arXiv subjects

Youran Sun

Publications and source records attributed to Youran Sun.

At least 19 recordsLinked to original sources

arXiv2026

Asking What a Director Sees Can Change What Vision-Language Models Choose

Vision-language models are now routinely tested on theory of mind: whether they can work out what another person sees, knows, and means. A common design asks the model to report that perspective and then to act for the person, and reads a correct report beside a wrong action as a failure to use the perspective. That reading rests on a quiet assumption: that asking for the report leaves the action unchanged. We show that it does not. In a rendered Director Task, a partition hides the largest block from a director who asks for "the largest block". Asking first which blocks the director can see raises the share of replies that choose this hidden block by 12 to 88 percentage points on six GPT models and by 17 on Qwen3.8-27B, even though the models mostly report correctly which blocks the director sees. Reasoning traces and prompt rewordings suggest why: the question shifts how the request is read, from the director's view toward the model's own. When the prompt states what the director knows and refers to, fewer than 1% of replies pair a correct report with a hidden-block choice. Any evaluation that elicits a perspective report before the action it scores needs a baseline without the question, and we deliver one: paired prompts that differ only in that question, a design any perspective-taking benchmark can adopt.

cs.CL↗

arXiv2026

Sci$Π$vot: Benchmarking Principal-Investigator Judgment at Pivotal Moments of Scientific AutoResearch Agents

Large language model (LLM) agents can keep scientific experiments running for months with little human supervision, but reliable autonomous research also requires scientific judgment. Evaluating scientific judgment is difficult because research decisions are open-ended and depend on project context. We construct Sci$Π$vot, a 248-case benchmark of scientific judgment based on human corrections. The benchmark draws on records from six months of research in which senior researchers ran LLM agents on 26 projects across six scientific fields and repeatedly intervened to correct the agents' research judgments. In each case, an agent chooses between the original LLM plan and the human-corrected alternative, using only the project record available before the correction as context. We compare LLM and human performance and examine the effects of reasoning effort, access to the project record, and reconsideration. At medium reasoning effort with the project record, models systematically prefer the original LLM plan over the human-corrected alternative. Four of five frontier models perform below chance on the selected cases. The highest accuracy in the project condition across the tested models and reasoning effort settings is 63.5%, compared with 73.4% for fresh researchers. At medium effort, every model achieves lower accuracy with the project record than without it, and reconsideration does not consistently improve overall accuracy. Even with a human-corrected alternative available, frontier models struggle with the selected research decisions, while Sci$Π$vot supports systematic evaluation of scientific judgment for LLM agents.

cs.AI↗

arXiv2026

Scoring the Wrong Question: Readout Failures in Constrained-Option Evaluation

Constrained-option scoring reads a model's probabilities for a fixed set of permitted answers, so it returns a score even when the model is about to write something else. We study a prompt that quotes a multiple-choice item, ending in the item's own answer instruction, and asks for a forecast of whether a reader model given only a short note will answer it correctly. The model can begin answering the quoted item: on three Qwen3 checkpoints an answer letter is the most probable token on nearly every passage, and the forecast ranks the reader's correctness indistinguishably from chance. Beyond prior work's argmax check and prefill repair, we contribute a label-free diagnostic of the declared options' support, their first-token mass in context; a deletion test tracing the Qwen3 answer-letter start to the quoted instruction; and rendering and tokenisation checks for dropped separators and coinciding first tokens. Prefilling an answer stem lifts the median option mass from below to above one half on all nine checkpoints scored with this prompt and, with options and renormalisation unchanged, raises the Qwen3-averaged AUC from 0.4929 to 0.6152. lm-polygraph's default P(True) estimator shows a related failure: it reads True where Qwen3 without thinking favours an answer letter (MMLU) or the prompt's (A)/(B) label (TriviaQA), and its Qwen3-averaged AUROC is indistinguishable from chance on MMLU and below chance on TriviaQA. Scoring and renormalising the (A)/(B) labels after an answer stem raises it to 0.632 and 0.868, renormalising True against False in place reaches 0.660 and 0.860, and on TriviaQA both exceed all fourteen default single-answer library estimators. The original forecast is renormalised too, yet stays indistinguishable from chance: option mass flags low support in both settings without labels, and only checking the intended target shows which score still ranks.

cs.CL↗

arXiv2026

When an Evaluation Rule Writes Training Labels: Measuring Human-Reference Forgiveness in NAVSIM

