Arson attacks at Ebola hospitals are a cry for regional development
Letter to the Editor
Academic Intelligence · Curated Daily
AcademicHub aggregates real-time literature from top journals and preprint platforms. Build your personal research radar and let large language models compile cross-disciplinary analysis briefings automatically.
Letter to the Editor
arXiv:2602.04155v2 Announce Type: replace-cross Abstract: When deploying a single predictor across multiple subpopulations, we propose a fundamentally different approach: interpreting group fairness as a bargaining problem among subpopulations. This game-theoretic perspective reveals that existing robust optimization methods such as minimizing worst-group loss or regret correspond to classical bargaining solutions and embody different fairness principles. We propose relative improvement, the ratio of actual risk reduction to potential reduction from a baseline predictor, which recovers the Kalai-Smorodinsky solution. Unlike absolute-scale methods that may not be comparable when groups have different potential predictability, relative improvement provides axiomatic justification including scale invariance and individual monotonicity. We establish finite-sample convergence guarantees under mild conditions.
arXiv:2606.20387v1 Announce Type: new Abstract: We investigate how interaction geometry shapes the low-energy phases of sparse tunable long-range quantum models. We focus on a class of graphs whose degree grows logarithmically with system size, and show how symmetry and frustration in graph connectivity can drive, suppress, and reshape ground-state phase transitions. The central examples are power-of-$p$ graphs, where even and odd values of $p$ exhibit qualitatively distinct behaviour: even-$p$ graphs inherit the rich phase structure of the power-of-two model, while odd-$p$ graphs are governed by geometric frustration. Fibonacci graphs provide a contrasting case, lacking the discrete self-similarity of the power-of-$p$ family but exhibiting a direct geometric mapping between the short- and long-range limits. Across our models, we find that phase structure and criticality are governed by the same effective-geometry principle, unifying our framework for experimentally motivated long-range quantum systems.
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
Research on bias in large language models (LLMs) has predominantly focused on third-person audits, which study how models represent or evaluate demographic groups as external subjects. However, this paradigm overlooks a structural blind spot because the user is absent from the audit. In practice, LLMs are used in open-ended, personal interactions, during which the model implicitly represents the user and adjusts its responses accordingly. When identical requests yield different responses depending on who is asking, bias manifests not in how the model describes others but in how it treats its interlocutor. We propose Situated Interaction Auditing (SIA), a user-centered framework for studying how user profile signals – implicit sociodemographic markers, writing style, and stated identity – systematically shape LLM response quality, content, and tone. We demonstrate the framework through a case study that intersects gender and socioeconomic status signals across multiple task domains and outline a research agenda for SIA as a new mission for natural language processing.
Optical coherence tomography (OCT) is essential in ophthalmology, but inconsistent image quality especially in low-cost devices hinders automated analysis. To address this, we introduce a flow-matching-based test-time adaptation method that generates high-quality surrogate images from noisy inputs. Typically, domain gaps between test and training data cause pixel distribution mismatches during the denoising process. We overcome this by matching the test image's histogram to synthetic reference trajectories, successfully aligning the input with expected distributions. Additionally, we remove the network's time conditioning to account for slight deviations in real-world noise distributions. Our approach achieves state-of-the-art performance in segmenting critical biomarkers for two stages of Age-related Macular Degeneration (AMD). Code is available: https://github.com/Veit21/tta-flow.
Background and Aims Genetic testing in dilated cardiomyopathy (DCM) guides risk stratification and family screening. Likely pathogenic or pathogenic (LP/P) variants are identified in approximately one-third of patients, leaving many without a genetic diagnosis. Cohort studies suggest that "gene-elusive" patients have a lower risk of adverse events. This study aims to better characterise this group and identify factors associated with adverse outcomes. Methods Consecutive and unrelated DCM patients undergoing genetic testing and returning no LP/P variants were retrospectively recruited and compared to two control cohorts of DCM patients carrying LP/P variants in LMNA and TTN for a primary composite endpoint of end-stage heart failure (ESHF) or malignant ventricular arrhythmia (MVA). Results Among patients without prior MVA, the composite endpoint occurred in 36/423 (8.5%) gene-elusive, 14/39 (35.9%) LMNA and 11/100 (11%) TTN cardiomyopathy patients (log-rank p
arXiv:2605.05167v2 Announce Type: replace Abstract: Absolutely maximally entangled (AME) states are fundamental resources in quantum information theory, yet their construction and certification remain a nontrivial problem. Within the family of quadratic phase quantum states, defined by symmetric matrices $P$ over finite fields $\mathbb{F}_{p^m}$, we show that the Rank-Purity Duality $\operatorname{Tr}(\rho_S^2) = |\mathbb{F}|^{-\operatorname{rk}_{\mathbb{F}}(P_{S,\bar{S}})}$ follows from additive character orthogonality and holds over all $\mathbb{F}_{p^m}$, yielding a polynomial-time AME certification criterion. For square-free dimensions $d = p_1\cdots p_r$, the Chinese Remainder Theorem induces a prime-field factorisation. This implies additivity of Rényi-2 entropy and yields sharp obstruction criteria that rule out cases such as $\operatorname{AME}(4,6)$ and constrain the open case $\operatorname{AME}(8,6)$. As a proof of concept, we construct an explicit $\operatorname{AME}(17,10001)$ state, certified across all $65{,}535$ bipartitions, demonstrating that the framework scales to large systems and previously inaccessible local dimensions.
