month-1 OCT features with tabular variables to capture disease status and early treatment response. ReVA integrates spatial OCT attention, dependency-aware tabular encoding, and cross-modal fusion to predict patient-specific long-term VA trajectories. The proposed framework achieves MAE=0.1246, RMSE=0.1621, and R^2=0.6064 for 24-month VA prediction, with consistent performance across all forecast horizons. Our findings show that incorporating early treatment-response signals enables clinically meaningful long-term visual acuity forecasting, supporting data-driven decision support for routine anti-VEGF management. Code and pretrained models will be released on https://github.com/nguyenpbui/ReVA.

12.
medRxiv (Medicine) 2026-06-15

Therapeutic efficacy study on shoulder impingement syndrome in swimmers: a network meta-analysis

Shoulder impingement syndrome (SIS), including subacromial impingement and rotator cuff tendinitis, is commonly caused by repetitive swimming movements and associated shoulder joint dysfunction. Despite numerous available treatment options, no consensus exists on the most effective treatment option. Therefore, this systematic review and network meta-analysis aimed to investigate treatment methods for SIS in swimmers. Using a frequentist framework and Cochrane PICOS principles, we compared SIS treatments, constructed network evidence diagrams, and assessed heterogeneity. A total of 45 studies were included in the qualitative synthesis, and 42 contributed to the network meta-analysis, comprising 1752 participants, 9 treatment categories, and outcome measures. For pain outcomes, some adjunctive interventions combined with exercise showed favorable ranking probabilities, although several estimates were accompanied by wide confidence intervals. For shoulder range-of-motion outcomes, taping, acupuncture, manual therapy, and sport-specific training showed favorable effects in selected comparisons, particularly for external and internal rotation. According to surface under the cumulative ranking curve (SUCRA) rankings, exercise combined with medium-frequency therapy ranked highly for pain reduction, whereas exercise combined with acupuncture or extracorporeal shock wave therapy ranked highly for shoulder flexion. Exercise combined with taping ranked highly for external rotation, and exercise combined with manual therapy ranked highly for internal rotation. However, the interpretation of ranking results should remain cautious because uncertainty and inconsistency were present in some comparisons. Exercise-based rehabilitation appears to remain central to the management of SIS in swimmers. Several adjunctive interventions showed favorable findings for selected outcomes, especially pain relief and shoulder rotational function. However, the available evidence was affected by heterogeneity, inconsistency, and imprecision across some treatment comparisons. More rigorously designed swimmer-specific randomized controlled trials are needed before firm treatment hierarchies can be established. Trial registration: The protocol for this systematic review is registered with PROSPERO (www.crd.york.ac.uk/PROSPERO; registration number: CRD42024498851). The first submission of PROSPERO was on January 15, 2024, and it was revised and updated on March 25, 2026.

13.
arXiv (math.PR) 2026-06-16

A small noise approximation for Muller's Ratchet

arXiv:2606.15842v1 Announce Type: new Abstract: We consider an infinite system of SDEs with Fleming-Viot noise indexed by $k=0,1,2,\dots$, whose parameters $\alpha,\lambda$, and $\nu$ are the (deleterious) selection coefficient, the (uni-directional) mutation rate, and a quantity which determines the size of the system's fluctuations. The SDE's unique weak solution $X(t) = (X_k(t))_{k=0,1,2,...}$ models what is known in population genetics as Muller's ratchet. Here, $X_k(t)$ stands for the frequency of individuals carrying $k$ deleterious mutations. Since the mutation process is uni-directional, $t\mapsto \inf\{k: X_k(t)> 0\}$ is non-decreasing for almost every path of $X$, and we refer to an increase as a click of Muller's ratchet. A long standing question concerns the clicking rate of Muller's ratchet. Using Duhamel's principle for semigroups, we give a partial answer by approximating $E(\sum_{k=1}^\infty kX_k(t) )$ and $E\big(X_0(t)\big)$ up to $O(1/\nu^2)$ for fixed $\alpha$, $\lambda$ and $t>0$. Our results suggest that $\psi:=\nu \alpha e^{-\lambda/\alpha}$ is a crucial quantity also when the mutation/selection ratio $\theta = \lambda/\alpha$ is moderately large: for large $\nu \alpha$, clicking of the ratchet on the time scale $\frac 1\alpha \log \theta$ becomes rare as soon as $\psi$ becomes large.