When the human reference scores zero on a metric, the released GTRS-Dense label generator for NAVSIM marks every candidate trajectory in the scene as passing it. NAVSIM's authors introduced this human-reference forgiveness to avoid penalizing contextually justified maneuvers when scoring one trajectory, and warned that it could overlook important failures. In label generation it sets a whole column of 16,384 candidate targets to passing. To measure the consequences for supervision, we re-run the generator with the overwrite disabled and compare the pre-overwrite targets with the released labels on all 103,288 navtrain scenes. The rule erases a candidate distinction that the training loss reads on 11,237 of them (10.8793%). Firing usually changes most of a column: lane keeping carries 9,982 of the 13,042 forgiven loss columns, and its median forgiven column had 14,391 of 16,384 candidates failing before the overwrite. On held-out navtest scenes forgiven on lane keeping, the released lane-keeping head's median AUC against the pre-overwrite outcome is 0.7095; on unforgiven scenes matched on failing-candidate count it is 0.9807. For the Hydra-MDP checkpoint released with GTRS, whose configuration takes the same label file, the two values are 0.6627 and 0.9761. Continuing the released GTRS-Dense checkpoint for 300 optimizer steps with three paired seeds, we observe the forgiven-scene AUC 0.1086-0.1251 higher with pre-overwrite than with published targets, and a narrower gap between matched groups, still above zero. Scoring with forgiveness disabled, we observe lane keeping higher by 2.478-3.524 points on navtest scenes forgiven on any of five loss metrics, with lower adjacent-frame plan consistency. Both changes are larger there than on the rest. EPDMS, scored the same way, does not separate the two target sets.

cs.LG↗

arXiv2026

Bringing Agentic Search to Earth Observation Data Discovery

NASA and its data centers hold thousands of geoscience datasets and tools like Worldview, Giovanni, the Science Discovery Engine, and Harmony. Finding the right one is hard even for domain experts. We present an agentic search framework for geoscience data discovery that takes a natural-language research query and returns matching datasets and tools. We demonstrate that, in the era of large language models, the latent value of knowledge graphs (KGs) can be substantially amplified through agentic search. From the NASA Earth Observation Knowledge Graph (NASA EO-KG) we derive NASA-EO-Bench, an open benchmark of 47k query-dataset pairs (21k task-based queries). A neural scorer fine-tuned on NASA-EO-Bench beats cosine and BM25 baselines. Further combining it with BM25 via score fusion raises both Recall@10 (R@10) and MRR to over 5x the unadapted cosine baseline. On top of this supervised pipeline, a zero-shot reranking stage lifts MRR by 16%, significant under a paired bootstrap, with no additional training, and autonomous web and arXiv tool use adds a further gain, showing that LLM reasoning is complementary to supervised retrieval.

cs.IR↗

arXiv2026

Staged Depth Training: A Representation Curriculum for PINNs

Representation quality is a central determinant of PINNs' performance, yet standard training leaves representations to emerge implicitly while fitting the final solution. We introduce \textbf{representation curriculum}, an ordered process in which representations are explicitly learned, transferred independently of their predictors, and progressively refined. We realize it with Staged Depth Training (SDT), which trains a shallow prefix under a temporary physics-informed head, discards the head, and freezes the learned prefix while adding depth, without equation-specific encodings or changes to the final architecture. Across the 20 default forward problems in PINNacle with three backbones, SDT improves 40 of 59 equal-budget problem--backbone cells by at least 5\% and remains within that band in the rest, with a 32.8\% geometric-mean error reduction on a PirateNet-style backbone. Mechanistic ablations suggest that the gain is not explained by optimizer restarts or shallow warm-starting alone. Representation visualizations and hyperparameter-basin analyses provide diagnostic evidence on representation geometry and local sensitivity to shared hyperparameters. On Poisson--Boltzmann 2D, SDT also more than doubles the fitted depth-scaling exponent for both backbones. These results support representation curriculum as a promising training strategy for improving PINNs while preserving the deployed architecture and inference cost.

