# paper-forge daily report — 2026-10-02

161 new papers screened; 10 analyzed in depth; 0 reproduction runs.

## Reproductions

- none today (reimplementation is opt-in: `forge daily --reimplement`)

## Top candidates

### [Exact information accounting for SGD methods](https://arxiv.org/abs/2610.00446)
`2610.00446` · score 61.6 · analysis: heuristic · feasibility 8/10

As an alternative to the standard geometric analyses, we give an exact, information-theoretic analysis of stochastic gradient descent (SGD) and its variants. We show that a preconditioned SGD step is the posterior-mean update of a Gaussian Bayes model, and that its one-step regret splits into an intrinsic-time cost and a change in comparator information.

Rubric evidence: + cheap/closed-form ("Convex"); + toy/synthetic experiments ("synthetic"); + small/synthetic data ("synthetic"); + concrete method ("closed-form"); + formal statement ("convergence"); no public code found; + keywords: optimizer

### [FERPO: Forward Entropy-Regularized Policy Optimization](https://arxiv.org/abs/2610.02198)
`2610.02198` · score 60.0 · analysis: heuristic · feasibility 6/10

Several state-of-the-art methods for online reinforcement learning in continuous control improve policies using action gradients of a learned critic. However, critics are typically trained to predict returns, and accurate value predictions do not necessarily yield accurate action derivatives, potentially leading to unreliable policy updates.

Rubric evidence: + public benchmarks ("benchmarks"); + explicit algorithm/pseudocode ("Algorithm 1"); + formal statement ("theorem"); + code: https://github.com/Atarilab/FERPO; + keywords: sampling, regularization

### [Counterfactual Generation via Flow Matching: Coupling-Sensitive End-to-End Rates](https://arxiv.org/abs/2610.01193)
`2610.01193` · score 58.4 · analysis: heuristic · feasibility 8/10

Counterfactual generation seeks to sample outcomes under a hypothetical intervention or decision using observational data collected under the factual assignment mechanism. We develop a flow-matching approach that combines a sample-split, doubly robust training objective with a learned coupling between observed source outcomes and target outcomes drawn from a fitted conditional outcome model.

Rubric evidence: + toy/synthetic experiments ("synthetic"); + small/synthetic data ("synthetic"); + public benchmarks ("benchmarks"); + explicit algorithm/pseudocode ("Algorithm 2"); + formal statement ("bound"); no public code found; + keywords: monte carlo

### [Hierarchical Continuous Diffusion Language Models](https://arxiv.org/abs/2610.02193)
`2610.02193` · score 56.8 · analysis: heuristic · feasibility 6/10

Discrete diffusion language models offer a compelling alternative to autoregressive generation for tasks demanding bidirectional reasoning and global constraint satisfaction. Yet they share a structural bottleneck: when decoding in parallel, each token is sampled independently from its marginal, severing the statistical dependencies among the tokens decoded together.

Claimed: perplexity 1.0 (abstract (regex extraction, unverified))
Rubric evidence: - diffusion/video ("Diffusion"); + concrete method ("closed-form"); + formal statement ("bound"); + claims simplicity ("simple"); + code: https://github.com/rhfeiyang/HC-DLM; + keywords: variational; + 50 HF upvotes

### [The Curvature of Regret in Contextual Linear Optimization](https://arxiv.org/abs/2610.01980)
`2610.01980` · score 54.7 · analysis: heuristic · feasibility 7/10

Decision-focused learning for linear optimization is complicated by the discontinuity of the optimizer, where small cost errors may leave the decision unchanged or move it to a different vertex. We show that this non-smooth pointwise behavior becomes locally quadratic after averaging over the data distribution, and we derive the curvature in closed form, specifically, a matrix-valued measure supported on the walls of the normal fan.

Rubric evidence: + cheap/closed-form ("closed form"); + explicit algorithm/pseudocode ("Algorithm 1"); + concrete method ("closed form"); + formal statement ("We prove"); + claims simplicity ("simple"); no public code found; + keywords: optimizer

### [Where-OPD: Spatially Guided On-Policy Self-Distillation of MLLMs with Synthetic Scenes](https://arxiv.org/abs/2610.02117)
`2610.02117` · score 53.2 · analysis: heuristic · feasibility 6/10

On-policy self-distillation has recently emerged as an effective approach for improving language-model reasoning by supervising students with a frozen or EMA version of themselves that receives privileged information. Its application to multimodal large language models (MLLMs), however, remains largely unexplored.

Rubric evidence: - LLM-based ("large language model"); + toy/synthetic experiments ("Synthetic"); + small/synthetic data ("Synthetic"); + public benchmarks ("benchmarks"); + claims simplicity ("simple"); + code: https://github.com/sirkosophia/Where-OPD

### [E-MoE: Enhanced Mixture-of-Experts for Non-Factorized Diffusion Language Models](https://arxiv.org/abs/2609.37533)
`2609.37533` · score 52.1 · analysis: heuristic · feasibility 6/10

Masked diffusion models (MDMs) generate sequences by progressively unmasking several tokens per denoising step, but their reverse process is typically factorized over positions, limiting sample quality in the few-step regime where diffusion's speed advantage over autoregressive decoding matters most. A recent line of work introduces a continuous Gaussian latent, trained as a variational autoencoder, to capture correlations across positions, but such approaches are prone to posterior collapse, where the latent is silently ignored.

Rubric evidence: - diffusion/video ("Diffusion"); + toy/synthetic experiments ("synthetic"); + small/synthetic data ("synthetic"); + public benchmarks ("benchmarks"); + formal statement ("bound"); no public code found; + keywords: variational; + 17 HF upvotes

### [Trust the Direction, Search the Step: Zero-and-First-Order Methods for LLM Fine-Tuning](https://arxiv.org/abs/2610.02190)
`2610.02190` · score 52.1 · analysis: heuristic · feasibility 5/10

Step-size selection remains a central challenge in large-scale neural network optimization; conservative steps slow convergence, while aggressive steps can destabilize it. We combine \textbf{Z}ero-and-\textbf{F}irst-\textbf{O}rder optimization~(ZFO) and propose a lightweight framework that decouples direction selection from step-size.

