absorb.md

Wes Roth

Chronological feed of everything captured from Wes Roth.

paper / wesroth / Jun 14

HAWC Observatory data yields no evidence for axion-like particles in M87, setting new constraints.

This research utilized 7.5 years of HAWC Observatory data to conduct an indirect search for axion-like particles (ALPs) by examining photon-ALP mixing in M87's environment. Although no definitive signal of photon-ALP conversion was detected, the study established constraints on ALPs within a specific mass and coupling constant range. These findings are consistent with prior research and contribute to the exclusion of ALPs in this parameter space.

axion-like-particlesdark-mattergamma-ray-astronomyhawc-observatorym87particle-astrophysicsphoton-axion-mixing
“The HAWC Observatory analyzed 7.5 years of data from M87 to search for axion-like particles (ALPs) by observing their photon-ALP mixing in magnetic fields.”
paper / wesroth / Jun 14

SOMA-SQL: Autonomous Ambiguity Resolution for Enhanced NL-to-SQL Translation

SOMA-SQL is a novel approach to Natural Language to SQL (NL-to-SQL) translation that addresses ambiguity in user queries and database schemas. It autonomously resolves these ambiguities by generating synthetic query logs to ground schema interpretations and employing targeted, ambiguity-driven probing queries for disambiguation and SQL repair. This method achieves significant improvements in execution accuracy, particularly for ambiguous questions, outperforming existing state-of-the-art baselines.

nl2sqlambiguity-resolutionsynthetic-datasql-generationllm-researchdatabase-interfaces
“SOMA-SQL autonomously resolves ambiguity in NL-to-SQL translation without human intervention.”
paper / wesroth / Jun 14

SPARC: Reliable and Scalable Spatial Annotations for Robot Demonstrations

SPARC is a risk-aware framework for generating structured spatial annotations from robot demonstrations, complete with reliability scores. Unlike existing methods that suffer from poorly calibrated confidence and data loss, SPARC utilizes the inherent spatio-temporal structure of robot tasks to produce a reliability signal. This approach significantly reduces noisy labels, retains more useful samples, and leads to state-of-the-art performance on object-grounding and pointing tasks, even outperforming baselines in challenging real-world scenarios.

roboticsspatial-annotationsrobot-demonstrationsreliability-calibrationcomputer-visionembodied-aiarxiv-paper
“SPARC generates structured spatial annotations for robot demonstrations with an associated reliability score.”
paper / wesroth / Apr 27

Variational Ansatz Bias Causes Discrepancies in Hubbard Model Ground State; Systematic Improvement Reveals Coexisting Superconducting and Stripe Orders

In the 2D Hubbard model at finite doping, different Transformer backflow fermionic wave functions (Slater determinant, particle-hole, Pfaffian) achieve near-degenerate state-of-the-art energies but initially converge to qualitatively distinct spin, charge, and pairing correlations due to ansatz bias. Upon symmetry restoration and variance reduction for improved accuracy, all ansatze converge to the same ground state featuring coexisting superconducting and stripe orders. This demonstrates that variational energy minimization alone cannot resolve competing phases, necessitating tracking of correlation functions during systematic wave function improvement.

hubbard-modelvariational-wavefunctionstransformer-backflowsuperconductivitystripe-ordersstrongly-correlated-electronsquantum-many-body
“Different variational ansatze in the 2D Hubbard model at finite doping converge to states with qualitatively different spin, charge, and pairing correlations despite nearly degenerate energies”
paper / wesroth / Apr 23

JTPRO Co-Optimizes Instructions and Tool Schemas to Boost LLM Agent Reliability in Large Toolsets

JTPRO iteratively refines global instructions and per-tool schemas via rollout-driven reflection to address tool mis-selection and slot-filling errors in LLM agents with large, domain-specific tool inventories. It preserves tool-local cues for disambiguation while evaluating on Tool Selection Accuracy (TSA), Slot Filling Accuracy (SFA), and Overall Success Rate (OSR). JTPRO surpasses CoT agents and GEPA by 5-20% relative OSR; ablations confirm joint optimization outperforms isolated tuning.

llm-agentstool-callingprompt-optimizationreflective-optimizationtool-selectionlarge-language-models
“LLM agents struggle with tool mis-selection and incorrect slot/value instantiation due to generic prompts and underspecified tool schemas”
paper / wesroth / Apr 23

MARCO Achieves SOTA Semantic Correspondence with Superior Generalization in 3x Smaller, 10x Faster Model