arXiv:2603.09555v2 Announce Type: replace-cross Abstract: High-throughput Mamba-2 inference is usually tied to fused CUDA and Triton kernels, limiting portability across accelerator backends. We show that the state space duality (SSD) recurrence has a compiler-friendly structure: diagonal per-head dynamics, fixed-size chunking, einsum-dominated compute, and static control flow. Expressing this structure in standard JAX primitives gives a single-source inference path with no custom kernels, a registered JAX PyTree cache, and a compiled on-device autoregressive loop. On a single Google Cloud TPU v6e, batch-1 prefill reaches approximately 140 TFLOPS, or 15% model FLOP utilisation (MFU), the roofline ceiling for this regime, and cached decode reaches up to 64% hardware bandwidth utilisation (HBU). At a 4096-token context, cached decode is 27x–36x faster than full-prefix recomputation across five Mamba-2 checkpoints from 130M to 2.7B parameters. The same source runs unmodified on NVIDIA L40S, where cached decode remains sequence-length independent across all model scales. WikiText-103 validation perplexity matches the Triton reference mamba_ssm v2.2.2 within +/-0.0005 points, and hidden states agree to float32 rounding tolerance. Code is available at https://github.com/CosmoNaught/mamba2-jax.
Chain-of-thought (CoT) improves large language models' performance in math and symbolic reasoning. But on planning, contested ethics, and tasks where the model cannot check itself, more reasoning makes things worse. Both effects are documented; what has been missing is a principled account of which property decides the outcome. We argue it is meta-uncertainty: how unsure the model is about the reliability of its own evidence. When that uncertainty is high, extra reasoning stops adding signal and starts manufacturing false confidence. We prove that the policy minimizing expected free energy under uncertain precision stops integrating cues after a finite number of high-validity ones when the precision prior is heavy-tailed (Theorem 2.6.1), and under a Descending Dominance condition, is sample-wise identical to take-the-best (Theorem 2.7.4). Fast-and-frugal heuristics and active inference are, then, two descriptions of the same computation. The prediction is that on high-meta-uncertainty items, longer CoT should degrade accuracy. We score the regime per item (simulate-and-recover rho > 0.96), build FEH-79, a benchmark of Knightian frames with matched controls, and run a pre-registered study across seven models (five open-weight 3B-32B, two frontier), five CoT lengths, and 7,875 responses. The gate, fixed before any data, required a negative interaction with posterior probability above 0.95 and an accuracy drop of more than 6 points. It held. The high-regime drop is 17.3 points (95% CI [7.7, 25.5]); matched items with definite answers show no cost. The effect is regime-dependent: decisive in capable mid-to-large models, directional in the two frontier systems, absent-to-reversed in the weakest. The framework answers when CoT helps and unifies the Bayesian and fast-and-frugal traditions: less-is-more effects are evidence about the meta-uncertainty regime, not against Bayesian cognition.
