pith. machine review for the scientific record. sign in

arxiv: 2108.07258 · v3 · submitted 2021-08-16 · 💻 cs.LG · cs.AI· cs.CY

Recognition: 1 theorem link

· Lean Theorem

On the Opportunities and Risks of Foundation Models

Aditi Raghunathan, Alex Tamkin, Ali Malik, Allen Nie, Ananya Kumar, Andy Shih, Annie Chen, Antoine Bosselut, Armin W. Thomas, Avanika Narayan, Ben Newman, Bohan Wu, Camilo Ruiz, Chelsea Finn, Chris Donahue, Chris Piech, Christopher D. Manning, Christopher Potts, Christopher R\'e, Dallas Card, Daniel E. Ho, Dan Jurafsky, Deepak Narayanan, Dora Demszky, Dorsa Sadigh, Drew A. Hudson, Ehsan Adeli, Emma Brunskill, Eric Mitchell, Erik Brynjolfsson, Esin Durmus, Eva Portelance, Faisal Ladhak, Fereshte Khani, Florian Tram\`er, Frieda Rong, Geoff Keeling, Giray Ogut, Hamed Nilforoshan, Hongyu Ren, Isabelle Levent, Isabel Papadimitriou, Jack Ryan, Jared Quincy Davis, Jeannette Bohg, Jenny Hong, Jiajun Wu, Jiaxuan You, Jing Huang, John Etchemendy, John Hewitt, Joon Sung Park, Juan Carlos Niebles, Julian Nyarko, Jure Leskovec, Kaitlyn Zhou, Karan Goel, Kathleen Creel, Kawin Ethayarajh, Keshav Santhanam, Krishnan Srinivasan, Kyle Hsu, Laurel Orr, Lauren Gillespie, Li Fei-Fei, Lucia Zheng, Mark Krass, Matei Zaharia, Michael S. Bernstein, Michael Zhang, Michihiro Yasunaga, Mina Lee, Moussa Doumbouya, Neel Guha, Niladri Chatterji, Noah Goodman, Omar Khattab, Pang Wei Koh, Percy Liang, Peter Henderson, Pratyusha Kalluri, Ranjay Krishna, Rishi Bommasani, Rob Reich, Rodrigo Castellon, Rohan Taori, Rohith Kuditipudi, Rose E. Wang, Russ Altman, Saahil Jain, Sang Michael Xie, Shelby Grossman, Shiori Sagawa, Shyamal Buch, Siddharth Karamcheti, Simran Arora, Stefano Ermon, Suraj Nair, Suvir Mirchandani, Sydney von Arx, Tatsunori Hashimoto, Tengyu Ma, Thomas Icard, Tianyi Zhang, Tony Lee, Trevor Gale, William Wang, Xiang Lisa Li, Xikun Zhang, Xuechen Li, Yuhuai Wu, Yuhui Zhang, Yusuf Roohani, Zanele Munyikwa

Pith reviewed 2026-05-10 16:10 UTC · model grok-4.3

classification 💻 cs.LG cs.AIcs.CY
keywords foundation modelsemergent capabilitieshomogenizationdeep learningtransfer learningAI riskssocietal impact
0
0 comments X

The pith

Foundation models trained on broad data at scale develop emergent capabilities that incentivize their use across many tasks while passing any defects to all downstream adaptations.

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

The paper defines foundation models as large systems like BERT and GPT-3 that are pretrained on wide-ranging data and then adapted to numerous specific tasks. It claims that the sheer scale of this pretraining produces new abilities in areas such as language, vision, and reasoning that smaller models do not show. This effectiveness pushes many applications to rely on the same base models, creating homogenization that amplifies both benefits and problems. The authors stress that current understanding of these models remains limited, especially regarding failure modes and true limits, and argue that addressing this requires coordinated work across technical and social fields.

Core claim

Foundation models are models trained on broad data at scale that prove adaptable to a wide range of downstream tasks. Their scale produces emergent capabilities beyond those of smaller models, while their versatility across tasks drives homogenization of the AI ecosystem. As a result, any flaws in the foundation model are inherited by every adapted system built on it. Despite impending widespread use, there is still no clear account of how these models function, when they break, or what they can ultimately do.