14.
arXiv (quant-ph) 2026-06-16

Grid-state deformation in a no-jump non-Hermitian bosonic dimer

arXiv:2606.17036v1 Announce Type: new Abstract: We study the no-jump evolution of ideal grid states in a lossy bosonic dimer with differential decay. The effective non-Hermitian quadratic dynamics induces a complex symplectic flow in phase space that deforms both the primitive lattice vectors and the origin seed. The average decay rate controls common attenuation, while coherent hopping and differential decay control the reduced dimer deformation. The reduced sector contains elliptic, parabolic, and hyperbolic regimes with imaginary spectra, an exceptional point, and real spectra, producing oscillatory, linear, and exponential lattice deformations. Although projected lattice areas can change, the deformation comes from a determinant-one complex symplectic flow on the full four-dimensional phase space. For a Gaussian regularization of the origin seed, we derive the associated complex width matrix and identify the positivity conditions that preserve Gaussian form. For an initial two-mode qunaught product state, the lossless limit recovers the standard beam-splitter generation of a square GKP$+$ Bell pair, while the no-jump dynamics produces its non-Hermitian deformation with a postselection cost set by the no-jump probability.

15.
Nature (Science) 2026-06-08

GPR15-guided CD8<sup>+</sup> T regulatory cells control intestinal inflammation

作者:

Inflammatory bowel disease (IBD) causes chronic suffering from gastrointestinal inflammation and dysfunction that can progress to colon cancer1,2. The disease prevalence is increasing and there is an urgent need to better understand its pathogenic mechanisms to improve treatment. We show that GPR15, a G protein-coupled receptor (GPCR) expressed in immune cells and previously described as an entry co-factor for human and simian immunodeficiency viruses3, is a marker and homing receptor for a subset of intramucosal GPR15-guided regulatory CD8+ T lymphocytes (CD8+ TIGR). Deleterious GPR15 gene variants in humans cause defective homing of CD8+ TIGR and are associated with severe early-onset IBD. Moreover, CD8+ TIGR cells are reduced in the intestinal mucosa of sporadic IBD patients. In mice, GPR15 deficiency impairs colonic homing of CD8+ TIGR cells, leading to accumulation of inflammatory macrophages and increased susceptibility to colitis. CD8+ TIGR cells potently kill macrophages activated by intestinal damage or disease using Fas ligand (FasL) and TNF-related weak inducer of apoptosis (TWEAK). The identification of CD8+ TIGR cells yields new insights into organ-specific immune regulation and potential therapeutics for IBD.

16.
arXiv (CS.CV) 2026-06-15

GarmentSketch: Large-scale Sketch-to-Fashion Benchmark

Fashion sketching is a cornerstone of design workflows, allowing rapid visualization of creative concepts prior to physical prototyping. Yet, progress in sketch-based fashion image synthesis has been hindered by the absence of large-scale, high-quality paired resources. To bridge this gap, we present GarmentSketch, a novel dataset comprising 26,249 fashion sketches across 21 garment categories, each paired with detailed textual descriptions. Captions were produced through a multi-stage pipeline that integrates multiple multimodal large language models (MLLMs) with human-in-the-loop refinement, ensuring both semantic accuracy and descriptive richness. We benchmark GarmentSketch on state-of-the-art generative models, providing baseline performance for sketch-guided text-to-image generation. Our experiments reveal both the promise and the current limitations of existing methods. By offering a comprehensive and richly annotated resource, GarmentSketch establishes a foundation for advancing sketch understanding, fine-grained fashion image generation, and creative human-AI collaboration in design. The dataset will be available at: https://khangbdd.github.io/garmentsketch.

18.
arXiv (CS.CL) 2026-06-16

Bridging Passive and Active: Enhancing Conversation Starter Recommendation via Active Expression Modeling

Large Language Model (LLM)-driven conversational search is shifting information retrieval from reactive keyword matching to proactive, open-ended dialogues. In this context, Conversation Starters are widely deployed to provide personalized query recommendations that help users initiate dialogues. Conventionally, recommending these starters relies on a closed "exposure-click" loop. Yet, this feedback loop mechanism traps the system in an echo chamber where, compounded by data sparsity, it fails to capture the dynamic nature of conversational search intents shaped by the open world. As a result, the system skews towards popular but generic suggestions. In this work, we uncover an untapped paradigm shift to shatter this harmful feedback loop: harnessing user "free will" through active user expressions. Unlike traditional recommendations, conversational search empowers users to bypass menus entirely through manually typed queries. The open-world intents in active queries hold the key to breaking this loop. However, incorporating them is non-trivial: (1) there exists an inherent distribution shift between active queries and formulated starters. (2) Furthermore, the "non-ID-able" nature of open text renders traditional item-based popularity statistics ineffective for large-scale industrial streaming training. To this end, we propose Passive-Active Bridge (PA-Bridge), a novel framework that employs an adversarial distribution aligner to bridge the distributional gap between passively recommended starters and active expressions. Moreover, we introduce a semantic discretizer to enable the deployment of popularity debiasing algorithms. Online A/B tests on our platform, demonstrate that PA-Bridge significantly boosts the Feature Penetration Rate by 0.54% and User Active Days by 0.04%.