cs.LG↗

arXiv2026

AutoSR: Automatic Symbolic Regression by Searching Research States

We introduce Automatic Symbolic Regression (AutoSR), a fully automated system that instantiates Research-Space Symbolic Regression by searching persistent scientific investigations rather than isolated equations. Finite, noisy data often yield numerically competitive expressions that imply very different behavior outside the observed regime, making numerical fit and syntactic complexity insufficient measures of scientific credibility. Existing approaches largely focus on improving expressions, yet the search typically retains little beyond the resulting formula and score, losing the scientific record, such as motivations and probes, that inform what to try next. AutoSR preserves this record in a \textbf{Research State}, coupling each candidate equation with the reasoning, computational evidence, and independent review developed along its branch. Proposer--reviewer agents develop these states under progressive-widening Monte Carlo tree search (PW-MCTS), which allocates computation across competing investigations, while the accumulated research record is ultimately synthesized into a final report that explains the leading relation and the basis for its selection. Across nine selected challenges from two benchmark suites, AutoSR recovers algebraically equivalent relations in every case, including three cp3-bench problems that no published system recovers and six structurally diverse LSR-Transform problems. Overall, AutoSR extends symbolic regression from equation-level search toward automated scientific investigation, allowing scientific knowledge and accumulated evidence to shape both what is explored and how the resulting equation is justified.

cs.SC↗

arXiv2026

AgonAlpha: Autonomous Alpha Discovery via Prompt Economy and Scalable Agentic Search

Language models can propose many plausible trading factors, but an autonomous research system must also allocate its evaluation budget, verify its own evidence, and preserve how each candidate was produced. We present AgonAlpha, an architecture that searches over frozen research artifacts---hypotheses, executable expressions, platform evidence, rationales, and review status---rather than formulas alone. To our knowledge, AgonAlpha is the first alpha-mining system to combine verified artifact search, a fresh-context adversarial reviewer with re-execution and veto authority, and pending-aware parallel budget allocation, together with a complete public evidence trail. Independent deployments on WorldQuant BRAIN produced SPECTACULAR-grade alphas across five users and six model backends, with Fitness reaching 9.50 and Sharpe reaching 3.48, while retaining prompt-to-expression provenance for every submission.

cs.AI↗

arXiv2026

Understanding Sparse Attention Selectivity in Long-Context Foundation Models via Counterfactual Evaluation

Sparse attention is widely deployed in long-context serving stacks, yet no framework audits how discarding blocks changes the influence of specific content on model output. We first establish that the phenomenon is real and causal: Block Sparse Flash Attention (BSFA) route replay across four architectures changes output decisions in 13 of 16 cells, with zero identity-replay label flips. We then introduce a dense-calibrated counterfactual audit using matched probe cards---Gold (carrying the correct answer label), Poison (carrying a target wrong label), and Benign (filler only)---under six-layout position symmetry, isolating the sparsification-specific effect. Two patterns compete. Signal concentration: the selector preserves Gold and Poison blocks far above filler-matched Benign blocks (G$\approx$P$\gg$B across all model--task pairs). Integration loss: discarding blocks severs cross-block attention---confirmed by an ablation where isolating the probe block collapses its influence from 4.48 logits to zero. Compression ratio governs the balance: a full sweep from mild ($c=0.25$) to aggressive ($c=0.75$) compression across four model--task pairs reveals that three of four cells move toward stronger sparse amplification at higher compression, with two exhibiting sign reversals. Three independent arms---BSFA route replay, controlled block-top-$k$, and KV-cache eviction---converge: sparsification changes content influence in ways aggregate accuracy cannot detect. We provide an open measurement framework deployable on any model exposing block identities.

cs.CL↗

arXiv2026

Cross-Task Dissociation in Frontier Vision-Language Model Theory of Mind

Do frontier vision-language models present a coherent Theory-of-Mind (ToM) profile across tasks, matching the same human reference group, or does that profile fragment from one paradigm to the next? We evaluate a shared panel of nine frontier VLMs on two psychology-derived benchmarks: the Keysar Director Task (visual perspective-taking under egocentric interference) and the Frith-Happé animated triangles scored with the Castelli rubric (intention attribution from pure motion). On the Director Task, without chain-of-thought, the panel makes the egocentric error on 78\% of trials like children rather than adults; variation is substantial across models, and reasoning rescues several models. On the triangles, the panel under-attributes intention: its ToM profile sits more than three times closer to the high-functioning-autistic-adult (HF-ASD) mean than to the typical-development-adult (TD) mean, while Goal-Directed and Random stay near TD. No model is nearest TD on both tasks; the model that looks adult-like on the Director Task falls on the HF-ASD side on the triangles, and the most TD-like model on the triangles is child-like on the Director Task. We report group-level descriptions, not diagnostic labels for any model.