Rubric evidence: - LLM-based ("LLM"); + CPU-friendly ("lightweight"); + explicit algorithm/pseudocode ("Algorithm 1"); + formal statement ("convergence"); - system/framework (many moving parts) ("framework"); + code: https://github.com/nizswan/Zeroth-First-Order-Framework; + keywords: optimizer

### [When Do Intrinsic Rewards Lead to Exploration?](https://arxiv.org/abs/2610.02159)
`2610.02159` · score 49.5 · analysis: heuristic · feasibility 5/10

Intrinsic rewards are designed to guide exploration in reinforcement learning by assigning value to an agent's experience, for example through prediction error or learning progress. However, maximizing these rewards need not produce the most informative experience available.

Rubric evidence: + formal statement ("bound"); + claims simplicity ("simple"); + code: https://github.com/scottviteri/what-is-exploration

### [Exact Distinguishability in Non-Markovian Decision Processes](https://arxiv.org/abs/2610.01527)
`2610.01527` · score 46.3 · analysis: heuristic · feasibility 5/10

Non-Markovian environments are often modeled as Regular Decision Processes (RDPs), where dynamics depend on the interaction history through a finite automaton. Existing offline guarantees for RDPs rely on a distinguishability assumption on the behaviour policy but provide no means of verifying it.

Rubric evidence: + formal statement ("We prove"); + code: https://github.com/Kcbir/pec