What carries the argument

Foundation models, defined as models pretrained on broad data at scale and then adapted for downstream tasks, which carry the argument by linking scale to emergence and by showing how shared bases transmit strengths and weaknesses to all uses.

If this is right

  • Capabilities in language understanding, image generation, and robotics improve as models grow larger.
  • Applications in law, healthcare, and education gain efficiency from shared bases but must account for inherited limitations.
  • Societal issues such as bias, misuse, and environmental costs become concentrated rather than distributed.
  • Technical work on evaluation, security, and theory must address the full model rather than isolated adaptations.
  • Interdisciplinary efforts are needed to study both technical behavior and broader impacts.

Where Pith is reading between the lines

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

  • If homogenization holds, oversight could shift from regulating individual applications to auditing and updating the small number of base models.
  • Developers might test whether fine-tuning or prompt changes can reliably isolate or correct base-model defects without full retraining.
  • The pattern suggests similar dynamics could appear in other scaled systems, such as large simulation models or scientific foundation models.
  • A practical next step would be systematic comparison of multiple independent foundation models to measure how much their defect profiles actually overlap.

Load-bearing premise

That training at current scales reliably produces capabilities that cannot be foreseen from smaller models and that most future AI work will converge on a few shared foundation models without separate safeguards.

What would settle it

A controlled experiment showing that performance gains on downstream tasks can be fully predicted by scaling laws fitted to smaller models, with no new qualitative behaviors appearing at foundation scale.

read the original abstract

AI is undergoing a paradigm shift with the rise of models (e.g., BERT, DALL-E, GPT-3) that are trained on broad data at scale and are adaptable to a wide range of downstream tasks. We call these models foundation models to underscore their critically central yet incomplete character. This report provides a thorough account of the opportunities and risks of foundation models, ranging from their capabilities (e.g., language, vision, robotics, reasoning, human interaction) and technical principles(e.g., model architectures, training procedures, data, systems, security, evaluation, theory) to their applications (e.g., law, healthcare, education) and societal impact (e.g., inequity, misuse, economic and environmental impact, legal and ethical considerations). Though foundation models are based on standard deep learning and transfer learning, their scale results in new emergent capabilities,and their effectiveness across so many tasks incentivizes homogenization. Homogenization provides powerful leverage but demands caution, as the defects of the foundation model are inherited by all the adapted models downstream. Despite the impending widespread deployment of foundation models, we currently lack a clear understanding of how they work, when they fail, and what they are even capable of due to their emergent properties. To tackle these questions, we believe much of the critical research on foundation models will require deep interdisciplinary collaboration commensurate with their fundamentally sociotechnical nature.

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

1 major / 2 minor

Summary. The paper introduces foundation models as large-scale models trained on broad data that adapt to diverse downstream tasks (e.g., BERT, DALL-E, GPT-3). It surveys capabilities in language, vision, robotics, reasoning, and human interaction; technical principles including architectures, training, data, systems, security, evaluation, and theory; applications in law, healthcare, and education; and societal impacts such as inequity, misuse, economic/environmental effects, and legal/ethical issues. The central thesis is that, while rooted in standard deep learning and transfer learning, scale produces emergent capabilities and incentivizes homogenization, yielding leverage but also risks since defects propagate to all adapted downstream models. It notes the current lack of understanding of their mechanisms, failure modes, and capabilities, and calls for interdisciplinary research.