cs.CL↗

arXiv2026

Effective Parameters, Real Behavior: Renormalization for Robotics -- From Infinite Electron Mass to Sim-to-Real Gap

Bridging the sim-to-real gap is a central problem in robotics, and the prevailing approach is to build increasingly accurate simulators. Here, we propose another approach based on renormalization: using effective, resolution-dependent parameters to absorb details omitted by the simulator and reproduce real behavior. These parameters may differ from measured physical values because they compensate for what the simulator leaves out. We demonstrate this mechanism analytically for proportional--derivative (PD) control at finite simulation frequency, where proportional feedback changes the effective derivative gain and derivative feedback changes the effective inertia. We then interpret dynamic rope manipulation and underwater swimming through the same perspective. Finally, we present a practical procedure for choosing observables, identifying omitted physics, and determining effective parameters. Renormalization offers robotics a complementary path across the sim-to-real gap: effective parameters, real behavior.

cs.RO↗

arXiv2026

RouterVLA: Budgeted Commissioning and Expert Onboarding for Growing VLA Pools

Robotic teams often maintain several vision-language-action policies but still deploy one global winner. We study two recurring decisions: which expert to deploy for a new condition and which candidate to add to the pool. RouterVLA combines a split-clean prior and outcome-disjoint probes with onboarding that credits only failures the incumbent pool cannot handle. Under an exactly cost-matched probe budget, it reaches 60.53\% held-out success, a $+1.64\pp$ gain over a semantic shortlist. Both criteria independently converge on the same five experts, confirming that the candidates best positioned to cover the base pool's blind spots are also broadly capable.

cs.RO↗

arXiv2026

PerspectiveGap: A Benchmark for Multi-Agent Orchestration Prompting

Real-world LLM applications are moving beyond single-agent workflows toward orchestrated multi-agent systems, yet current models still struggle to determine what each sub-agent needs to know. To measure this, we introduce PerspectiveGap, a benchmark for evaluating LLMs' ability to compose orchestration prompts for multi-agent systems. PerspectiveGap contains 110 scenarios, each evaluated through two distractor-mixed task formats: role-fragment assignment and free-form prompt writing. These scenarios are organized into 10 topologies, which are distilled from the authors' real-world engineering practice and framed by the Prompt Economy principle: building loop-centered orchestrations that maximize utility with minimal role and engineering overhead. In experiments with 33 commercial models from 10 companies, GPT-5.5 substantially outperforms all competitors, whereas Opus 4.8 shows a notable weakness in orchestration prompting despite its strong coding performance. Nevertheless, PerspectiveGap remains challenging: the evaluated models achieve an average combined pass rate of only 17.2\% (GPT-5.5 62.0\%) and an average overall leakage rate of 217.9\% (a per-scenario information leak-event count, not a proportion; GPT-5.5 49.1\%). These findings suggest that multi-agent orchestration prompting is a distinct and under-evaluated capability, and PerspectiveGap provides a foundation for measuring and improving it systematically.

cs.CL↗

arXiv2026

Agon: An Autonomous Large-Scale Omnidisciplinary Research System Built on Prompt Economy

Large language models are making research production scalable, shifting the bottleneck from producing artifacts to judging claims. We present \textsc{Agon}, a research orchestrator that validates what can be checked inside the workflow and leaves the remaining judgments to human scientists. \textsc{Agon} is built on six design principles: Prompt Economy, Future-Facing, Minimal Prompts, OmniDisciplinary, Massive Parallelism, and Zero-Code. We ran \textsc{Agon} across domains for 444 iterations of Prompt Economy loops, using only small starting topics and no human-written experimental code. These deployments demonstrate scalability while exposing new classes of failure. We organize these failures into a taxonomy along severity, fixability, visibility, and capability locus. The taxonomy separates failures the loops can see and fix from those that require human judgment. Together, these results show that \textsc{Agon} is pushing research toward a new paradigm: machine scales, human steers.

cs.SE↗

arXiv2026

Correcting Mean Bias in Text Embeddings: A Refined Renormalization with Training-Free Improvements on MMTEB