## All screened papers

| score | paper | compute | data | clarity | code | interest | analysis | laptop-feasible |
|---:|---|---:|---:|---:|---:|---:|---|---|
| 61.6 | [Exact information accounting for SGD methods](https://arxiv.org/abs/2610.00446) `2610.00446` | 0.85 | 0.75 | 0.75 | 0.00 | 0.20 | heuristic | True |
| 60.0 | [FERPO: Forward Entropy-Regularized Policy Optimization](https://arxiv.org/abs/2610.02198) `2610.02198` | 0.50 | 0.45 | 0.85 | 1.00 | 0.40 | heuristic | True |
| 58.4 | [Counterfactual Generation via Flow Matching: Coupling-Sensitive End-to-End Rates](https://arxiv.org/abs/2610.01193) `2610.01193` | 0.65 | 0.80 | 0.85 | 0.00 | 0.20 | heuristic | True |
| 56.8 | [Hierarchical Continuous Diffusion Language Models](https://arxiv.org/abs/2610.02193) `2610.02193` | 0.35 | 0.40 | 0.90 | 1.00 | 0.50 | heuristic | False |
| 54.7 | [The Curvature of Regret in Contextual Linear Optimization](https://arxiv.org/abs/2610.01980) `2610.01980` | 0.70 | 0.40 | 1.00 | 0.00 | 0.20 | heuristic | True |
| 53.2 | [Where-OPD: Spatially Guided On-Policy Self-Distillation of MLLMs with Synthetic Scenes](https://arxiv.org/abs/2610.02117) `2610.02117` | 0.45 | 0.80 | 0.55 | 1.00 | 0.00 | heuristic | False |
| 52.1 | [E-MoE: Enhanced Mixture-of-Experts for Non-Factorized Diffusion Language Models](https://arxiv.org/abs/2609.37533) `2609.37533` | 0.50 | 0.80 | 0.55 | 0.00 | 0.50 | heuristic | True |
| 52.1 | [Trust the Direction, Search the Step: Zero-and-First-Order Methods for LLM Fine-Tuning](https://arxiv.org/abs/2610.02190) `2610.02190` | 0.45 | 0.40 | 0.75 | 1.00 | 0.20 | heuristic | False |
| 49.5 | [When Do Intrinsic Rewards Lead to Exploration?](https://arxiv.org/abs/2610.02159) `2610.02159` | 0.50 | 0.40 | 0.70 | 1.00 | 0.00 | heuristic | True |
| 46.3 | [Exact Distinguishability in Non-Markovian Decision Processes](https://arxiv.org/abs/2610.01527) `2610.01527` | 0.50 | 0.40 | 0.55 | 1.00 | 0.00 | heuristic | True |
| 45.8 | [CorrGRPO: Correlation-Normalized GRPO for Multi-Reward Learning](https://arxiv.org/abs/2609.36820) `2609.36820` | 0.50 | 0.40 | 0.30 | 1.00 | 0.30 | — | — |
| 45.3 | [Sharp Non-Asymptotic Analysis of the Penalized Challenger in $β$-EB-TCI for Bernoulli Bandits](https://arxiv.org/abs/2610.01951) `2610.01951` | 0.50 | 0.40 | 0.70 | 0.00 | 0.40 | — | — |
| 45.3 | [Muon meets Tamed Langevin: Momentum Preconditioning beyond Convex and gradient-Lipschitz Potentials](https://arxiv.org/abs/2610.02158) `2610.02158` | 0.70 | 0.40 | 0.55 | 0.00 | 0.20 | — | — |
| 44.7 | [No Model Required: Text Entropy Rate Filtering Mitigates Iterative Fine-Tuning Collapse](https://arxiv.org/abs/2610.01493) `2610.01493` | 0.65 | 0.75 | 0.40 | 0.00 | 0.00 | — | — |
| 44.7 | [JoinGR: Learning to Traverse Join Graphs for Table Retrieval](https://arxiv.org/abs/2610.01064) `2610.01064` | 0.65 | 0.45 | 0.40 | 0.00 | 0.40 | — | — |
| 44.7 | [Zero Flux: Flow-Based Comparison of High-Dimensional Discrete Distributions](https://arxiv.org/abs/2610.01472) `2610.01472` | 0.65 | 0.75 | 0.40 | 0.00 | 0.00 | — | — |
| 44.7 | [Heteroskedastic Canonical Polyadic Tensor Decomposition](https://arxiv.org/abs/2610.00498) `2610.00498` | 0.65 | 0.75 | 0.40 | 0.00 | 0.00 | — | — |
| 43.2 | [Generative modeling of intrinsically disordered protein regions by reinforcing sparse autoencoder features](https://arxiv.org/abs/2610.02189) `2610.02189` | 0.50 | 0.40 | 0.40 | 1.00 | 0.00 | — | — |
| 43.2 | [Inference for stochastic differential equations driven by weighted sub-fractional Brownian motion using neural networks and the Euler approximation](https://arxiv.org/abs/2610.00793) `2610.00793` | 0.50 | 0.75 | 0.40 | 0.00 | 0.20 | — | — |
| 42.6 | [Inferring Multi-Timescale Neural Dynamics with Switching Linear Dynamical Systems](https://arxiv.org/abs/2610.01786) `2610.01786` | 0.65 | 0.75 | 0.30 | 0.00 | 0.00 | — | — |
| 42.1 | [Cost-augmented Schrödinger bridges on graphs are exactly solvable: a Feynman-Kac tilt replaces learned control](https://arxiv.org/abs/2610.02195) `2610.02195` | 0.70 | 0.40 | 0.40 | 0.00 | 0.20 | — | — |
| 42.1 | [Lower Bounds for Stochastic First-Order Algorithms with Variance Reduction in Nonconvex--Concave Minimax Optimization](https://arxiv.org/abs/2610.01662) `2610.01662` | 0.70 | 0.40 | 0.55 | 0.00 | 0.00 | — | — |
| 42.1 | [IQS-BO: In-Context Query Selection for Bayesian Optimisation](https://arxiv.org/abs/2610.01269) `2610.01269` | 0.40 | 0.80 | 0.30 | 0.00 | 0.40 | — | — |
| 41.6 | [Make Sparse Rewards Count: Density-Aware Reward Aggregation for Multi-Reward RL](https://arxiv.org/abs/2610.00574) `2610.00574` | 0.30 | 0.40 | 0.40 | 1.00 | 0.30 | — | — |
| 40.5 | [Multifidelity Formulations for Triangular Transport](https://arxiv.org/abs/2610.00698) `2610.00698` | 0.65 | 0.40 | 0.40 | 0.00 | 0.20 | — | — |
| 40.5 | [How the Audit Rule Shapes Faithful Factor Explanations in LLMs](https://arxiv.org/abs/2610.01514) `2610.01514` | 0.45 | 0.80 | 0.45 | 0.00 | 0.00 | — | — |