Significance. If the observations on emergence and homogenization hold, the report is significant as a timely, broad synthesis that frames the foundation-model paradigm and its sociotechnical implications. It consolidates existing work, identifies gaps in understanding emergence and failures, and advocates caution plus collaboration, serving as a reference point for the field at a time of rapid deployment.

major comments (1)
  1. [Abstract] Abstract: the claim that scale 'results in new emergent capabilities' is load-bearing for the paradigm-shift framing yet is presented without quantitative evidence, formal bounds, or specific citations to studies showing behaviors unpredictable from smaller models; this weakens the distinction from standard scaling laws.
minor comments (2)
  1. [Abstract] Abstract: the long compound sentence enumerating topics reduces readability; splitting into shorter sentences would improve clarity.
  2. [Throughout] Throughout: the term 'homogenization' is used repeatedly but receives no explicit initial definition or scope clarification, which could confuse readers unfamiliar with the concept.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their constructive and insightful review. We address the single major comment below and have revised the manuscript to strengthen the abstract's claim while remaining faithful to the empirical evidence and open questions discussed in the body of the paper.

read point-by-point responses
  1. Referee: [Abstract] Abstract: the claim that scale 'results in new emergent capabilities' is load-bearing for the paradigm-shift framing yet is presented without quantitative evidence, formal bounds, or specific citations to studies showing behaviors unpredictable from smaller models; this weakens the distinction from standard scaling laws.

    Authors: We agree that the abstract statement is central to the paper's framing and benefits from explicit support. The body of the manuscript already cites and discusses empirical evidence for emergence, including in-context learning and other capabilities in GPT-3 (Brown et al., 2020) that were not observed or predictable from smaller-scale models, as well as scaling-law analyses (Kaplan et al., 2020). Formal bounds on emergence are indeed unavailable and are highlighted as an open research question throughout the paper. To directly address the referee's concern, we have revised the abstract to include a specific citation to this literature and a brief qualifier noting the empirical nature of the observed behaviors. This revision preserves the distinction from standard scaling laws without overstating theoretical guarantees. revision: yes

Circularity Check

0 steps flagged

No circularity: position paper with no derivations or fitted predictions

full rationale

The document is a discursive survey and position paper that defines foundation models, surveys capabilities/risks, and calls for interdisciplinary research. It contains no equations, no parameter fitting, no 'predictions' of quantities, and no derivation chain. All technical claims are framed as observations from external literature or as open questions (e.g., 'we currently lack a clear understanding'). No step reduces to a self-definition, self-citation load-bearing premise, or renamed known result by construction. The central statements about emergence and homogenization are presented as motivating observations rather than proven results, consistent with the paper's explicit acknowledgment of incomplete understanding.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The report rests on domain assumptions from deep learning rather than new axioms or entities; no free parameters or invented physical entities are introduced.

axioms (2)
  • domain assumption Training scale produces emergent capabilities not present or predictable in smaller models.
    Invoked in the abstract and introduction as a defining property of foundation models.
  • domain assumption Widespread effectiveness will drive homogenization of AI development around a few base models.
    Stated as a consequence of adaptability across tasks.

pith-pipeline@v0.9.0 · 6048 in / 1371 out tokens · 58825 ms · 2026-05-10T16:10:14.335079+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

  • IndisputableMonolith/Foundation/RealityFromDistinction reality_from_one_distinction unclear

    Though foundation models are based on standard deep learning and transfer learning, their scale results in new emergent capabilities, and their effectiveness across so many tasks incentivizes homogenization. Homogenization provides powerful leverage but demands caution, as the defects of the foundation model are inherited by all the adapted models downstream.

Forward citations

Cited by 60 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Contrastive Identification and Generation in the Limit

    cs.LG 2026-05 unverdicted novelty 8.0

    Contrastive pair presentations yield exact identifiability characterizations via a geometric refinement of Angluin's condition, a new contrastive closure dimension for generation, mutual incomparability with text iden...

  2. Prism: Symbolic Superoptimization of Tensor Programs

    cs.PL 2026-04 unverdicted novelty 8.0

    Prism is the first symbolic superoptimizer for tensor programs that uses sGraph for compact representation of program families, two-level search, e-graph equivalence checking, and auto-tuning to achieve up to 2.2x spe...

  3. DSPy: Compiling Declarative Language Model Calls into Self-Improving Pipelines

    cs.CL 2023-10 conditional novelty 8.0

    DSPy compiles short declarative programs into LM pipelines that self-optimize and outperform both standard few-shot prompting and expert-written chains on math, retrieval, and QA tasks.

