pith. machine review for the scientific record. sign in

arxiv: 2605.05573 · v1 · submitted 2026-05-07 · 🌌 astro-ph.IM · cs.AI

Recognition: unknown

AstroAlertBench: Evaluating the Accuracy, Reasoning, and Honesty of Multimodal LLMs in Astronomical Classification

Alice Zou, Ashish Mahabal, Claire Chen, Facundo Perez Paolino, Jiabao Sean Xiao, Luke Handley, Matthew Graham, Ricky Nilsson, Shuze Daniel Liu, Theophile Jegou Du Laz

Authors on Pith no claims yet

Pith reviewed 2026-05-08 05:18 UTC · model grok-4.3

classification 🌌 astro-ph.IM cs.AI
keywords multimodal LLMsastronomical transientsbenchmark evaluationmodel honestyscientific classificationZTF alertsreasoning assessment
0
0 comments X

The pith

Multimodal LLMs show a disconnect between classification accuracy and honest self-evaluation on astronomical transient alerts

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper presents a benchmark that tests how well multimodal large language models can review and classify real astronomical events from image and metadata inputs. It structures the test around three linked steps: confirming the provided data, applying scientific logic to interpret it, and assigning the event to one of five categories. When thirteen different models are run on fifteen hundred alerts from a wide-field survey, the results show that models reaching high classification accuracy often fail to correctly judge the strength of their own reasoning. This separation matters because an assistant used in live observatory workflows must signal when its output rests on shaky grounds rather than simply guessing right most of the time. The work also begins a human review loop so experts can later refine the benchmark at larger scale.

Core claim

AstroAlertBench evaluates thirteen frontier multimodal LLMs on a pilot set of 1,500 real ZTF alerts by requiring each model to complete metadata grounding, scientific reasoning, and hierarchical classification; the evaluation finds that high accuracy scores frequently do not coincide with the models' ability to self-assess the validity of their reasoning chains.

What carries the argument

The three-stage logical chain of metadata grounding, scientific reasoning, and hierarchical classification, together with an independent honesty metric that scores how accurately a model evaluates its own reasoning quality.

If this is right

  • High-accuracy models cannot be treated as automatically reliable for live astronomical data review because they may still misjudge their own reasoning.
  • The benchmark supplies a repeatable protocol for measuring both performance and self-awareness in scientific classification tasks.
  • A human-in-the-loop review stage allows experts to annotate model outputs and thereby improve future iterations of the benchmark.
  • Separate tracking of accuracy and honesty creates a concrete target for training methods that aim to make model self-assessment match actual correctness.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Surveys with different instruments or wavelengths could be added to test whether the accuracy-honesty gap persists across data types.
  • Models that perform well on both metrics could be used as an initial filter to surface only the most uncertain alerts for human astronomers.
  • The same staged evaluation plus honesty check might be adapted to other image-plus-text scientific domains such as medical or materials imaging.

Load-bearing premise

The 1,500-alert pilot sample from one survey stands in for the full diversity of astronomical data and that the honesty metric validly tracks real-world reliability of LLM outputs.

What would settle it

A larger and more diverse set of alerts in which every high-accuracy model also scores high on the honesty metric would show that the observed misalignment between accuracy and honesty is not a general feature.

Figures

Figures reproduced from arXiv: 2605.05573 by Alice Zou, Ashish Mahabal, Claire Chen, Facundo Perez Paolino, Jiabao Sean Xiao, Luke Handley, Matthew Graham, Ricky Nilsson, Shuze Daniel Liu, Theophile Jegou Du Laz.