arXiv:2606.17931v1 Announce Type: new Abstract: In recent years, electronic (E) commerce services have rapidly increased in the daily lives of people, which helpsthem to purchase products online. However, retail platforms have struggled to understand customer behavior and make it difficult to predict their future purchases. To overcome these challenges, this study proposes a hybrid Retail Deep NeuralNetwork (Ret-DNN) with an Extreme Gradient Boosting(XGBoost) model for capturing temporal features and tabular dynamics of retail data. First, data were sourced from a UnitedKingdom (UK)-based online retailer that contains transactions with almost 500,000 records. Then, the collected data were pre-processed using a series of techniques, such as data cleaning, outlier handling, temporal feature extraction, feature encoding, and z-score normalization, to ensure that the data were ready for model training and testing. Subsequently, the preprocessed data were fed into the Ret-DNN model, which acts as a feature extractor to understand the complete context of customer transactions. Further, the extracted data were fed as input into the XGBoost model, which predicted the final output as the purchase probability of customers. Finally, the proposed Ret-DNN XGBoost model achieved better results by attaining aMean Absolute Error (MAE) 0.2193 when compared to the existing Ret-DNN model. Keywords: Customer behavior forecasting, extreme gradientboosting, electronic commerce, predictive analytic, retail deepneural networks.
As Large Language Model (LLM) APIs become ubiquitous, users increasingly rely on black-box fingerprinting to verify that providers are serving the advertised premium models. However, these methods may overlook adversarial providers who manipulate model weights to cheat the fingerprint process. We introduce a novel threat termed fingerprint spoofing, where a malicious provider stealthily serves a weaker model that has been parameter-efficiently fine-tuned to mimic a stronger model, thereby evading user-side fingerprinting. We first formally prove that user-side resource constraints (i.e., finite query budgets and weak fingerprinting classifiers) make current fingerprinting vulnerable to fingerprint spoofing. Guided by this theoretical analysis, we propose GhostPrint, a cost-effective attack framework leveraging surrogate modeling, reward-ranked fine-tuning, and knowledge distillation. Extensive evaluations in both static and continual fingerprinting settings demonstrate that GhostPrint allows weak models to consistently bypass representative fingerprint methods while maintaining utility at a low fine-tuning cost, exposing a critical vulnerability in current LLM fingerprinting pipelines.
Abstract Background Evaluating antenatal care (ANC) quality is critical to reducing maternal and neonatal mortality. In Zambia, despite high basic ANC attendance, comprehensive national evidence on the clinical content and quality of services remains limited. This study assessed the coverage of WHO-recommended ANC interventions and identified factors associated with care quality using the latest national data. Methods A cross-sectional analysis was conducted using data from the 2024 Zambia Demographic and Health Survey. The final analytic sample comprised 4,829 women aged 15-49 with a live birth in the preceding 5 years. A composite index of 15 selected, equally weighted WHO-recommended components evaluated clinical assessment, counseling/screening, preventive interventions, and utilization. Survey-weighted Poisson regression estimated adjusted incidence rate ratios (aIRRs) for the count of ANC components received. Results The mean ANC quality score was 12.5 out of 15 (95% CI: 12.4-12.6), and 78.5% (95% CI: 77.0-80.0) of women achieved adequate ANC ([≥] 12/15 components). While individual clinical and counseling coverage generally exceeded 90%, only 47.2% (95% CI: 45.3-49.0) of women initiated care during the first trimester, and just 4.8% (95% CI: 4.1-5.6) achieved [≥] 8 ANC contacts. Maternal education was the strongest and most stable predictor of quality across all models. Compared to no education, higher education was associated with an 8.0% higher expected quality score (aIRR = 1.080, 95% CI: 1.051-1.110). Lower ANC quality was significantly associated with unwanted pregnancies (aIRR = 0.970, 95% CI: 0.956-0.993) and with residence in Western (aIRR = 0.923, 95% CI: 0.897-0.951) and North Western (aIRR = 0.966, 95% CI: 0.937-0.996) provinces. Absence of distance barriers and residence in Eastern, Luapula, and Copperbelt provinces were associated with higher quality scores. Conclusion While average ANC component coverage in Zambia is high, critical gaps persist in early initiation and total contact frequency. Care adequacy is strongly influenced by maternal education, relationship status, pregnancy intention, and regional inequities. These findings underscore the need for interventions targeted at uneducated women, preventing unintended pregnancies, and underserved regions such as Western and North Western Provinces. Keywords: Antenatal care quality, ANC content, Zambia, maternal education.