  4. Generative Agents: Interactive Simulacra of Human Behavior

    cs.HC 2023-04 accept novelty 8.0

    Generative agents with memory streams, reflection, and planning using LLMs exhibit believable individual and emergent social behaviors in a simulated town.

  5. Decoupled and Divergence-Conditioned Prompt for Multi-domain Dynamic Graph Foundation Models

    cs.LG 2026-05 conditional novelty 7.0

    DyGFM introduces decoupled pre-training and divergence-conditioned prompts to create the first multi-domain dynamic graph foundation model that outperforms baselines on node classification and link prediction.

  6. Pretraining Strategies and Scaling for ECG Foundation Models: A Systematic Study

    eess.SP 2026-05 unverdicted novelty 7.0

    Contrastive predictive coding pretraining combined with structured state space models yields the strongest ECG foundation models, with continued gains from scaling data to 11 million samples.

  7. AcuityBench: Evaluating Clinical Acuity Identification and Uncertainty Alignment

    cs.AI 2026-05 unverdicted novelty 7.0

    AcuityBench harmonizes five datasets into a four-level acuity framework to evaluate LLMs on clinical urgency identification, error patterns, and uncertainty alignment across QA and conversational formats.

  8. TokaMind for Power Grid: Cross-Domain Transfer from Fusion Plasma

    physics.plasm-ph 2026-05 unverdicted novelty 7.0

    TokaMind, pre-trained on MAST tokamak data, transfers to power grid PMU data for severe event classification with F1 0.837, where difficulty depends on grid topology and CSD indicators boost early-warning performance ...

  9. BiAxisAudit: A Novel Framework to Evaluate LLM Bias Across Prompt Sensitivity and Response-Layer Divergence

    cs.CL 2026-05 unverdicted novelty 7.0

    BiAxisAudit measures LLM bias on two axes—across-prompt sensitivity via factorial grids and within-response divergence via split coding—revealing that task format explains as much variance as model choice and that 63....

  10. Curated Synthetic Data Doesn't Have to Collapse: A Theoretical Study of Generative Retraining with Pluralistic Preferences

    cs.LG 2026-05 unverdicted novelty 7.0

    Recursive generative retraining with pluralistic preferences converges to a stable diverse distribution that satisfies a weighted Nash bargaining solution.

  11. Learning to Communicate Locally for Large-Scale Multi-Agent Pathfinding

    cs.AI 2026-05 unverdicted novelty 7.0

    LC-MAPF uses multi-round local communication between neighboring agents in a pre-trained model to outperform prior learning-based MAPF solvers on diverse unseen scenarios while preserving scalability.

  12. How Many Iterations to Jailbreak? Dynamic Budget Allocation for Multi-Turn LLM Evaluation

    cs.LG 2026-05 unverdicted novelty 7.0

    DAPRO provides the first dynamic, theoretically guaranteed way to allocate interaction budgets across test cases for bounding time-to-event in multi-turn LLM evaluations, achieving tighter coverage than static conform...

  13. On the Safety of Graph Representation Learning

    cs.LG 2026-05 unverdicted novelty 7.0

    GRL-Safety benchmark shows that safety in graph representation learning depends on interactions between method design and specific graph stresses rather than broad method families.

  14. Graphlets as Building Blocks for Structural Vocabulary in Knowledge Graph Foundation Models

    cs.AI 2026-05 unverdicted novelty 7.0

    Graphlets mined as structural tokens improve zero-shot inductive and transductive link prediction in knowledge graph foundation models across 51 diverse graphs.

  15. LUCAS-MEGA: A Large-Scale Multimodal Dataset for Representation Learning in Soil-Environment Systems

    cs.LG 2026-05 unverdicted novelty 7.0

    LUCAS-MEGA fuses 68 soil-environment datasets into a 70k-sample multimodal resource that supports self-supervised pretraining of SoilFormer, whose representations align with known soil processes.

  16. LUCAS-MEGA: A Large-Scale Multimodal Dataset for Representation Learning in Soil-Environment Systems

    cs.LG 2026-05 unverdicted novelty 7.0

    LUCAS-MEGA fuses 68 heterogeneous soil datasets into a 70k-sample multimodal collection and demonstrates its value by pretraining a tabular transformer whose representations align with established soil processes.