Figure 1
Figure 1. Figure 1: Overview of the AstroAlertBench pipeline. Starting from a first-detection alert, we construct view at source ↗
Figure 2
Figure 2. Figure 2: AstroAlertBench taxonomy. The inner ring shows the five benchmark classes with equal weight: bogus, asteroid, supernova, AGN, and variable star. The outer ring illustrates representative subtype categories; this is not an exhaustive classification tree. Wedge sizes are schematic and do not encode subtype counts. This section details the construction and composition of AstroAlertBench, focusing on the trans… view at source ↗
Figure 3
Figure 3. Figure 3: Hierarchical classification cascade. The five benchmark categories are mapped through three sequential decision stages: artifact rejection (Stage-1), physical origin (Stage-2), and astrophys￾ical subclass assignment (Stage-3). (Part B) and Decision (Part C) stages to scientific logic rather than cross-modal serialization or input-parsing errors. The full per-field accuracy results for each model are report… view at source ↗
Figure 4
Figure 4. Figure 4: Per-class accuracy heatmap across 13 runs. Rows are the 13 evaluated models, sorted top-to-bottom by absolute 5-class accuracy; columns are the five benchmark classes (supernova, AGN, variable star, asteroid, bogus). Five-class classification performance. As shown in view at source ↗
Figure 5
Figure 5. Figure 5: Inverse correlation between self-assessment and accuracy. Each point represents one of the 13 evaluated runs (n = 1, 500). The dashed line indicates an OLS fit with a slope of −31.2±10.0 percentage points of accuracy per unit on the 0–5 self-reasoning scale. Models with higher empirical accuracy tend to be more modest in their self-evaluations view at source ↗
Figure 6
Figure 6. Figure 6: Second-rollout retry on the low-confidence subset. Comparison of first-pass and second￾pass accuracy over the same 35 alerts per model, drawn from each model’s original low-confidence alert pool. Gains (∆) are reported in absolute percentage points view at source ↗
Figure 7
Figure 7. Figure 7: End-to-end 5-class accuracy ranked across the 13 evaluated configurations. Error bars denote 1σ binomial standard errors. Claude Opus 4.7 think establishes the current bench￾mark ceiling at 60.60 ± 1.26%, opening a +9.53 ± 1.80 pp gap over the second-ranked GPT-5.4 high-think. Stage-wise cascade view at source ↗
Figure 8
Figure 8. Figure 8: Stage-wise cascade accuracy across the three Part C stages. Four grouped bars per run: Stage-1 (real-vs-artifact), Stage-2 (solar-system-vs-astrophysical), Stage-3 (astrophysical subclass), and Stage-3 conditional (restricted to rows where Stages 1 and 2 are both correct). Stage-1 is essentially solved across all model scales (77.15–87.07%); Stage-2 separates closed-source from open-source; Stage-3 conditi… view at source ↗
Figure 9
Figure 9. Figure 9: Predicted-class distribution among the 300 true-AGN alerts (top four models). Across all four models, 91–93% of real AGN alerts are predicted as variable_star; AGN recall ranges from 5.3% to 7.3%. The AGN-collapse pattern holds independently of model family, scale, and reasoning mode. D Reasoning-mode analysis: think vs. no-think To quantify the impact of reasoning on the end-to-end 5-class accuracy report… view at source ↗
Figure 10
Figure 10. Figure 10: Reasoning-mode dial across five model families (n = 1, 500 per run). For each family, blue bars show absolute 5-class accuracy with thinking enabled and orange bars show the same with thinking disabled; ∆ annotations report the gap and the two-sample z-statistic. The think-win margin grows with backbone capacity: Claude Opus 4.7 shows the largest gap (+11.73 pp), Qwen3.5-397B-A17B and GPT-5.4 follow with … view at source ↗
Figure 11
Figure 11. Figure 11: Per-run Pearson correlation between self-confidence and correctness, sorted descend￾ing. Larger values mean the model’s self-confidence rank-orders its right answers from its wrong ones; near-zero values mean the confidence dial is uninformative. Error bars are 1σ standard error. E.3 Per-bin accuracy at the τ = 4 threshold view at source ↗
Figure 12
Figure 12. Figure 12: 5-class accuracy on rows the model labels high-confidence (self-mean ≥ 4) versus low￾confidence (self-mean < 4). Claude Opus 4.7 nothink has both a populated low-confidence bin and a meaningful accuracy gap; GPT-5.4 high-think has a strong gap on a much narrower low-confidence population. F.1 Methodology For each closed-source run we constructed a low-confidence subset by selecting first-pass benchmark ro… view at source ↗
Figure 13
Figure 13. Figure 13: Calibration gap vs. Pearson r across 13 runs. Each marker is one run. The horizontal axis is the calibration gap defined in §E.1; the vertical axis is the per-row Pearson correlation also shown as a sorted bar chart in main-text view at source ↗
Figure 14
Figure 14. Figure 14: Confidence-bin distribution per run. Most runs park more than 90% of predictions in the high-confidence bin; only the two Opus 4.7 runs come anywhere near a balanced split (think 67/33, nothink 67/33). For the high-mass-only runs there is no usable low-confidence subset to calibrate against. Model W → C′ C → W′ GPT-5.4 high-think 6 3 GPT-5.4 no-think 4 3 Gemini 2.5 Flash no-think 0 0 Claude Opus 4.7 think… view at source ↗
Figure 15
Figure 15. Figure 15: Correction, damage, and persistence rates per model. Tall green (correction) plus zero red (damage) on both Claude Opus 4.7 variants; GPT-5.4 high-think trades correction against a visible damage segment. 35 view at source ↗
Figure 16
Figure 16. Figure 16: Normalized 2×2 paired-outcome composition. Stacked bars sum to 100% within each model. Opus 4.7 runs devote a large slice to W → C ′ with no visible C → W′ band; Gemini 2.5 Flash no-think is dominated by W → W′ plus the rows that stay correct view at source ↗
Figure 17
Figure 17. Figure 17: McNemar two-sided p-values on − log10 scale. The test evaluates the asymmetry of accuracy changes using only discordant pairs (W → C ′ vs. C → W); the dashed line marks p = 0.05. Both Claude Opus 4.7 variants sit above this threshold, indicating statistically significant improvements. GPT-5.4 configurations remain below the line, while Gemini 2.5 Flash has zero discordant pairs, rendering the test degener… view at source ↗
Figure 18
Figure 18. Figure 18: Science–Reference–Difference montage for view at source ↗
Figure 19
Figure 19. Figure 19: Science–Reference–Difference montage for view at source ↗
Figure 20
Figure 20. Figure 20: Three human-accuracy definitions on the 15-alert subset. Effective (30.67% ± 5.32, abstention scored as incorrect, n = 75), selective (34.33% ± 5.80, abstention excluded, n = 67), and ensemble majority (26.67% ± 11.42, strict 3/5 non-DK majority matches gold, n = 15). Error bars are 1σ binomial standard errors on each metric’s denominator. G.3 Same-15-alert head-to-head with models To enable an apples-to-… view at source ↗
Figure 21
Figure 21. Figure 21: Frontier models versus human references on the 15-alert subset. Each horizontal bar reports a model run’s absolute 5-class accuracy on the same 15 alerts the human cohort evaluated (error bars: ±1σ binomial). Three dashed vertical lines mark the strict ensemble majority (26.67%, n = 15), the effective human rate (30.67%, n = 75), and the best individual expert (46.67%, n = 15). Claude Opus 4.7 think excee… view at source ↗
Figure 22
Figure 22. Figure 22: Q2 vs. Q3 highlight composition on ZTF26aargnnp. Stacked bars give the share of expert-marked green (“fully correct”), yellow (“partially correct”), red (“flawed or hallu￾cinatory”), and unhighlighted characters in the model’s leading_interpretation (Q2) and alternative_analysis (Q3) fields, pooled over the 13 audited model blocks. The green share rises from 23.19% in Q2 to 44.48% in Q3 while the red shar… view at source ↗
Figure 23
Figure 23. Figure 23: Self-score vs. human mean per model on ZTF26aargnnp. Paired bars show the blended human grade (five raters) and the model’s mean Part B self-score (leading_interpretation + alternative_analysis) for each of the 13 audited runs. 7One model (Qwen3.5-397B-A17B nothink) is excluded from correlation and calibration metrics because it did not output Part B self-score fields in this example; however, its reasoni… view at source ↗
Figure 24
Figure 24. Figure 24: Per-LLM calibration on ZTF26aargnnp. Mean Part B self-score (vertical axis) against the blended human mean (horizontal axis) for the 12 models with complete Part B fields. The dashed y = x line marks perfect calibration; marker fill encodes Part C correctness (green = correct class, red = incorrect class, gray = unparsed) view at source ↗
Figure 25
Figure 25. Figure 25: Self-rating bias on ZTF26aargnnp. Mean Part B self-score minus the blended human grade, sorted. Positive values (red bars) indicate the model rates its own reasoning above the human consensus; negative values (blue bars) indicate the model is stricter than the experts. 49 view at source ↗
Figure 26
Figure 26. Figure 26: Science–Reference–Difference montage for view at source ↗
read the original abstract