| 40.0 | [Stochastic Rounding in Low-Precision Transformer Inference: A Variable-Precision Emulation Study of a Small GPT-2](https://arxiv.org/abs/2610.01889) `2610.01889` | 0.30 | 0.40 | 0.55 | 1.00 | 0.00 | — | — |
| 40.0 | [Pragmatic DML with AI-Learned Representations](https://arxiv.org/abs/2610.01935) `2610.01935` | 0.70 | 0.40 | 0.45 | 0.00 | 0.00 | — | — |
| 40.0 | [Block Optimism for Nonstationary Bandits with Latent Linear Dynamics](https://arxiv.org/abs/2610.00911) `2610.00911` | 0.50 | 0.45 | 0.55 | 0.00 | 0.20 | — | — |
| 40.0 | [Removing the NEEDLE in the Haystack: Backdoor Removal in LLMs via Weight Orthogonalisation](https://arxiv.org/abs/2610.00348) `2610.00348` | 0.30 | 0.40 | 0.40 | 1.00 | 0.20 | — | — |
| 39.5 | [Generalized Engression Models](https://arxiv.org/abs/2610.01823) `2610.01823` | 0.65 | 0.45 | 0.30 | 0.00 | 0.20 | — | — |
| 38.9 | [Finetuning with Sampling: SFT Learns Better Than You Think](https://arxiv.org/abs/2610.02140) `2610.02140` | 0.50 | 0.40 | 0.40 | 0.00 | 0.40 | — | — |
| 38.9 | [Are We Recovering Mechanisms? Objective-Level Recovery Gaps in Mechanistic Interpretability](https://arxiv.org/abs/2610.02098) `2610.02098` | 0.50 | 0.40 | 0.40 | 0.00 | 0.40 | — | — |
| 38.9 | [Bellman Meets Lyapunov: Unsupervised Reinforcement Learning via Mastering Chaos](https://arxiv.org/abs/2610.02012) `2610.02012` | 0.50 | 0.40 | 0.70 | 0.00 | 0.00 | — | — |
| 38.9 | [Mem++: Non-Destructive Memory for Long-Term Organizational LLM Agents](https://arxiv.org/abs/2610.02002) `2610.02002` | 0.30 | 0.45 | 0.30 | 1.00 | 0.20 | — | — |
| 38.9 | [On skew-symmetric distributions and their use in Monte Carlo sampling algorithms: coordinate-free, Gibbs-style and manifold versions of the Barker proposal](https://arxiv.org/abs/2610.01448) `2610.01448` | 0.50 | 0.40 | 0.40 | 0.00 | 0.40 | — | — |
| 38.9 | [Sharpening Tax in Post-Training](https://arxiv.org/abs/2610.01509) `2610.01509` | 0.05 | 0.45 | 0.45 | 1.00 | 0.50 | — | — |
| 38.4 | [Joint Branch-Space Transform Coding for Diffusion Activation Quantization with Classifier-Free Guidance](https://arxiv.org/abs/2610.00930) `2610.00930` | 0.55 | 0.40 | 0.60 | 0.00 | 0.00 | — | — |
| 38.4 | [A Missing Piece for Trustworthy AI Reviewers: From Benchmarking Rhetorical Robustness to SciCore Review](https://arxiv.org/abs/2609.39027) `2609.39027` | 0.50 | 0.45 | 0.40 | 0.00 | 0.30 | — | — |
| 38.4 | [AgSpec: Pushing the Limits of Retrieval-Based Speculative Decoding in Coding Agent Pipelines](https://arxiv.org/abs/2610.01108) `2610.01108` | 0.50 | 0.45 | 0.30 | 0.00 | 0.43 | — | — |
| 37.9 | [From Knowledge Access to Source Learning: Developing Source-Specific Competence](https://arxiv.org/abs/2610.02150) `2610.02150` | 0.30 | 0.45 | 0.40 | 1.00 | 0.00 | — | — |
| 37.9 | [Yo-ByT5: Efficient and High-Fidelity Diacritic Restoration for Yorùbá](https://arxiv.org/abs/2610.01634) `2610.01634` | 0.30 | 0.45 | 0.40 | 1.00 | 0.00 | — | — |
| 37.4 | [SoftServe: A Scalable Quasi-Newton Method for Deep Learning](https://arxiv.org/abs/2610.02182) `2610.02182` | 0.55 | 0.40 | 0.40 | 0.00 | 0.20 | — | — |
| 37.4 | [Sim+Real: Joint Simulation - Experiment Training Improves Balanced Prediction in Physical Systems](https://arxiv.org/abs/2610.01974) `2610.01974` | 0.65 | 0.40 | 0.40 | 0.00 | 0.00 | — | — |
| 37.4 | [Typological Alignment of Stack-Based Language Models on Mildly Context-Sensitive Artificial Languages](https://arxiv.org/abs/2610.02040) `2610.02040` | 0.65 | 0.40 | 0.40 | 0.00 | 0.00 | — | — |
| 37.4 | [Better Supervision Is Nearby: Neighborhood On-Policy Self-Distillation](https://arxiv.org/abs/2609.39687) `2609.39687` | 0.50 | 0.45 | 0.40 | 0.00 | 0.23 | — | — |
| 36.8 | [Detecting Inconsistencies in Model Specifications with LLM-as-Verifier Reasoning](https://arxiv.org/abs/2610.01847) `2610.01847` | 0.30 | 0.40 | 0.40 | 1.00 | 0.00 | — | — |
| 36.8 | [Compound interpretation is based on analogy](https://arxiv.org/abs/2610.01688) `2610.01688` | 0.50 | 0.40 | 0.60 | 0.00 | 0.00 | — | — |
| 36.8 | [GAW-PO: Preference Optimization with Gradient-Aligned Token Weights](https://arxiv.org/abs/2610.01511) `2610.01511` | 0.50 | 0.45 | 0.40 | 0.00 | 0.20 | — | — |
| 36.8 | [Grand Canonical Generators](https://arxiv.org/abs/2610.00683) `2610.00683` | 0.50 | 0.40 | 0.30 | 0.00 | 0.40 | — | — |
| 36.8 | [Fractional Laplace Neural Operators: Exact Architectures, an Expressivity Frontier at Criticality, and Certified Stability for Memory-Driven Network Dynamics](https://arxiv.org/abs/2610.00515) `2610.00515` | 0.50 | 0.45 | 0.55 | 0.00 | 0.00 | — | — |
| 36.3 | [When Does Correction Become Repair? Mechanistic Auditing of Internal Interventions in Tool-Using LLMs](https://arxiv.org/abs/2609.36138) `2609.36138` | 0.30 | 0.40 | 0.30 | 1.00 | 0.10 | — | — |
| 35.8 | [From Gradients to Capabilities: Understanding Multi-Teacher On-Policy Distillation](https://arxiv.org/abs/2610.02179) `2610.02179` | 0.50 | 0.40 | 0.40 | 0.00 | 0.20 | — | — |