  17. Erase Persona, Forget Lore: Benchmarking Multimodal Copyright Unlearning in Large Vision Language Models

    cs.CV 2026-05 unverdicted novelty 7.0

    CoVUBench is the first benchmark framework for evaluating multimodal copyright unlearning in LVLMs via synthetic data, systematic variations, and a dual protocol for forgetting efficacy and utility preservation.

  18. CUE: Concept-Aware Multi-Label Expansion to Mitigate Concept Confusion in Long-Tailed Learning

    cs.CV 2026-05 unverdicted novelty 7.0

    CUE mitigates concept confusion in long-tailed visual recognition by expanding supervision with multi-label concept sets from zero-shot CLIP and LLMs, using auxiliary Binary Logit-Adjustment losses to achieve stronger...

  19. Self-Evolving Software Agents

    cs.SE 2026-04 unverdicted novelty 7.0

    Self-evolving agents use BDI-LLM architecture to autonomously discover new goals and generate executable code from minimal prior knowledge in dynamic multi-agent settings.

  20. LLM-Assisted Empirical Software Engineering: Systematic Literature Review and Research Agenda

    cs.SE 2026-04 unverdicted novelty 7.0

    A systematic review of 50 studies identifies 69 LLM-assisted tasks in empirical software engineering, concentrated in data processing and analysis with gaps in human-centered integration and reproducibility reporting.

  21. A Systematic Survey of Security Threats and Defenses in LLM-Based AI Agents: A Layered Attack Surface Framework

    cs.CR 2026-04 unverdicted novelty 7.0

    A new 7x4 taxonomy organizes agentic AI security threats by architectural layer and persistence timescale, revealing under-explored upper layers and missing defenses after surveying 116 papers.

  22. Participatory provenance as representational auditing for AI-mediated public consultation

    cs.AI 2026-04 unverdicted novelty 7.0

    Participatory provenance auditing of Canada's AI strategy consultation shows official AI summaries exclude 15-17% of participants more than random baselines, with 33-88% exclusion for dissent clusters.

  23. Rethinking Scale: Deployment Trade-offs of Small Language Models under Agent Paradigms

    cs.CL 2026-04 unverdicted novelty 7.0

    Single-agent systems with tools provide the optimal performance-efficiency trade-off for small language models, outperforming base models and multi-agent setups.

  24. Faster by Design: Interactive Aerodynamics via Neural Surrogates Trained on Expert-Validated CFD

    cs.LG 2026-04 unverdicted novelty 7.0

    A graph-based neural operator trained on expert-validated race-car CFD data reaches accuracy levels usable for early-stage interactive aerodynamic design exploration.

  25. TacticGen: Grounding Adaptable and Scalable Generation of Football Tactics

    cs.AI 2026-04 conditional novelty 7.0

    TacticGen generates realistic, adaptable football tactics via a multi-agent diffusion transformer trained on 3.3M events and 100M frames, supporting rule-, language-, or model-based guidance at inference time.

  26. From UAV Imagery to Agronomic Reasoning: A Multimodal LLM Benchmark for Plant Phenotyping

    cs.CV 2026-04 unverdicted novelty 7.0

    PlantXpert benchmark shows fine-tuned VLMs reach up to 78% accuracy on plant phenotyping but scaling gains plateau and quantitative biological reasoning remains weak.

  27. SafeAdapt: Provably Safe Policy Updates in Deep Reinforcement Learning

    cs.LG 2026-04 unverdicted novelty 7.0

    SafeAdapt certifies a Rashomon set of safe policies from demonstration data and projects updates from arbitrary RL algorithms onto it to guarantee preservation of safety on source tasks.

  28. The Shrinking Lifespan of LLMs in Science

    cs.DL 2026-04 unverdicted novelty 7.0

    LLM adoption in science follows a compressing inverted-U trajectory where release year predicts time-to-peak and lifespan better than model attributes.