MARCO builds on DINOv2 with a novel training framework combining coarse-to-fine objectives for precise localization and self-distillation to expand sparse keypoint supervision into dense, semantically coherent correspondences. This addresses poor generalization of prior dual-encoder diffusion models to unseen keypoints and categories. MARCO sets new state-of-the-art on SPair-71k, AP-10K, and PF-PASCAL benchmarks, with amplified gains at fine-grained thresholds (+8.9 PCK@0.01), strongest improvements on unseen data (+5.1 SPair-U, +4.7 MP-100), while being 3x smaller and 10x faster than diffusion baselines.

semantic-correspondencecomputer-visiondino-v2generalizable-correspondenceself-distillationcoarse-to-fine
“MARCO sets a new state-of-the-art on SPair-71k, AP-10K, and PF-PASCAL benchmarks”
paper / wesroth / Apr 23

T16 Pipeline Doubles TESS Exoplanet Candidates with 10,000 New Detections from Cycle 1 FFIs Down to 16th Magnitude

The T16 project processed 83.7 million TESS Cycle 1 FFI light curves down to T=16 mag using uniform detrending and systematics correction, enabling a semi-automated ML-assisted transit search that identified 11,554 planet candidates with periods 0.5-27 days. This yielded 10,091 new candidates, including 411 single-transit events, more than doubling the prior TESS candidate count, with emphasis on faint stars where occurrence rates predict abundant planets. Pipeline validation confirmed a new hot Jupiter around metal-poor thick-disk star TIC 183374187 via Magellan/PFS radial velocities.

exoplanet-detectiontess-missiontransit-searchplanet-candidateshot-jupitermachine-learning-astronomyastro-ph-ep
“T16 processed 83,717,159 TESS Cycle 1 FFI light curves down to T=16 mag”
paper / wesroth / Apr 23

HST/STIS Observations Disprove Localized H2O Aurora on Europa, Confirm Global H Exosphere

Reanalysis of HST/STIS Lyα observations of Europa from 1999 and 2012-2020 detects a global atomic hydrogen exosphere at all epochs, with no evidence of localized H2O auroral emissions, including in prior images interpreted as south pole outgassing. The exosphere shows velocity-dependent attenuation from Earth's H absorption, yielding a temperature of ~1000 K (upper limit 5100 K). For 2014-2015, vertical H column density is 1.4e12 cm^-2 and source rate 1.1e27 s^-1. Discrepancies with earlier H2O claims stem from incorrect disk positioning and omission of exospheric signal.

europalyman-alphahst-stish-exosphereplanetary-auroraastro-ph-epspace-physics
“No localized emission enhancements, such as H2O aurora, detected in any STIS observations including the 1999 south pole image.”
paper / wesroth / Apr 18

COMPOSITE-STEM: A New Benchmark for Evaluating AI in Scientific Discovery

COMPOSITE-STEM is a new benchmark designed to evaluate AI agents' reasoning capabilities in accelerating scientific discovery. It comprises 70 expert-written tasks across physics, biology, chemistry, and mathematics, curated by doctoral-level researchers. The benchmark utilizes a hybrid grading approach combining exact-match grading and criterion-based rubrics with an LLM-as-a-jury protocol, enabling flexible assessment of scientifically meaningful outputs.

ai-agentsscientific-discoveryllm-evaluationbenchmarkingmultimodal-modelsfrontier-modelsreproducibility
“Existing AI benchmarks focusing on scientific reasoning are becoming saturated and only measure performance on constrained outputs.”
paper / wesroth / Apr 13

MT-OSC: Background Chat History Condensation Cuts LLM Multi-Turn Token Costs by 72%

LLMs degrade in performance across multi-turn conversations when full chat history is naively appended to prompts, exhausting context windows and inflating latency and cost. MT-OSC (One-off Sequential Condensation) addresses this by running a background Condenser Agent — combining a few-shot inference-based Condenser and a lightweight Decider — that selectively compresses chat history without interrupting the user experience. Evaluated across 13 state-of-the-art LLMs and multiple multi-turn benchmarks, the framework reduces token counts by up to 72% in 10-turn dialogues while maintaining or improving accuracy and demonstrating robustness to distractors and irrelevant turns.

llm-infrastructuremulti-turn-conversationcontext-windowprompt-compressionconversational-aitoken-efficiencyllm-optimization
“MT-OSC reduces token counts by up to 72% in 10-turn dialogues compared to naive full-history appending.”
Older entries →