| 35.8 | [AI Emulation of Stochastic Sudden Stratospheric Warming with Interpretable Latent Structure](https://arxiv.org/abs/2610.02069) `2610.02069` | 0.50 | 0.40 | 0.40 | 0.00 | 0.20 | — | — |
| 35.8 | [Sequential Capacity of Quantum Processes with Finite Memory](https://arxiv.org/abs/2610.02068) `2610.02068` | 0.50 | 0.40 | 0.55 | 0.00 | 0.00 | — | — |
| 35.8 | [Distributionally Robust Schrödinger Bridge](https://arxiv.org/abs/2610.02043) `2610.02043` | 0.50 | 0.40 | 0.40 | 0.00 | 0.20 | — | — |
| 35.8 | [The Asymptotics of Language Model Alignment with Memory](https://arxiv.org/abs/2610.01828) `2610.01828` | 0.50 | 0.40 | 0.40 | 0.00 | 0.20 | — | — |
| 35.8 | [A Safe Prototype Is Not a Safety Direction: Reference Dependence and Prompt Confounds in Response-Safety Embeddings](https://arxiv.org/abs/2610.01801) `2610.01801` | 0.50 | 0.40 | 0.40 | 0.00 | 0.20 | — | — |
| 35.8 | [From Rules to Neural Graphs: Scalable Structured Prediction for Patent Prior Art Search](https://arxiv.org/abs/2610.01553) `2610.01553` | 0.50 | 0.40 | 0.40 | 0.00 | 0.20 | — | — |
| 35.8 | [Do Multilingual Encoders Produce Language-Consistent Semantic IDs?](https://arxiv.org/abs/2610.01139) `2610.01139` | 0.50 | 0.40 | 0.40 | 0.00 | 0.20 | — | — |
| 35.8 | [Enterprise Representation Simplification (ERS): Reducing Representational Complexity for Enterprise AI](https://arxiv.org/abs/2610.00791) `2610.00791` | 0.50 | 0.40 | 0.40 | 0.00 | 0.20 | — | — |
| 35.8 | [Tight Transition Time Bounds for Separable Logistic Regression at the Edge of Stability](https://arxiv.org/abs/2610.01459) `2610.01459` | 0.50 | 0.40 | 0.55 | 0.00 | 0.00 | — | — |
| 35.8 | [The Price of Correlated Tests: How Strict Should a Model Release Gate Be?](https://arxiv.org/abs/2610.00993) `2610.00993` | 0.50 | 0.40 | 0.55 | 0.00 | 0.00 | — | — |
| 35.8 | [Learning Multiple Timescales for Goal-Conditioned Reinforcement Learning](https://arxiv.org/abs/2610.00849) `2610.00849` | 0.50 | 0.40 | 0.40 | 0.00 | 0.20 | — | — |
| 35.8 | [Signal-Noise Factorization Isolates Nuisance Variation into Removable Subspaces](https://arxiv.org/abs/2610.00751) `2610.00751` | 0.50 | 0.40 | 0.40 | 0.00 | 0.20 | — | — |
| 35.8 | [How Divergence Becomes Decision Flips in Compressed Language Models](https://arxiv.org/abs/2610.00694) `2610.00694` | 0.50 | 0.40 | 0.40 | 0.00 | 0.20 | — | — |
| 35.3 | [BranchIP: Learning Adaptive Equivariant Computation for Interatomic Potentials](https://arxiv.org/abs/2610.02013) `2610.02013` | 0.65 | 0.40 | 0.30 | 0.00 | 0.00 | — | — |
| 35.0 | [SILSA: Sliding-Window Slice Latents for Topology-Preserving High-Resolution 3D Generation](https://arxiv.org/abs/2610.02201) `2610.02201` | 0.50 | 0.40 | 0.30 | 0.00 | 0.28 | — | — |
| 34.7 | [Foundations without Fundamentals: Zero-Shot Blind Spots in Time Series FMs](https://arxiv.org/abs/2610.02058) `2610.02058` | 0.20 | 0.80 | 0.55 | 0.00 | 0.00 | — | — |
| 34.7 | [ScholarCatalyst: A Benchmark for Retrieving Papers That Inspire New Research](https://arxiv.org/abs/2610.02202) `2610.02202` | 0.50 | 0.45 | 0.30 | 0.00 | 0.20 | — | — |
| 34.7 | [Task-Oriented Rank Adaptation for Continual Learning in Text Classification](https://arxiv.org/abs/2610.01702) `2610.01702` | 0.50 | 0.45 | 0.45 | 0.00 | 0.00 | — | — |
| 33.7 | [Relative Transitions, Not Absolute Destinations: A Transfer-and-Ground Framework for Target-Trajectory-Free Human Mobility Generation](https://arxiv.org/abs/2610.02033) `2610.02033` | 0.50 | 0.40 | 0.45 | 0.00 | 0.00 | — | — |
| 33.7 | [Weather-Aware Domain Adaptation for Street-View Weather Recognition](https://arxiv.org/abs/2610.02000) `2610.02000` | 0.50 | 0.45 | 0.40 | 0.00 | 0.00 | — | — |
| 33.7 | [SIEVE: Selective attention-value Suppression for Vision-Language Models Unlearning](https://arxiv.org/abs/2610.01962) `2610.01962` | 0.50 | 0.40 | 0.45 | 0.00 | 0.00 | — | — |
| 33.7 | [A Comparative Explainability Framework for DeBERTa-v3 in Zero-Shot Medical Abstract Classification](https://arxiv.org/abs/2610.02116) `2610.02116` | 0.50 | 0.40 | 0.45 | 0.00 | 0.00 | — | — |
| 33.7 | [TabJoinBench: A Benchmark for Joinable Table Discovery](https://arxiv.org/abs/2610.00817) `2610.00817` | 0.50 | 0.45 | 0.40 | 0.00 | 0.00 | — | — |
| 33.7 | [Adaptive Conformal Prediction for Image Regression Models with Application to an Inertial Confinement Fusion Emulator](https://arxiv.org/abs/2610.00535) `2610.00535` | 0.50 | 0.40 | 0.30 | 0.00 | 0.20 | — | — |
| 32.6 | [Posterior sampling by source-space MCMC via prior-based few-step transport maps](https://arxiv.org/abs/2610.01034) `2610.01034` | 0.40 | 0.35 | 0.30 | 0.00 | 0.40 | — | — |
| 32.6 | [Effective Resistance and Graph Neural Network Reliability in Tissue-Specific Interactomes](https://arxiv.org/abs/2610.02175) `2610.02175` | 0.50 | 0.40 | 0.40 | 0.00 | 0.00 | — | — |