arXiv:2606.13608v1 Announce Type: new Abstract: Agent systems are advancing quickly across domains, but their evaluation remains fragmented. Most benchmarks rely on fixed, LLM-centric harnesses that require heavy integration, create test-production mismatch, and limit fair comparison across diverse agent designs. The root problem is the lack of an open, agent-agnostic assessment interface. We advocate Agentified Agent Assessment (AAA), where evaluation is performed by judge agents and all participants interact through standardized protocols: A2A for task management and MCP for tool access. Conventional benchmarking defines two separate interfaces, one for the benchmark and one for the agent, while AAA only needs one; this yields a generic, unified framework that separates assessment logic from agent implementation and enables reproducible, interoperable, and multi-agent evaluation. We further introduce AgentBeats as a concrete realization of AAA: we identify five practical operation modes that make standardized assessment compatible with real-world constraints on openness, privacy, and reproducibility. To evaluate our design at scale, we conduct two studies: a five-month open competition that drew 298 judge agents across 12 categories together with 467 subject agents from independent participants, showing that AAA applies across a heterogeneous range of benchmarks; and a case study on coding agents that confirms agentified evaluation preserves fidelity with the public record while surfacing previously missing head-to-head results, yielding research insights about agent design. Combining a community-scale field study and a controlled coding case study, we verify that AAA delivers coverage, practicality, and fidelity across heterogeneous scenarios at scale. Together, AAA and AgentBeats offer a clear path toward open, standardized, and reproducible agent assessment.
by Md Istiaq Ansari, Khandakar Tanvir Ahmed, Debby D. Wang, Kirill Medvedev, Wei Zhang Protein isoforms produced from the same gene through post-transcriptional regulatory mechanisms, such as alternative splicing, can substantially alter protein structure and function, including drug-binding properties. However, most existing drug-target interaction (DTI) and drug-target affinity (DTA) prediction models rely exclusively on a single representative protein sequence per gene, typically the canonical or longest isoform, thereby overlooking the functional diversity introduced by alternative isoforms. This assumption can introduce bias, limit generalizability, and compromise the biological validity of model predictions. In this study, we systematically investigate the impact of protein isoform variation on DTI prediction accuracy. Our results show that substituting the canonical sequence with an alternative isoform often leads to substantial declines in predictive performance. Structural and binding affinity analyses further reveal that these discrepancies are frequently associated with changes in predicted binding-site configurations, which we further examine through controlled perturbations of binding-site residues. These experiments suggest that even subtle alterations in binding regions can lead to inconsistent DTI predictions. Overall, our findings uncover a critical limitation in current DTI modeling frameworks and underscore the importance of incorporating isoform-specific information to better reflect biological reality and improve therapeutic relevance. The codes and datasets are available at https://github.com/compbiolabucf/DTIVariant.
Emotional intelligence (EI) in Large Language Models (LLMs) is often evaluated through static understanding tasks or single-response dialogue generation. However, emotion management is interactive: a good model should not only recognize a user's emotion, but also improve the user's emotional and relational state over several turns. We introduce EIBench, a simulator-based benchmark for interactive emotion management. EIBench contains 2,222 scenarios, with 2,009 for training and 213 for held-out testing. The scenarios are organized by a 2x2 taxonomy covering Support, Defense, Repair, and Charm, which together capture different forms of support, boundary maintenance, trust repair, and rapport building. In each scenario, an LLM simulator plays the user, updates an emotion-relation state after each turn, and maps the final state to an anchor-based score. This design makes EIBench both an evaluation benchmark and a training environment: the final state gives the outcome reward, while the per-turn state updates provide dense feedback for RL. We evaluate 15 open- and closed-source LLMs. Current models perform well on support and rapport-building scenes, but struggle with boundary maintenance under user pressure. To improve the EI ability of LLMs, we propose Centered Turn-Credit GRPO (CTC-GRPO), a GRPO extension that reuses the simulator's per-turn state updates as dense turn-level feedback while preserving the final outcome reward. CTC-GRPO improves Qwen3-8B from -22.4 to +22.4 on EIBench and also improves on out-of-distribution evaluations including SAGE (+12.4) and EQBench3 (+20.9%). Our results show that simulator-tracked user states can support both evaluation and training for multi-turn emotion management.
Markerless, single-RGB-D-camera motion capture provides a low-cost and non-invasive alternative to conventional marker-based systems for robot teleoperation; however, depth estimation often degrades in the presence of self-occlusion, particularly during upper-limb motion. This paper presents an Arm Kinematic Correction (AKC) method that improves depth estimation by enforcing geometric constraints based on constant arm lengths. The proposed approach reconstructs occluded joint depths by leveraging wrist positions and predefined arm lengths via a deterministic formulation based on the Pythagorean theorem, thereby avoiding the need for complex probabilistic modeling or parameter tuning. Experimental validation against a Vicon reference system demonstrates reliable performance for both static and dynamic joint motions, evaluated using root-mean-square error (RMSE) and Pearson correlation. Furthermore, motion-mapping teleoperation is successfully demonstrated in both simulated and physical robot environments. The results show that AKC enhances robustness and preserves anatomical consistency under long-duration, severe self-occlusion, even when paired with less reliable temporal filters, highlighting its practicality for real-time applications such as robot teleoperation and human-robot interaction.