Modern astronomical observatories generate a massive volume of multimodal data, creating a critical bottleneck for expert human review. While multimodal large language models (LLMs) have shown promise in interpreting complex visual and textual inputs, their ability to perform specialized scientific classification while providing interpretable reasoning remains understudied. We introduce AstroAlertBench, a comprehensive multimodal benchmark designed to evaluate LLM performance in astronomical event review along a three-stage logical chain: metadata grounding, scientific reasoning, and hierarchical classification over five categories. We use a pilot sample of 1,500 real-world alerts from the Zwicky Transient Facility (ZTF), a wide-field survey that scans the northern sky to detect transient astronomical events. On this dataset, we benchmark 13 frontier closed-source and open-weight LLMs that support visual input. Our results reveal that high accuracy does not always align with model ``honesty,'' defined as the ability to self-evaluate its reasoning, which affects its reliability as a real-world assistant. We further initialize a human-in-the-loop evaluation protocol as a precursor to future community-scale participation. Together, AstroAlertBench provides a framework for developing calibrated and interpretable astronomical assistants.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper introduces AstroAlertBench, a multimodal benchmark for evaluating 13 frontier LLMs on astronomical transient classification using a pilot sample of 1,500 real ZTF alerts. Models are assessed along a three-stage logical chain (metadata grounding, scientific reasoning, hierarchical classification into five categories). The central result is that high accuracy does not always align with model 'honesty' (self-evaluation of reasoning), which the authors argue affects reliability as real-world assistants; a human-in-the-loop protocol is also initialized as a precursor to community evaluation.