| 32.6 | [Linear Programming Representations and Strongly Polynomial Algorithms for Robust Markov Decision Processes](https://arxiv.org/abs/2610.02131) `2610.02131` | 0.50 | 0.40 | 0.40 | 0.00 | 0.00 | — | — |
| 32.6 | [Comparing a gradient boosting algorithm to the GOES FDC for wildfire detection](https://arxiv.org/abs/2610.01994) `2610.01994` | 0.50 | 0.40 | 0.40 | 0.00 | 0.00 | — | — |
| 32.6 | [Universal interpolation for deep residual self-attention networks](https://arxiv.org/abs/2610.01981) `2610.01981` | 0.50 | 0.40 | 0.40 | 0.00 | 0.00 | — | — |
| 32.6 | [AutoCompact: Learning When to Compact Context in Long-Horizon Coding Agents](https://arxiv.org/abs/2610.02163) `2610.02163` | 0.50 | 0.40 | 0.40 | 0.00 | 0.00 | — | — |
| 32.6 | [VETO: Video Efficient Token Optimization for Vision Language Models](https://arxiv.org/abs/2610.01785) `2610.01785` | 0.50 | 0.40 | 0.40 | 0.00 | 0.00 | — | — |
| 32.6 | [AURAL: Adaptive Latent Reasoning with Joint Chunk for Speech Language Models](https://arxiv.org/abs/2610.01560) `2610.01560` | 0.50 | 0.40 | 0.40 | 0.00 | 0.00 | — | — |
| 32.6 | [Q-SPT: Learnable Query-Based Compression for Low-Frame-Rate Speech Tokenization](https://arxiv.org/abs/2610.01492) `2610.01492` | 0.50 | 0.40 | 0.40 | 0.00 | 0.00 | — | — |
| 32.6 | [Not All Is Lost: Repairing Lossy User Preference States of Personalization Encoders](https://arxiv.org/abs/2610.01270) `2610.01270` | 0.50 | 0.40 | 0.40 | 0.00 | 0.00 | — | — |
| 32.6 | [Clifford Sheaf Neural Networks](https://arxiv.org/abs/2610.01322) `2610.01322` | 0.50 | 0.40 | 0.40 | 0.00 | 0.00 | — | — |
| 32.6 | [Gradient-Guided Density Peak Clustering](https://arxiv.org/abs/2610.01050) `2610.01050` | 0.50 | 0.40 | 0.40 | 0.00 | 0.00 | — | — |
| 32.6 | [Platonic Task Arithmetic](https://arxiv.org/abs/2610.00929) `2610.00929` | 0.50 | 0.40 | 0.40 | 0.00 | 0.00 | — | — |
| 32.6 | [Scalable Multi-Task Inverse Reinforcement Learning](https://arxiv.org/abs/2610.00758) `2610.00758` | 0.50 | 0.40 | 0.40 | 0.00 | 0.00 | — | — |
| 32.6 | [Learning to Price Electricity for Optimal Demand Response](https://arxiv.org/abs/2610.00755) `2610.00755` | 0.50 | 0.40 | 0.40 | 0.00 | 0.00 | — | — |
| 32.1 | [Generalization Is Stability, Not Accuracy: Multi-Axis Evaluation of LLMs](https://arxiv.org/abs/2610.01428) `2610.01428` | 0.30 | 0.45 | 0.30 | 0.00 | 0.43 | — | — |
| 31.6 | [Higher-Order Molecular Grammars for Generative and Foundation Models in Chemistry](https://arxiv.org/abs/2610.02186) `2610.02186` | 0.30 | 0.45 | 0.45 | 0.00 | 0.20 | — | — |
| 31.6 | [Optimizing Effective Training Time for Large-Scale Recommendation Systems](https://arxiv.org/abs/2610.02057) `2610.02057` | 0.50 | 0.45 | 0.30 | 0.00 | 0.00 | — | — |
| 31.1 | [Sample complexity bounds for categorical Markov random fields via Discrete Diffusions](https://arxiv.org/abs/2610.02128) `2610.02128` | 0.35 | 0.40 | 0.40 | 0.00 | 0.20 | — | — |
| 31.1 | [The hidden advantage of mask resampling: a theory of masked autoencoders](https://arxiv.org/abs/2610.01578) `2610.01578` | 0.25 | 0.40 | 0.55 | 0.00 | 0.20 | — | — |
| 30.5 | [TACO: Ternary Absolute-max Column-wise One-sparse Optimizer for LLM Fine-Tuning](https://arxiv.org/abs/2610.02199) `2610.02199` | 0.00 | 0.40 | 0.40 | 1.00 | 0.20 | — | — |
| 30.5 | [Decoding Looped Transformers Better for (Almost) Free](https://arxiv.org/abs/2610.02185) `2610.02185` | 0.50 | 0.40 | 0.30 | 0.00 | 0.00 | — | — |
| 30.5 | [Kolmogorov-Arnold Networks for Free-Boundary Partial Differential Equations](https://arxiv.org/abs/2610.02084) `2610.02084` | 0.50 | 0.40 | 0.30 | 0.00 | 0.00 | — | — |
| 30.5 | [Controllable Multi-label Video Safety Detection via Adaptive Tversky Policy Optimization](https://arxiv.org/abs/2610.02019) `2610.02019` | 0.50 | 0.40 | 0.30 | 0.00 | 0.00 | — | — |
| 30.5 | [The Other Half of Workflow Portability: Evidence-Backed HPC Site Profiles with Agentic Discovery](https://arxiv.org/abs/2610.00971) `2610.00971` | 0.50 | 0.40 | 0.30 | 0.00 | 0.00 | — | — |
| 30.5 | [Langevin-Informed Transfer Learning: Replacing Target Samples by Black-Box Feedback](https://arxiv.org/abs/2610.01522) `2610.01522` | 0.50 | 0.40 | 0.30 | 0.00 | 0.00 | — | — |
| 30.5 | [Tolerance-Based Fairness Auditing: Violation Certification and Sensitivity Screening](https://arxiv.org/abs/2610.01005) `2610.01005` | 0.50 | 0.40 | 0.30 | 0.00 | 0.00 | — | — |
| 30.5 | [CARM: Cancellation-Aware Response Masking for LLM Reinforcement Learning](https://arxiv.org/abs/2610.02039) `2610.02039` | 0.30 | 0.45 | 0.55 | 0.00 | 0.00 | — | — |
| 30.5 | [Devils in Question Relay: Source-Conditioned Relay Steering to Mitigate Hallucinations in Audio-visual Large Language Models](https://arxiv.org/abs/2609.37568) `2609.37568` | 0.30 | 0.45 | 0.40 | 0.00 | 0.20 | — | — |
| 30.0 | [When Users Change Their Minds: Measuring and Repairing Intent Drift in LLM Agents](https://arxiv.org/abs/2609.32520) `2609.32520` | 0.30 | 0.45 | 0.30 | 0.00 | 0.30 | — | — |