19.
arXiv (quant-ph) 2026-06-17

Entanglement dynamics for atoms near a reflecting boundary: Enhancement and suppression by environment-induced interactions

arXiv:2602.23773v2 Announce Type: replace Abstract: We investigate how environment-induced interactions influence the entanglement dynamics of two atoms held at fixed positions near a perfectly reflecting boundary. Within the framework of open quantum systems, we explicitly incorporate the environment-induced energy shifts, including both atom-boundary contributions and an environment-induced atom-atom interaction, which are often neglected in previous studies. We show that, for any initial two-atom state, these energy-shift effects qualitatively and quantitatively modify the entanglement dynamics relative to treatments that omit them. Depending on the geometry and parameter regime, the environment-induced interactions can either enhance entanglement generation – yielding a larger maximum concurrence and a longer entanglement lifetime – or suppress it, reducing both the peak concurrence and the survival time. This behavior contrasts sharply with the free-space case, where the environment-induced atom-atom interaction affects entanglement generation only for a restricted class of initial states and does so in an exclusively assisting manner.

20.
arXiv (CS.LG) 2026-06-18

Neural Network Implementation of the Renormalization Group for Fault Diagnosis with Class Imbalance

arXiv:2606.18326v1 Announce Type: new Abstract: The application of machine learning models in practical tasks faces challenges such as class imbalance and multidimensional noise. This paper proposes RGNet, a neural network architecture based on the concept of the renormalization group (RG), for hierarchical coarse-graining of the feature space. The model sequentially compresses the input dimensionality and concatenates all scales before classification, allowing it to capture both local details and global patterns. The notion of RG-flows is introduced - interpretable low-dimensional representations whose visualization via t-SNE reveals a discrete curvilinear structure confirming the effectiveness of coarse-graining. Experimental results are presented on the imbalanced AI4I dataset. The obtained results demonstrate that RGNet is a universal, interpretable, and competitive solution for fault prediction in applications with imbalanced classes.

21.
arXiv (CS.AI) 2026-06-18

Where Did the Variability Go? From Vibe Coding to Product Lines by Regeneration

arXiv:2606.19042v1 Announce Type: cross Abstract: In vibe coding, an emerging AI-driven paradigm, an LLM generates an entire program from a natural language prompt, but what happens to the variability that traditional software engineering carefully builds into code? To answer this question, we conducted an exploratory analysis on 10 vibe coded C/C++ projects, which suggests that there is near-zero in-artifact variability, i.e., at compile and runtime. All variability decisions are resolved at a single new binding time, generation time, the moment the LLM produces the source code. Rather than treating this as a defect to fix, we propose Variability by Regeneration (VbR), to our knowledge the first product-line approach in which the LLM acts as the derivation engine, generating a purpose-built, free of dead code binary for each variant from a declarative specification, while a variant dispatcher transparently routes user requests to the matching binary. We formalise VbR, contrast it with classical SPL derivation, and demonstrate its full pipeline on a wc product family. For SPL engineering, variability in AI-generated software belongs in the specification, not in the code.

22.
arXiv (CS.AI) 2026-06-12

ARMOR-MAD: Adaptive Routing for Heterogeneous Multi-Agent Debate in Large Language Model Reasoning

arXiv:2606.13197v1 Announce Type: new Abstract: Multi-agent debate (MAD) can improve large language model reasoning, but fixed debate pipelines often waste computation and can amplify correlated errors among similar agents. We propose ARMOR-MAD, a training-free heterogeneous MAD framework that treats debate as conditional computation. ARMOR-MAD combines three components: Pre-debate Agreement Routing (PAR) decides whether independently generated Round-0 answers require debate; Early Agreement Stopping Evaluator (EASE) stops debate after convergence; and Semantic Outlier Detection (SOD) down-weights abnormal final answers during aggregation. Across MATH Level 5, GSM8K, MMLU, and MMLU-Pro, ARMOR-MAD consistently improves over fixed-round heterogeneous debate with the same model pool, reaching 65.5\%, 96.5\%, 90.0\%, and 81.5\% accuracy, respectively. The results suggest that genuine model heterogeneity and agreement-based control are both important for making MAD more accurate and efficient.