  29. A Mixture of Experts Foundation Model for Scanning Electron Microscopy Image Analysis

    cs.LG 2026-04 unverdicted novelty 7.0

    A mixture-of-experts transformer foundation model pretrained on diverse SEM images enables generalization across materials and outperforms SOTA on unsupervised defocus-to-focus restoration.

  30. AI Agents Under EU Law

    cs.CY 2026-04 unverdicted novelty 7.0

    AI agent providers face an exhaustive inventory requirement for actions and data flows, as high-risk systems with untraceable behavioral drift cannot meet the AI Act's essential requirements.

  31. Many Preferences, Few Policies: Towards Scalable Language Model Personalization

    cs.CL 2026-04 unverdicted novelty 7.0

    PALM produces a small portfolio of LLMs that contains a near-optimal model for any user preference weight vector, with theoretical bounds on portfolio size and approximation quality.

  32. Training a Student Expert via Semi-Supervised Foundation Model Distillation

    cs.CV 2026-04 conditional novelty 7.0

    A semi-supervised framework distills vision foundation models into compact instance segmentation experts that outperform their teachers by up to 11.9 AP on Cityscapes and 8.6 AP on ADE20K while being 11 times smaller.

  33. Hidden Reliability Risks in Large Language Models: Systematic Identification of Precision-Induced Output Disagreements

    cs.AI 2026-04 unverdicted novelty 7.0

    PrecisionDiff is a differential testing framework that uncovers widespread precision-induced behavioral disagreements in aligned LLMs, including safety-critical jailbreak divergences across precision formats.

  34. TextGrad: Automatic "Differentiation" via Text

    cs.CL 2024-06 unverdicted novelty 7.0

    TextGrad performs automatic differentiation for compound AI systems by backpropagating natural-language feedback from LLMs to optimize variables ranging from code to molecular structures.

  35. VoxPoser: Composable 3D Value Maps for Robotic Manipulation with Language Models

    cs.RO 2023-07 unverdicted novelty 7.0

    VoxPoser uses LLMs to compose 3D value maps via VLM interaction for model-based synthesis of robust robot trajectories on open-set language-specified manipulation tasks.

  36. Voyager: An Open-Ended Embodied Agent with Large Language Models

    cs.AI 2023-05 unverdicted novelty 7.0

    Voyager achieves superior lifelong learning in Minecraft by combining an automatic exploration curriculum, a library of executable skills, and iterative LLM prompting with environment feedback, yielding 3.3x more uniq...

  37. QLoRA: Efficient Finetuning of Quantized LLMs

    cs.LG 2023-05 conditional novelty 7.0

    QLoRA finetunes 4-bit quantized LLMs via LoRA adapters to match full-precision performance while using far less memory, enabling 65B-scale training on single GPUs and producing Guanaco models near ChatGPT level.

  38. Segment Anything

    cs.CV 2023-04 unverdicted novelty 7.0

    A promptable model trained on 1B masks achieves competitive zero-shot segmentation performance across tasks and is released publicly with its dataset.

  39. A Time Series is Worth 64 Words: Long-term Forecasting with Transformers

    cs.LG 2022-11 conditional novelty 7.0

    PatchTST uses subseries patching and channel-independent Transformers to deliver significantly better long-term multivariate time series forecasting and strong self-supervised transfer performance.

  40. Beyond the Imitation Game: Quantifying and extrapolating the capabilities of language models

    cs.CL 2022-06 accept novelty 7.0

    BIG-bench is a 204-task benchmark that measures scaling trends, calibration, and absolute limitations of language models across knowledge, reasoning, and social domains.

  41. A Generalist Agent

    cs.AI 2022-05 accept novelty 7.0

    Gato is a multi-modal, multi-task, multi-embodiment generalist policy using one transformer network to handle text, vision, games, and robotics tasks.

  42. OPT: Open Pre-trained Transformer Language Models

    cs.CL 2022-05 unverdicted novelty 7.0

    OPT releases open decoder-only transformers up to 175B parameters that match GPT-3 performance at one-seventh the carbon cost, along with code and training logs.