| 30.0 | [Madeleine: Learning Involuntary Recall for Conversational Memory from Simulated Lives](https://arxiv.org/abs/2610.01118) `2610.01118` | 0.45 | 0.75 | 0.00 | 0.00 | 0.00 | — | — |
| 29.5 | [One Basis to Animate Them All: Gaussian Blendshape Distillation for Real-Time Avatars](https://arxiv.org/abs/2610.02207) `2610.02207` | 0.40 | 0.40 | 0.40 | 0.00 | 0.00 | — | — |
| 29.5 | [On Language Drift during RLVR Post-Training](https://arxiv.org/abs/2610.02015) `2610.02015` | 0.30 | 0.40 | 0.55 | 0.00 | 0.00 | — | — |
| 29.5 | [KaliBench: A Fine-Grained Benchmark for Cybersecurity Tool Use on Kali Linux with Runtime-Free Verifiable Rewards](https://arxiv.org/abs/2610.02206) `2610.02206` | 0.00 | 0.45 | 0.30 | 1.00 | 0.20 | — | — |
| 29.5 | [Old Ideas, Novel Problems: The Instability of LLM-Based Novelty Evaluation](https://arxiv.org/abs/2610.02022) `2610.02022` | 0.30 | 0.40 | 0.40 | 0.00 | 0.20 | — | — |
| 29.3 | [Do Audio LLMs Listen Before They Act? Diagnosing Acoustic-Context Gating in Voice Agents](https://arxiv.org/abs/2609.32536) `2609.32536` | 0.30 | 0.45 | 0.30 | 0.00 | 0.26 | — | — |
| 28.9 | [Acmite: Mitigating Gender Bias in LLMs through Concept-Guided Mutual Information](https://arxiv.org/abs/2610.01696) `2610.01696` | 0.45 | 0.40 | 0.30 | 0.00 | 0.00 | — | — |
| 28.8 | [Argo-Bench: Evaluating Data Agents on Enterprise-Scale Workflows](https://arxiv.org/abs/2610.02122) `2610.02122` | 0.00 | 0.45 | 0.30 | 1.00 | 0.16 | — | — |
| 28.4 | [A Matryoshka Hierarchical RAG for Efficient Multi-Hop Question Answering](https://arxiv.org/abs/2610.01767) `2610.01767` | 0.30 | 0.45 | 0.30 | 0.00 | 0.20 | — | — |
| 28.4 | [CANOPY: Adaptive-Granularity Evidence Compression for Multimodal RAG](https://arxiv.org/abs/2610.00923) `2610.00923` | 0.30 | 0.45 | 0.30 | 0.00 | 0.20 | — | — |
| 27.9 | [In-context Learning of Single-index Targets: Comparing Kernel and Feature Learners](https://arxiv.org/abs/2610.01712) `2610.01712` | 0.25 | 0.40 | 0.40 | 0.00 | 0.20 | — | — |
| 27.9 | [Training-Free Diffusion Planning with Analytical Local Scores](https://arxiv.org/abs/2610.01959) `2610.01959` | 0.35 | 0.40 | 0.40 | 0.00 | 0.00 | — | — |
| 27.4 | [Universal Byte-Level Encoding: UTF-8/UTF-16 Routing to Reduce Cross-Script Token-Budget Disparities](https://arxiv.org/abs/2610.01984) `2610.01984` | 0.30 | 0.45 | 0.40 | 0.00 | 0.00 | — | — |
| 27.4 | [A rubric landscape for evaluating clinical reasoning in large language models: what exists, what is missing, and what needs to be combined](https://arxiv.org/abs/2610.01938) `2610.01938` | 0.30 | 0.45 | 0.40 | 0.00 | 0.00 | — | — |
| 27.4 | [Pay for the Fault, Not the Flow: Label-Free In-Flow Multi-Agent Workflow Optimization](https://arxiv.org/abs/2610.01017) `2610.01017` | 0.30 | 0.45 | 0.40 | 0.00 | 0.00 | — | — |
| 26.8 | [FlexRouter: Learning Complementary Model Sets for Flexible LLM Routing](https://arxiv.org/abs/2609.38585) `2609.38585` | 0.30 | 0.45 | 0.30 | 0.00 | 0.10 | — | — |
| 26.7 | [Prompt2Skill: Unsupervised Skill Optimization From Natural Language Instructions](https://arxiv.org/abs/2609.38593) `2609.38593` | 0.30 | 0.40 | 0.30 | 0.00 | 0.16 | — | — |
| 26.3 | [Mingbird: A Local-First Agent Harness Enabling Small Open Models to Complete Real Tasks](https://arxiv.org/abs/2610.02001) `2610.02001` | 0.00 | 0.45 | 0.30 | 1.00 | 0.00 | — | — |
| 26.3 | [Where LLMs Fail with Visualization DSLs](https://arxiv.org/abs/2610.01873) `2610.01873` | 0.30 | 0.40 | 0.40 | 0.00 | 0.00 | — | — |
| 26.3 | [Beyond Linear Concepts: Discovering and Aligning Non-Linear Concept Manifolds in Large Language Models](https://arxiv.org/abs/2610.01821) `2610.01821` | 0.30 | 0.40 | 0.40 | 0.00 | 0.00 | — | — |
| 26.3 | [What Makes Something Hard(er)? Explaining Question Difficulty in Natural Language](https://arxiv.org/abs/2610.01627) `2610.01627` | 0.30 | 0.40 | 0.40 | 0.00 | 0.00 | — | — |
| 26.3 | [Auditing Web Agent Evaluation on WebArena-Lite: Human Review of Outcomes and Trajectories](https://arxiv.org/abs/2610.01491) `2610.01491` | 0.30 | 0.40 | 0.40 | 0.00 | 0.00 | — | — |
| 25.8 | [Faynt: Scaling and Optimizing Policies for Competitive Melee](https://arxiv.org/abs/2610.02144) `2610.02144` | 0.25 | 0.45 | 0.40 | 0.00 | 0.00 | — | — |
| 25.8 | [Wasserstein Gradient Flows and Forward-Only Diffusion Are Not Enough for Multimodal Sampling](https://arxiv.org/abs/2610.02081) `2610.02081` | 0.35 | 0.40 | 0.15 | 0.00 | 0.20 | — | — |
| 25.3 | [Do Your Own Research: Learning to Forecast by Learning to Search](https://arxiv.org/abs/2610.01955) `2610.01955` | 0.00 | 0.40 | 0.30 | 1.00 | 0.00 | — | — |
| 25.3 | [The Missing Primitive: Diagnosing and Repairing Mathematical Reasoning in Large Language Models](https://arxiv.org/abs/2610.02191) `2610.02191` | 0.30 | 0.45 | 0.30 | 0.00 | 0.00 | — | — |
| 25.3 | [MoLE: Mixture of Latent Experts for Complementary Visual Reasoning](https://arxiv.org/abs/2610.01917) `2610.01917` | 0.50 | 0.45 | 0.00 | 0.00 | 0.00 | — | — |