We find that current sentence-embedding models produce outputs with a consistent bias: every embedding $e$ decomposes as $\tilde e + μ$, where the mean $μ$ is near-identical across all sentences. We study two training-free corrections -- subtracting $μ$ directly (R1), or projecting each embedding off the mean direction (R2) -- and show, via a first-order error-propagation argument, that R2 cancels the parallel component of mean-estimation error that R1 retains. Across 38 models on the Massive Multilingual Text Embedding Benchmark (MMTEB)~\citep{MMTEB}, R2 yields consistent classification gains (paired $\bar t = 3.31$, 29 of 38 models with $t>2$, zero losses), and the per-model mean norm $\Vertμ\Vert$ correlates with which models benefit most. A nine-method dose-response ablation on five models further reveals that mild single-direction removal helps, but full principal component analysis (PCA) whitening hurts every model we test, and that R2 and All-but-the-Top with depth one agree within $0.18$ pp downstream despite weak geometric alignment between $\hatμ$ and the centered top principal component.

cs.CL↗

arXiv2026

In-depth Analysis of Graph-based RAG in a Unified Framework

Graph-based Retrieval-Augmented Generation (RAG) has proven effective in integrating external knowledge into large language models (LLMs), improving their factual accuracy, adaptability, interpretability, and trustworthiness. A number of graph-based RAG methods have been proposed in the literature. However, these methods have not been systematically and comprehensively compared under the same experimental settings. In this paper, we first summarize a unified framework to incorporate all graph-based RAG methods from a high-level perspective. We then extensively compare representative graph-based RAG methods over a range of questing-answering (QA) datasets -- from specific questions to abstract questions -- and examine the effectiveness of all methods, providing a thorough analysis of graph-based RAG approaches. As a byproduct of our experimental analysis, we are also able to identify new variants of the graph-based RAG methods over specific QA and abstract QA tasks respectively, by combining existing techniques, which outperform the state-of-the-art methods. Finally, based on these findings, we offer promising research opportunities. We believe that a deeper understanding of the behavior of existing methods can provide new valuable insights for future research.

cs.IR↗

arXiv2026

Semantic Data Processing with Holistic Data Understanding

Semantic operators have increasingly become integrated within data systems to enable processing data using Large Language Models (LLMs). Despite significant recent effort in improving these operators, their accuracy is limited due to a critical flaw in their implementation: lack of holistic data understanding. In existing systems, semantic operators often process each data record independently using an LLM, without considering data context, only leveraging LLM's dataset-agnostic interpretation of the user-provided task. However, natural language is imprecise, so a task can only be accurately performed if it is correctly interpreted in the context of the dataset. For example, for classification and scoring tasks, which are typical semantic map tasks, the standard method of processing each record row by row yields inaccurate results in a wide range of datasets. We propose HoldUp, a new method for semantic data processing with holistic data understanding. HoldUp processes records jointly, leveraging cross-record relationships to correctly interpret the task within the data context. Enabling holistic data understanding, however, is challenging due to what we call LLM data understanding paradox: while large representative data subsets are necessary to provide context, feeding long inputs to LLMs causes quality degradation due to well-known long-context issues. To resolve this paradox, we develop a novel clustering algorithm to identify the latent structure within the dataset through judicious use of LLMs, inspired by bagging. Using this approach as a primitive, we develop novel clustering-based classification and scoring methods to perform these two tasks with high accuracy. Experiments across 15 real-world datasets show that HoldUp consistently outperforms existing solutions, providing up to 33% higher accuracy for classification and 30% higher accuracy for scoring and clustering tasks.

cs.DB↗

arXiv2026

ReSearch: A Multi-Stage Machine Learning Framework for Earth Science Data Discovery

The rapid expansion of Earth Science data from satellite observations, reanalysis products, and numerical simulations has created a critical bottleneck in scientific discovery, namely identifying relevant datasets for a given research objective. Existing discovery systems are primarily retrieval-centric and struggle to bridge the gap between high-level scientific intent and heterogeneous metadata at scale. We introduce \textbf{ReSearch}, a multi-stage, reasoning-enhanced search framework that formulates Earth Science data discovery as an iterative process of intent interpretation, high-recall retrieval, and context-aware ranking. ReSearch integrates lexical search, semantic embeddings, abbreviation expansion, and large language model reranking within a unified architecture that explicitly separates recall and precision objectives. To enable realistic evaluation, we construct a literature-grounded benchmark by aligning natural language intent with datasets cited in peer-reviewed Earth Science studies. Experiments demonstrate that ReSearch consistently improves recall and ranking performance over baseline methods, particularly for task-based queries expressing abstract scientific goals. These results demonstrate the importance of intent-aware, multi-stage search as a foundational capability for reproducible and scalable Earth Science research.

cs.DB↗