  43. Do As I Can, Not As I Say: Grounding Language in Robotic Affordances

    cs.RO 2022-04 accept novelty 7.0

    SayCan combines an LLM's high-level semantic knowledge with robot skill value functions to select only feasible actions, enabling completion of abstract natural-language instructions on a real mobile manipulator.

  44. Multitask Prompted Training Enables Zero-Shot Task Generalization

    cs.LG 2021-10 conditional novelty 7.0

    Multitask fine-tuning of an encoder-decoder model on prompted datasets produces zero-shot generalization that often beats models up to 16 times larger on standard benchmarks.

  45. Correcting Influence: Unboxing LLM Outputs with Orthogonal Latent Spaces

    cs.LG 2026-05 unverdicted novelty 6.0

    A latent mediation framework with sparse autoencoders enables non-additive token-level influence attribution in LLMs by learning orthogonal features and back-propagating attributions.

  46. No One Knows the State of the Art in Geospatial Foundation Models

    cs.CV 2026-05 accept novelty 6.0

    An audit of 152 papers reveals that geospatial foundation models lack standardized evaluations, training controls, and weight releases, so no one knows the state of the art.

  47. SAGE: A Self-Evolving Agentic Graph-Memory Engine for Structure-Aware Associative Memory

    cs.AI 2026-05 unverdicted novelty 6.0

    SAGE is a self-evolving agentic graph-memory engine that dynamically constructs and refines structured memory graphs via writer-reader feedback, yielding performance gains on multi-hop QA, open-domain retrieval, and l...

  48. A CAP-like Trilemma for Large Language Models: Correctness, Non-bias, and Utility under Semantic Underdetermination

    cs.AI 2026-05 unverdicted novelty 6.0

    Under semantic underdetermination, LLMs cannot always guarantee strong correctness, strict non-bias, and high utility at once.

  49. DECO: Sparse Mixture-of-Experts with Dense-Comparable Performance on End-Side Devices

    cs.LG 2026-05 unverdicted novelty 6.0

    DECO sparse MoE matches dense Transformer performance at 20% expert activation with a 3x hardware inference speedup.

  50. DECO: Sparse Mixture-of-Experts with Dense-Comparable Performance on End-Side Devices

    cs.LG 2026-05 conditional novelty 6.0

    DECO matches dense model performance at 20% expert activation via ReLU-based routing with learnable scaling and the NormSiLU activation, plus a 3x real-hardware speedup.

  51. Quantifying Concentration Phenomena of Mean-Field Transformers in the Low-Temperature Regime

    math.AP 2026-05 unverdicted novelty 6.0

    In the low-temperature regime, the token distribution in mean-field transformers concentrates onto the push-forward under a key-query-value projection with Wasserstein distance scaling as √(log(β+1)/β) exp(Ct) + exp(-ct).

  52. When Should Teachers Control AI Generation for Mathematics Visuals?

    cs.HC 2026-05 conditional novelty 6.0

    Post-generation control in AI-assisted math visual creation yields higher teacher ratings for predictability and correctness than pre- or mid-generation control, with qualitative trade-offs in agency and effort.

  53. Valid Best-Model Identification for LLM Evaluation via Low-Rank Factorization

    cs.LG 2026-05 unverdicted novelty 6.0

    Doubly robust estimators that incorporate low-rank predictions enable valid finite-sample confidence intervals for best-model identification under adaptive sampling and without-replacement example selection in LLM evaluation.

  54. Biosignal Fingerprinting: A Cross-Modal PPG-ECG Foundation Model

    cs.LG 2026-05 unverdicted novelty 6.0

    A cross-modal masked autoencoder creates reusable biosignal fingerprints that match or exceed specialist models on seven cardiovascular tasks using only single-modality input.

  55. Learning to Communicate Locally for Large-Scale Multi-Agent Pathfinding

    cs.AI 2026-05 unverdicted novelty 6.0

    LC-MAPF is a decentralized MAPF solver that uses a learnable multi-round communication module among nearby agents to outperform prior IL and RL methods while preserving scalability.