Significance. If the results hold, the work is significant for astroinformatics by supplying a real-data framework to probe LLM limitations beyond raw accuracy in a high-volume scientific domain. Use of authentic ZTF alerts and the start of a human-in-the-loop protocol are strengths that support reproducible benchmarking and future community-scale participation. The focus on reasoning honesty directly addresses a practical barrier to deploying multimodal models as reliable astronomical assistants.

major comments (2)
  1. [§3 (Benchmark Design)] §3 (Benchmark Design): The honesty metric (self-evaluation of reasoning) is introduced as a key indicator that misalignment with accuracy reduces real-world reliability, yet the manuscript shows no independent validation or correlation with external criteria such as expert agreement on classifications, uncertainty calibration, or performance in a deployed human-in-the-loop setting. This is load-bearing for the central claim.
  2. [§4 (Results)] §4 (Results): The analysis of the 1,500-alert pilot sample reports no statistical significance tests, error analysis, confidence intervals on accuracy-honesty scores, or assessment of selection biases in the ZTF alerts; these omissions leave the observed misalignment only moderately supported as a general finding.
minor comments (2)
  1. [Abstract] The abstract states that five classification categories are used but does not enumerate them, which would improve immediate clarity for readers.
  2. [§2 (Related Work)] Additional references to prior LLM evaluation benchmarks in scientific domains would better situate AstroAlertBench within the existing literature.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback on our manuscript. The comments highlight important aspects of validation and statistical rigor that we will address in revision. We respond point-by-point to the major comments below.

read point-by-point responses

  1. Referee: [§3 (Benchmark Design)] The honesty metric (self-evaluation of reasoning) is introduced as a key indicator that misalignment with accuracy reduces real-world reliability, yet the manuscript shows no independent validation or correlation with external criteria such as expert agreement on classifications, uncertainty calibration, or performance in a deployed human-in-the-loop setting. This is load-bearing for the central claim.

    Authors: We acknowledge that the honesty metric, based on model self-evaluation of reasoning quality, lacks direct correlation with external validators such as expert agreement or calibrated uncertainty in the current pilot results. The manuscript presents the observed accuracy-honesty misalignment as an empirical finding from the 1,500 ZTF alerts and positions the human-in-the-loop protocol as an initial step toward such validation. We agree this aspect requires clearer framing. In revision, we will expand the discussion in §3 and add a limitations subsection to explicitly note the absence of independent validation in this pilot, describe how the human-in-the-loop setup is designed to enable future expert correlations, and outline planned analyses for uncertainty calibration. These changes will qualify the central claim as preliminary while preserving the observed misalignment as a motivating observation. revision: partial

  2. Referee: [§4 (Results)] The analysis of the 1,500-alert pilot sample reports no statistical significance tests, error analysis, confidence intervals on accuracy-honesty scores, or assessment of selection biases in the ZTF alerts; these omissions leave the observed misalignment only moderately supported as a general finding.