| 25.3 | [Which LLM to pick? Online Active Model Selection for Large Language Models](https://arxiv.org/abs/2610.01592) `2610.01592` | 0.30 | 0.45 | 0.30 | 0.00 | 0.00 | — | — |
| 25.3 | [OverAct: Measuring and Mitigating Proactive Over-Authorization in LLM Tool-Calling Agents](https://arxiv.org/abs/2610.01508) `2610.01508` | 0.30 | 0.45 | 0.30 | 0.00 | 0.00 | — | — |
| 24.2 | [Error-Corrected Inference-Time Scaling for Imperfect Diffusion Models](https://arxiv.org/abs/2610.01933) `2610.01933` | 0.10 | 0.40 | 0.30 | 0.00 | 0.40 | — | — |
| 24.2 | [Every Ablation Is a Dose: Counterweights and the Semblance of Self-Repair](https://arxiv.org/abs/2610.02173) `2610.02173` | 0.50 | 0.40 | 0.00 | 0.00 | 0.00 | — | — |
| 24.2 | [Scalable, Transferable Meta-network for Data Selection Requires a Different Loss (and Why the Obvious Choice is Problematic)](https://arxiv.org/abs/2610.02092) `2610.02092` | 0.30 | 0.40 | 0.30 | 0.00 | 0.00 | — | — |
| 24.2 | [Learn the Directions, Normalize the Gains: Post-Training Normalization for LoRA](https://arxiv.org/abs/2610.02067) `2610.02067` | 0.50 | 0.40 | 0.00 | 0.00 | 0.00 | — | — |
| 24.2 | [FastCI: Efficient GPU-Intensive CI for LLM Training Frameworks](https://arxiv.org/abs/2610.01967) `2610.01967` | 0.30 | 0.40 | 0.30 | 0.00 | 0.00 | — | — |
| 24.2 | [LLM2Jev: LLMs Are Already Jev-Style Decision Models -- When and How to Fine-Tune Them](https://arxiv.org/abs/2610.02076) `2610.02076` | 0.30 | 0.40 | 0.30 | 0.00 | 0.00 | — | — |
| 24.2 | [Latent JEPA: Abstract Future Prediction for Latent Reasoning in Chemistry](https://arxiv.org/abs/2610.01947) `2610.01947` | 0.30 | 0.40 | 0.30 | 0.00 | 0.00 | — | — |
| 24.2 | [AgentWebRec: Compact Evidence Fusion over the Agent Web for Personalized Recommendation](https://arxiv.org/abs/2610.01705) `2610.01705` | 0.30 | 0.40 | 0.30 | 0.00 | 0.00 | — | — |
| 24.2 | [Neither Black nor White: Balancing Semantic and Collaborative Signals with Graph-Informed Semantic IDs (GrIS)](https://arxiv.org/abs/2610.01533) `2610.01533` | 0.50 | 0.40 | 0.00 | 0.00 | 0.00 | — | — |
| 24.2 | [Learning to structure data from user-generated thematic corpora](https://arxiv.org/abs/2610.01463) `2610.01463` | 0.30 | 0.40 | 0.30 | 0.00 | 0.00 | — | — |
| 24.2 | [Optimal Transport Meets Reinforcement Learning: A Survey](https://arxiv.org/abs/2610.01413) `2610.01413` | 0.50 | 0.40 | 0.00 | 0.00 | 0.00 | — | — |
| 23.7 | [RPTune: Learned Context Curation for LLM Catalog Search](https://arxiv.org/abs/2610.00964) `2610.00964` | 0.30 | 0.00 | 0.30 | 0.00 | 0.50 | — | — |
| 23.2 | [Smaller Models, Better Rejects: Preference Distillation Scaling](https://arxiv.org/abs/2609.38987) `2609.38987` | 0.00 | 0.40 | 0.55 | 0.00 | 0.20 | — | — |
| 22.6 | [Cross-Lingual Alignment for Decoder-Only Models using MoE Routers](https://arxiv.org/abs/2610.01921) `2610.01921` | 0.05 | 0.40 | 0.60 | 0.00 | 0.00 | — | — |
| 22.6 | [Embedding Prediction Helps Image Generation](https://arxiv.org/abs/2610.02203) `2610.02203` | 0.35 | 0.00 | 0.40 | 0.00 | 0.20 | — | — |
| 22.6 | [Keyword Harnesses Fail Open: A Cheap Diagnostic Ladder for Tool-Use Claims in Small Language Models](https://arxiv.org/abs/2610.02142) `2610.02142` | 0.15 | 0.45 | 0.40 | 0.00 | 0.00 | — | — |
| 22.1 | [ChainLoRA: Geometry-Preserving Task Vector Merging for Continual Learning in LLMs](https://arxiv.org/abs/2610.00431) `2610.00431` | 0.30 | 0.45 | 0.00 | 0.00 | 0.20 | — | — |
| 21.1 | [QK-Wanda: Coupling Queries and Keys for Unstructured Pruning](https://arxiv.org/abs/2610.01554) `2610.01554` | 0.00 | 0.40 | 0.60 | 0.00 | 0.00 | — | — |
| 19.5 | [Prefill-Free Cross-Family KV Cache Transfer for Heterogeneous Multi-Agent LLMs](https://arxiv.org/abs/2609.32259) `2609.32259` | 0.00 | 0.45 | 0.40 | 0.00 | 0.10 | — | — |
| 17.9 | [Local Support Learning](https://arxiv.org/abs/2610.02126) `2610.02126` | 0.00 | 0.40 | 0.30 | 0.00 | 0.20 | — | — |
| 17.9 | [Can LLMs Reliably Annotate Bioassay Metadata to Improve Data Readiness?](https://arxiv.org/abs/2610.01616) `2610.01616` | 0.30 | 0.00 | 0.40 | 0.00 | 0.00 | — | — |
| 17.9 | [Comparison of Common Crawl News & GDELT](https://arxiv.org/abs/2610.00587) `2610.00587` | 0.30 | 0.00 | 0.40 | 0.00 | 0.00 | — | — |
| 16.3 | [Beyond Decodability: Do Acoustic Factors Drive Predictions in Speech-Based Alzheimer's Assessment?](https://arxiv.org/abs/2610.01846) `2610.01846` | 0.25 | 0.40 | 0.00 | 0.00 | 0.00 | — | — |
| 14.7 | [Counterfactual Auditing of Bias in Open-Source Large Language Models for Clinical Triage](https://arxiv.org/abs/2610.01963) `2610.01963` | 0.00 | 0.40 | 0.30 | 0.00 | 0.00 | — | — |
| 14.7 | [Sequential Functional Structured Tucker Compression for Large Language Model Attentions](https://arxiv.org/abs/2610.00717) `2610.00717` | 0.00 | 0.40 | 0.30 | 0.00 | 0.00 | — | — |

Scores: weighted mean of rubric criteria (0–1 each) × 100; see README for the rubric.