Text-to-image (T2I) systems increasingly rely on Large Language Model (LLM)-based text conditioning to interpret and expand user prompts. While this improves prompt understanding and text-image alignment, we find that it can also introduce implicit demographic assumptions, even when demographic attributes are unspecified. To systematically investigate this behavior across varying levels of prompt ambiguity and complexity, we construct a comprehensive benchmark covering diverse prompt settings. Evaluations on eight recent T2I models show that LLM-based systems consistently exhibit stronger demographic skew than non-LLM-based baselines. We further analyze system prompts, a component unique to LLM-based T2I systems that guides prompt interpretation and expansion. Our analyses show that these instructions strongly influence text embeddings, which subsequently leads to biased image generations. Motivated by these findings, we propose FairPro, a training-free debiasing framework that adaptively generates fairness-aware instructions while preserving user intent. Experiments demonstrate that FairPro substantially reduces demographic disparities while maintaining prompt fidelity.
arXiv:2601.05746v2 Announce Type: replace Abstract: Recent years have witnessed the rapid development of Large Language Model-based Multi-Agent Systems (MAS), which excel at collaborative decision-making and complex problem-solving. Researchers have further investigated Multi-Agent Debate (MAD) frameworks, which enhance the reasoning and collaboration capabilities of MAS through information exchange and debate among multiple agents. However, existing approaches often rely on unguided initialization, causing agents to adopt identical reasoning paths that lead to the same errors. As a result, effective debate among agents is hindered, and the final outcome frequently degenerates into simple majority voting. To solve the above problem, we introduce Dynamic Multi-Agent Debate (DynaDebate), which enhances the effectiveness of multi-agent debate through three key mechanisms: (1) Dynamic Path Generation and Allocation, which employs a dedicated Path Generation Agent to generate diverse and logical solution paths with adaptive redundancy; (2) Process-Centric Debate, which shifts the focus from surface-level outcome voting to rigorous step-by-step logic critique to ensure process correctness; (3) A Trigger-Based Verification Agent, which is activated upon disagreement and uses external tools to objectively resolve deadlocks. Experiments show that DynaDebate achieves superior or highly competitive performance across the majority of benchmarks\footnote{The code is at https://github.com/nwpuLee2021/brianstorm.}.
arXiv:2603.21283v3 Announce Type: replace Abstract: For a traveling salesperson problem (TSP) of n cities, we present a compact quantum encoding based on a time-register representation of tours. A candidate route is represented as a sequence of n-1 city labels over discrete time steps, with one fixed start city and the remaining cities encoded in binary registers. We describe three ingredients of the construction: uniform route generation over the route register, a reversible validity oracle, and a phase oracle that encodes the total tour cost. The validity oracle checks both that the non-start city labels form a permutation and, for incomplete graphs, that every directed edge used by the route exists. The cost oracle then accumulates the start-edge, intermediate-transition, and return-edge costs into a tour-dependent phase for valid routes. This yields a coherent superposition of candidate routes with feasibility and tour-length information embedded directly in the quantum state. The complete construction uses O(n log n) qubits, while a naive implementation has worst-case elementary-gate complexity O(n^3 log n). The encoding is compatible with amplitude amplification or spectral filtering techniques such as the quantum singular value transform (QSVT) or Grover's algorithm. However, due to the exponentially small fraction of valid tours, the overall complexity remains exponential even when combined with amplitude amplification.