    Authors: The 1,500-alert dataset is presented throughout as a pilot sample whose primary purpose is to demonstrate the benchmark framework rather than support broad statistical generalizations. We agree that the results section would benefit from additional quantitative support. In the revised manuscript, we will add bootstrap-derived confidence intervals for accuracy and honesty scores, include basic error analysis with examples of misalignment cases, perform appropriate significance tests (e.g., McNemar or chi-squared tests for paired accuracy-honesty differences where applicable), and discuss potential selection biases in the ZTF alerts (such as magnitude limits, sky coverage, or transient type distributions). These updates will be incorporated into §4 and the supplementary material. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical benchmark grounded in external data

full rationale

The paper is a pure empirical evaluation study that introduces AstroAlertBench as an operational framework for testing LLMs on 1,500 real ZTF alerts. The three-stage logical chain and honesty metric are defined as part of the benchmark protocol and applied to external model outputs and human review; results are reported as direct observations from this data rather than derived from any equations, fitted parameters, or self-referential predictions. No self-citation chains, ansatzes, or uniqueness theorems are invoked as load-bearing steps. The derivation chain is therefore self-contained against external benchmarks and does not reduce to its inputs by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the validity of the defined three-stage evaluation framework and the operational definition of honesty as self-evaluation of reasoning; no free parameters or invented entities are introduced.

axioms (1)
  • domain assumption The three-stage logical chain (metadata grounding, scientific reasoning, and hierarchical classification) captures the essential aspects of astronomical event review.
    Presented in the abstract as the core evaluation structure without additional justification or validation against expert processes.

pith-pipeline@v0.9.0 · 5543 in / 1289 out tokens · 56236 ms · 2026-05-08T05:18:30.325505+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

300 extracted references · 21 canonical work pages · 13 internal anchors

  1. [1]

    2025 , howpublished =

    Tinker , author =. 2025 , howpublished =

    2025

  2. [2]

    2026 , month = feb, howpublished =

    Qwen3.5: Towards Native Multimodal Agents , author =. 2026 , month = feb, howpublished =

    2026

  3. [3]

    Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities

    Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities , author=. arXiv preprint arXiv:2507.06261 , year=

    work page internal anchor Pith review arXiv

  4. [4]

    2026 , month = mar, howpublished =

    GPT-5.4 Thinking System Card , author =. 2026 , month = mar, howpublished =

    2026

  5. [5]

    2026 , month = apr, howpublished =

    Claude Opus 4.7 System Card , author =. 2026 , month = apr, howpublished =

    2026

  6. [6]

    ResearchBench: Benchmarking LLMs in Scientific Discovery via Inspiration-Based Task Decomposition

    Researchbench: Benchmarking llms in scientific discovery via inspiration-based task decomposition , author=. arXiv preprint arXiv:2503.21248 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  7. [7]

    MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts

    Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts , author=. arXiv preprint arXiv:2310.02255 , year=

    work page internal anchor Pith review arXiv

  8. [8]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  9. [9]

    The Zwicky Transient Facility Bright Transient Survey. III. BTSbot: automated identification and follow-up of bright transients with deep learning , author=. The Astrophysical Journal , volume=. 2024 , publisher=

    2024

  10. [10]

    Publications of the Astronomical Society of the Pacific , volume=

    The zwicky transient facility alert distribution system , author=. Publications of the Astronomical Society of the Pacific , volume=. 2019 , publisher=

    2019

  11. [11]

    The Astronomical Journal , volume=

    Alert classification for the ALeRCE broker system: the real-time stamp classifier , author=. The Astronomical Journal , volume=. 2021 , publisher=

    2021

  12. [12]

    Kimi K2.5: Visual Agentic Intelligence

    Kimi K2. 5: Visual Agentic Intelligence , author=. arXiv preprint arXiv:2602.02276 , year=

    work page internal anchor Pith review arXiv

  13. [13]