23.
arXiv (CS.AI) 2026-06-18

Rescaling MLM-Head for Neural Sparse Retrieval

arXiv:2606.18811v1 Announce Type: cross Abstract: Learned sparse retrieval (LSR) models such as SPLADE have traditionally used BERT-style masked language models as backbone encoders. A natural expectation is that replacing BERT with stronger pretrained encoders should improve retrieval effectiveness. However, we find that under standard SPLADE training recipes, backbones with large MLM-head L2 norms can suffer performance degradation and even training collapse under standard SPLADE training recipes. We identify this failure as a scale mismatch in the MLM head: SPLADE directly uses MLM-head outputs to construct sparse lexical representations, and query-document relevance is computed by an unnormalized dot product over these representations. As a result, an inflated MLM-head scale can amplify sparse activations, distort matching scores, and destabilize contrastive training under common training settings. To address this issue, we introduce a simple initialization-time correction that rescales the MLM-head projection by a constant factor before SPLADE training. This zero-cost adjustment improves training stability without modifying the model architecture or training objective. Across both in-domain and out-of-domain retrieval benchmarks, this simple correction substantially improves large-norm backbones such as ModernBERT and Ettin, turning unstable training runs into competitive sparse retrievers. In several settings, the corrected models further match or surpass the classic BERT-SPLADE baseline. These findings suggest that the bottleneck in adapting pretrained encoders to LSR is not encoder capacity alone, but the calibration of the MLM-head scale used to construct sparse lexical representations.

24.
arXiv (CS.LG) 2026-06-16

Inference-Time Decision Calibration for Temporal Classification

arXiv:2606.16034v1 Announce Type: new Abstract: Temporal classification errors are often treated as representation failures, but they can also arise from how available evidence is converted into decisions. This paper proposes a representation–calibration decomposition for temporal classification. We keep a trained native classifier frozen and separate two inference-time interventions: a conservative residual multi-scale branch that adds auxiliary logits to the native prediction, and a post-hoc branch-aware calibrator that recombines native and residual evidence at decision time. This design distinguishes missing temporal evidence from underused decision-level evidence without retraining the backbone. Across FI-2010, PTB-XL, UCI-HAR, MHEALTH, and HARTH, we find that gains are strongly regime-dependent. Residual multi-scale evidence is most useful in noisy or representation-limited settings, especially short-horizon FI-2010 and weaker recurrent backbones, while branch-aware calibration helps when native and auxiliary logits contain complementary evidence not fully exploited by the raw decision rule. Near-saturated settings show limited gains from either intervention. These results suggest that temporal classification should be understood not only as representation learning, but also as the problem of trusting, combining, and calibrating evidence from multiple views.

25.
arXiv (CS.AI) 2026-06-18

Learning from Own Solutions: Self-Conditioned Credit Assignment for Reinforcement Learning with Verifiable Rewards

arXiv:2606.18810v1 Announce Type: cross Abstract: Reinforcement learning with verifiable rewards (RLVR) has driven substantial progress in training LLMs for reasoning tasks, but representative methods such as GRPO assign uniform credit across all tokens, wasting gradient on routine tokens while under-crediting pivotal reasoning steps. Existing token-level credit assignment methods require resources beyond the model's own rollouts. GRPO variants rely on process reward models or ground-truth answers. Knowledge distillation assigns credit through per-token divergence but requires external teachers (On-Policy Distillation) or privileged information (On-Policy Self Distillation). However, these dependencies limit applicability in the pure RLVR setting. We observe that conditioning the model on its own verified trajectories induces a measurable per-token KL divergence between the original and conditioned distributions, and prove that distilling from a self-teacher constructed by verified trajectories leads to infeasible weighted-average solutions when multiple verified trajectories exist. We propose SC-GRPO (Self-Conditioned GRPO), which uses KL divergence mentioned before as a multiplicative weight on GRPO gradients. Across five benchmarks spanning math, code, and agentic tasks, SC-GRPO consistently outperforms 8.1% over GRPO and 5.9% over DAPO with stronger OOD performance. Moreover, SC-GRPO achieves higher performance than OPD.