We present MaxProof, a population-level test-time scaling framework for competition-level mathematical proof in the MiniMax-M3 series. M3 first trains three proof-oriented capabilities – proof generation, proof verification, and critique-conditioned proof repair – using a defense-in-depth generative verifier engineered for low false-positive rate. These capabilities are merged into a single released M3 model. At test time, MaxProof treats the model as a generator, verifier, refiner, and ranker, searches over a population of candidate proofs, and returns one final proof through tournament selection. With MaxProof test-time scaling, the M3 model reaches 35/42 on IMO 2025 and 36/42 on USAMO 2026, exceeding the human gold-medal threshold on both.
arXiv:2602.04037v3 Announce Type: replace Abstract: Learning domain adaptive policies that can generalize to unseen transition dynamics, remains a fundamental challenge in learning-based control. Substantial progress has been made through domain representation learning to capture domain-specific information, thus enabling domain-aware decision making. We analyze the process of learning domain representations through dynamical prediction and find that selecting contexts adjacent to the current step causes the learned representations to entangle static domain information with varying dynamical properties. Such mixture can confuse the conditioned policy, thereby constraining zero-shot adaptation. To tackle the challenge, we propose DADP (Domain Adaptive Diffusion Policy), which achieves robust adaptation through unsupervised disentanglement and domain-aware diffusion injection. First, we introduce Lagged Context Dynamical Prediction, a strategy that conditions future state estimation on a historical offset context; by increasing this temporal gap, we unsupervisedly disentangle static domain representations by filtering out transient properties. Second, we integrate the learned domain representations directly into the generative process by biasing the prior distribution and reformulating the diffusion target. Extensive experiments on challenging benchmarks across locomotion and manipulation demonstrate the superior performance, and the generalizability of DADP over prior methods. More visualization results are available on the https://outsider86.github.io/DomainAdaptiveDiffusionPolicy/.
arXiv:2605.28968v2 Announce Type: replace Abstract: We demonstrate atom–photon entanglement using a single cesium atom trapped in an optical tweezer. Entanglement is generated by resonant excitation and subsequent spontaneous decay, which entangles the atomic Zeeman state with photon polarization. The photon is collected with a high numerical aperture objective (NA = 0.55) and coupled into a single-mode fiber, enabling atom photon measurements and measurement of the Bell-state fidelity. We obtain raw entanglement fidelity of ${\mathcal F} = 0.942(16)$ and inferred fidelity of ${\mathcal F}_inf = 0.962(26)$ after correcting independently characterized atom measurement errors. Compared with related free-space experiments using $^{87}$Rb, the multilevel structure of the relevant excited state in $^{133}$Cs requires the use of a single short excitation pulse in each entanglement attempt in order to suppress unwanted re-excitation. These results establish a free-space Cs atom–photon interface and provide a step toward dual-species Rb–Cs quantum networking.
arXiv:2606.18307v1 Announce Type: cross Abstract: Optimizing the training data distribution for Supervised Fine-Tuning (SFT) dictates the capability of Large Language Models (LLMs). While existing data curation methods excel at accelerating training under constrained budgets, they are less suited to elevating the capability upper bound. The challenge here is no longer to identify a smaller subset that preserves performance, but to refine the data distribution toward instances most capable of improving the final model. To address this problem, we explore instance-level data attribution using Influence Functions (IF). We identify that standard IF formulations struggle in this setting due to two structural limitations: a proximity gap caused by off-policy validation targets, and a severe bias towards gradient norm. We propose DRIFT (Data Refinement via On-Policy Influence Functions for Supervised Fine-Tuning). Instead of relying on external reference data, DRIFT utilizes the model's on-policy rollouts as validation targets, which empirically minimizes the parameter proximity gap and better aligns with the local neighborhood assumption of IF. It further applies signed weighting based on trajectory correctness and debiases influence scores against the gradient hacking issue, allowing a small set of validation queries to act as reliable anchors for attributing the full dataset. Experiments on 7B-parameter instruction and reasoning models show that DRIFT consistently raises the performance ceiling on both, outperforming existing data curation baselines.
arXiv:2606.17006v1 Announce Type: cross Abstract: We introduce TuneJury, an open, instance-level pairwise reward model for text-to-music that predicts a music preference score from a text prompt and an audio clip. The released checkpoint is trained on publicly available human-preference labels covering arena-style (A vs. B) votes, metric-alignment preference pairs, crowdsourced pairwise comparisons, and expert aesthetic ratings. The predicted score margin between two clips is well calibrated on our held-out test split, supporting data filtering via a simple score threshold. TuneJury generalizes to both held-out test pairs and out-of-distribution benchmarks, remaining competitive with prior baselines on the latter. For generators released after training, we introduce anchor calibration, a post-hoc, per-system Bradley-Terry calibration that recovers agreement at substantially better data efficiency than from-scratch retraining. The same frozen reward drives consistent reward-axis gains across three downstream applications: inference-time best-of-N selection, DITTO-style latent optimization, and expert-iteration post-training. TuneJury is available at https://github.com/yonghyunk1m/TuneJury.