    International conference on machine learning , pages=

    Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models , author=. International conference on machine learning , pages=. 2023 , organization=

    2023

  14. [14]

    International conference on machine learning , pages=

    Learning transferable visual models from natural language supervision , author=. International conference on machine learning , pages=. 2021 , organization=

    2021

  15. [15]

    Qwen3 Technical Report

    Qwen3 technical report , author=. arXiv preprint arXiv:2505.09388 , year=

    work page internal anchor Pith review arXiv

  16. [16]

    Qwen Technical Report

    Qwen technical report , author=. arXiv preprint arXiv:2309.16609 , year=

    work page internal anchor Pith review arXiv

  17. [17]

    arXiv preprint arXiv:2304.05376 , year=

    Chemcrow: Augmenting large-language models with chemistry tools , author=. arXiv preprint arXiv:2304.05376 , year=

    work page arXiv

  18. [18]

    arXiv preprint arXiv:2304.05332 , year=

    Emergent autonomous scientific research capabilities of large language models , author=. arXiv preprint arXiv:2304.05332 , year=

    work page arXiv

  19. [19]

    SC24-W: Workshops of the International Conference for High Performance Computing, Networking, Storage and Analysis , pages=

    AstroMLab 2: AstroLLaMA-2-70B model and benchmarking specialised LLMs for astronomy , author=. SC24-W: Workshops of the International Conference for High Performance Computing, Networking, Storage and Analysis , pages=. 2024 , organization=

    2024

  20. [20]

    Nature Astronomy , volume=

    Textual interpretation of transient image classifications from large language models , author=. Nature Astronomy , volume=. 2025 , publisher=

    2025

  21. [21]

    Publications of the Astronomical Society of the Pacific , volume=

    Machine learning for the zwicky transient facility , author=. Publications of the Astronomical Society of the Pacific , volume=. 2019 , publisher=

    2019

  22. [22]

    The Astronomical Journal , volume=

    Alert classification for the ALeRCE broker system: The light curve classifier , author=. The Astronomical Journal , volume=. 2021 , publisher=

    2021

  23. [23]

    The Astronomical Journal , volume=

    The automatic learning for the rapid classification of events (ALeRCE) alert broker , author=. The Astronomical Journal , volume=. 2021 , publisher=

    2021

  24. [24]

    Publications of the Astronomical Society of the Pacific , volume=

    The zwicky transient facility: science objectives , author=. Publications of the Astronomical Society of the Pacific , volume=. 2019 , publisher=

    2019

  25. [25]

    Monthly Notices of the Royal Astronomical Society , volume=

    Real-bogus classification for the Zwicky Transient Facility using deep learning , author=. Monthly Notices of the Royal Astronomical Society , volume=. 2019 , publisher=

    2019

  26. [26]

    Publications of the Astronomical Society of the Pacific , volume=

    The zwicky transient facility: Data processing, products, and archive , author=. Publications of the Astronomical Society of the Pacific , volume=. 2019 , publisher=

    2019

  27. [27]

    Publications of the Astronomical Society of the Pacific , volume=

    The Zwicky Transient Facility: system overview, performance, and first results , author=. Publications of the Astronomical Society of the Pacific , volume=. 2019 , publisher=

    2019

  28. [28]

    The twelfth international conference on learning representations , year=

    Let's verify step by step , author=. The twelfth international conference on learning representations , year=

  29. [29]

    International conference on machine learning , pages=

    On calibration of modern neural networks , author=. International conference on machine learning , pages=. 2017 , organization=

    2017

  30. [30]

    Language Models (Mostly) Know What They Know

    Language models (mostly) know what they know , author=. arXiv preprint arXiv:2207.05221 , year=

    work page internal anchor Pith review arXiv

  31. [31]

    Advances in Neural Information Processing Systems , volume=

    Can llms solve molecule puzzles? a multimodal benchmark for molecular structure elucidation , author=. Advances in Neural Information Processing Systems , volume=

  32. [32]

    Nature , volume=

    Mathematical discoveries from program search with large language models , author=. Nature , volume=. 2024 , publisher=

    2024

  33. [33]

    Galactica: A Large Language Model for Science

    Galactica: A large language model for science , author=. arXiv preprint arXiv:2211.09085 , year=

    work page internal anchor Pith review arXiv

  34. [34]

    Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context

    Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context , author=. arXiv preprint arXiv:2403.05530 , year=

    work page internal anchor Pith review arXiv

  35. [35]

    Advances in neural information processing systems , volume=

    Flamingo: a visual language model for few-shot learning , author=. Advances in neural information processing systems , volume=

  36. [36]

    Advances in neural information processing systems , volume=

    Visual instruction tuning , author=. Advances in neural information processing systems , volume=

  37. [37]

    The Llama 3 Herd of Models

    The llama 3 herd of models , author=. arXiv preprint arXiv:2407.21783 , year=

    work page internal anchor Pith review arXiv

  38. [38]

    Advances in neural information processing systems , volume=

    Language models are few-shot learners , author=. Advances in neural information processing systems , volume=

  39. [39]

    Games and economic behavior , volume=

    Potential games , author=. Games and economic behavior , volume=. 1996 , publisher=

    1996

  40. [40]

    Concurrent submission to NeurIPS 2026 , year=

    Pessimism-Free Offline Learning in General-Sum Games via KL Regularization , author=. Concurrent submission to NeurIPS 2026 , year=

    2026

  41. [41]

    International Conference on Machine Learning , pages=

    Independent policy gradient for large-scale markov potential games: Sharper rates, function approximation, and game-agnostic convergence , author=. International Conference on Machine Learning , pages=. 2022 , organization=

    2022

  42. [42]

    IEEE Control Systems Letters , volume=

    Learning Nash in constrained Markov games with an -potential , author=. IEEE Control Systems Letters , volume=. 2024 , publisher=

    2024

  43. [43]

    2025 IEEE 64th Conference on Decision and Control (CDC) , pages=

    Markov potential game construction and multi-agent reinforcement learning with applications to autonomous driving , author=. 2025 IEEE 64th Conference on Decision and Control (CDC) , pages=. 2025 , organization=

    2025

  44. [44]

    arXiv preprint arXiv:2106.01969 , year=

    Global convergence of multi-agent policy gradient in markov potential games , author=. arXiv preprint arXiv:2106.01969 , year=

    work page arXiv

  45. [45]

    IEEE Transactions on Automatic Control , year=

    Markov -Potential Games , author=. IEEE Transactions on Automatic Control , year=

  46. [46]

    2025 , publisher=

    An -Potential Game Framework for Non-Cooperative Dynamic Games: Theory and Algorithms , author=. 2025 , publisher=

    2025

  47. [47]

    SIAM Journal on Control and Optimization , volume=

    An-Potential Game Framework for-Player Dynamic Games , author=. SIAM Journal on Control and Optimization , volume=. 2025 , publisher=

    2025

  48. [48]

    arXiv preprint arXiv:2305.12553 , year=

    Markov -Potential Games , author=. arXiv preprint arXiv:2305.12553 , year=

    work page arXiv

  49. [49]

    arXiv preprint arXiv:2310.06243 , year=

    Sample-efficient multi-agent rl: An optimization perspective , author=. arXiv preprint arXiv:2310.06243 , year=

    work page arXiv

  50. [50]

    Corruption-robust Offline Multi-agent Reinforcement Learning From Human Feedback

    Corruption-robust Offline Multi-agent Reinforcement Learning From Human Feedback , author=. arXiv preprint arXiv:2603.28281 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  51. [51]

    Journal of the ACM (JACM) , volume=

    Settling the complexity of computing two-player Nash equilibria , author=. Journal of the ACM (JACM) , volume=. 2009 , publisher=

    2009

  52. [52]

    Journal of Computer and system Sciences , volume=

    On the complexity of the parity argument and other inefficient proofs of existence , author=. Journal of Computer and system Sciences , volume=. 1994 , publisher=

    1994

  53. [53]

    Communications of the ACM , volume=

    The complexity of computing a Nash equilibrium , author=. Communications of the ACM , volume=. 2009 , publisher=

    2009

  54. [54]

    International conference on machine learning , pages=

    Pessimistic q-learning for offline reinforcement learning: Towards optimal sample complexity , author=. International conference on machine learning , pages=. 2022 , organization=

    2022

  55. [55]

    Operations research , volume=

    Model-based reinforcement learning for offline zero-sum Markov games , author=. Operations research , volume=. 2024 , publisher=

    2024

  56. [56]

    Beyond Pessimism: Offline Learning in KL-regularized Games

    Beyond Pessimism: Offline Learning in KL-regularized Games , author=. arXiv preprint arXiv:2604.06738 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  57. [57]

    Advances in neural information processing systems , volume=

    Direct preference optimization: Your language model is secretly a reward model , author=. Advances in neural information processing systems , volume=

  58. [58]

    2016 , publisher=

    Twenty lectures on algorithmic game theory , author=. 2016 , publisher=

    2016

  59. [59]

    2006 , publisher=

    Prediction, learning, and games , author=. 2006 , publisher=

    2006

  60. [60]

    Theory of computing , volume=

    The multiplicative weights update method: a meta-algorithm and applications , author=. Theory of computing , volume=. 2012 , publisher=

    2012

  61. [61]

    Machine Intelligence Research , volume=

    Offline pre-trained multi-agent decision transformer , author=. Machine Intelligence Research , volume=. 2023 , publisher=

    2023

  62. [62]

    arXiv preprint arXiv:2102.04402 , year=

    Contrasting centralized and decentralized critics in multi-agent reinforcement learning , author=. arXiv preprint arXiv:2102.04402 , year=

    work page arXiv

  63. [63]

    International Journal of Group Decision and Negotiation , volume=

    Automated negotiation: prospects, methods and challenges , author=. International Journal of Group Decision and Negotiation , volume=

  64. [64]

    2001 , publisher=

    Strategic negotiation in multiagent environments , author=. 2001 , publisher=

    2001

  65. [65]

    Communications of the ACM , volume=

    Algorithmic game theory , author=. Communications of the ACM , volume=. 2010 , publisher=

    2010

  66. [66]

    Econometrica: Journal of the Econometric Society , pages=

    A theory of auctions and competitive bidding , author=. Econometrica: Journal of the Econometric Society , pages=. 1982 , publisher=

    1982

  67. [67]

    Games and Economic Behavior , volume=

    On the value of information in a strategic conflict , author=. Games and Economic Behavior , volume=. 1990 , publisher=

    1990

  68. [68]

    1995 , publisher=

    Repeated games with incomplete information , author=. 1995 , publisher=

    1995

  69. [69]

    Bayesian

    Games with incomplete information played by “Bayesian” players, I--III Part I. The basic model , author=. Management science , volume=. 1967 , publisher=

    1967

  70. [70]

    Mathematics of operations research , volume=

    Optimal auction design , author=. Mathematics of operations research , volume=. 1981 , publisher=

    1981

  71. [71]

    The Journal of finance , volume=

    Counterspeculation, auctions, and competitive sealed tenders , author=. The Journal of finance , volume=. 1961 , publisher=

    1961

  72. [72]

    Proceedings of the national academy of sciences , volume=

    Stochastic games , author=. Proceedings of the national academy of sciences , volume=. 1953 , publisher=

    1953

  73. [73]

    Behavior Regularized Offline Reinforcement Learning

    Behavior regularized offline reinforcement learning , author=. arXiv preprint arXiv:1911.11361 , year=

    work page internal anchor Pith review arXiv 1911

  74. [74]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    Adaptive trust region policy optimization: Global convergence and faster rates for regularized mdps , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

  75. [75]

    Foundations and Trends

    Online learning and online convex optimization , author=. Foundations and Trends. 2025 , publisher=

    2025

  76. [76]

    2000 , publisher=

    Empirical Processes in M-estimation , author=. 2000 , publisher=

    2000

  77. [77]

    IEEE transactions on information theory , volume=

    Minimum complexity density estimation , author=. IEEE transactions on information theory , volume=. 2002 , publisher=

    2002

  78. [78]

    , author=

    On general minimax theorems. , author=

  79. [79]

    Dynamic Games and Applications , volume=

    Upper and lower values in zero-sum stochastic games with asymmetric information , author=. Dynamic Games and Applications , volume=. 2021 , publisher=

    2021

  80. [80]

    Games and Economic Behavior , volume=

    Adaptive game playing using multiplicative weights , author=. Games and Economic Behavior , volume=. 1999 , publisher=

    1999

Showing first 80 references.