  56. Post-training makes large language models less human-like

    cs.CL 2026-05 unverdicted novelty 6.0

    Post-training reduces LLMs' behavioral alignment with humans across families and sizes, with the misalignment increasing in newer generations while persona induction fails to improve individual-level predictions.

  57. Revisiting Transformer Layer Parameterization Through Causal Energy Minimization

    cs.LG 2026-05 unverdicted novelty 6.0

    CEM recasts Transformer layers as energy minimization steps, enabling constrained parameterizations like weight sharing and low-rank interactions that match standard baselines in 100M-scale language modeling.

  58. WeatherSyn: An Instruction Tuning MLLM For Weather Forecasting Report Generation

    cs.CL 2026-05 unverdicted novelty 6.0

    WeatherSyn is the first instruction-tuned MLLM for weather forecasting report generation, outperforming closed-source models on a new dataset of 31 US cities across 8 weather aspects.

  59. A renormalization-group inspired lattice-based framework for piecewise generalized linear models

    stat.ME 2026-05 unverdicted novelty 6.0

    RG-inspired lattice models for piecewise GLMs provide explicit interpretable partitions and a replica-analysis-derived scaling law for regularization that allows increasing complexity without expected rise in generali...

  60. The Geopolitics of AI Safety: A Causal Analysis of Regional LLM Bias

    cs.AI 2026-05 unverdicted novelty 6.0

    Causal analysis of LLMs finds standard bias metrics overestimate demographic effects due to context toxicity, with Western models showing higher refusal rates for certain groups and Eastern models showing targeted reg...

Reference graph

Works this paper leans on

18 extracted references · 18 canonical work pages · cited by 158 Pith papers · 2 internal anchors

  1. [1]

    Invariant Risk Minimization

    Vqa: Visual question answering. In Proceedings of the IEEE international conference on computer vision . 2425–2433. Maria Antoniak and David Mimno. 2021. Bad Seeds: Evaluating Lexical Methods for Bias Measurement. In Proceedings of ACL 2021. Hannah Arendt. 1987. Collective Responsibility. Springer Netherlands, Dordrecht, 43–50. Martin Arjovsky, Léon Botto...

  2. [2]

    Neural Machine Translation by Jointly Learning to Align and Translate

    On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation. PloS one 10, 7 (2015), e0130140. Claudine Badue, Rânik Guidolini, Raphael Vivacqua Carneiro, Pedro Azevedo, Vinicius Brito Cardoso, Avelino Forechi, Luan Jesus, Rodrigo Berriel, Thiago Meireles Paixão, Filipe Mutz, et al. 2020. Self-driving cars: A survey. ...

  3. [3]

    Documentation Debt

    Systematic Generalization: What Is Required and Can It Be Learned?. In International Conference on Learning Representations. Annette Baier. 1986. Trust and Antitrust. Ethics 96, 2 (1986), 231–260. http://www.jstor.org/stable/2381376 164 Center for Research on Foundation Models (CRFM) Andrea Bajcsy, Dylan P. Losey, M. O’Malley, and A. Dragan. 2017. Learnin...

  4. [4]

    cost disease

    Interactive perception: Leveraging action in perception and perception in action. IEEE Transactions on Robotics 33, 6 (2017), 1273–1291. Tolga Bolukbasi, Kai-Wei Chang, James Y Zou, Venkatesh Saligrama, and Adam T Kalai. 2016. Man is to Computer Programmer as Woman is to Homemaker? Debiasing Word Embeddings. InAdvances in Neural Information Processing Sys...

  5. [5]

    Gritsenko, et al

    The Benchmark Lottery. ArXiv abs/2107.07002 (2021). Mauricio Delbracio, Damien Kelly, Michael S Brown, and Peyman Milanfar. 2021. Mobile Computational Photography: A Tour. arXiv preprint arXiv:2102.09000 (2021). Dina Demner-Fushman, Yassine Mrabet, and Asma Ben Abacha. 2020. Consumer health information and question answering: helping consumers find answer...

  6. [6]

    ArXiv abs/2107.03451 (2021)

    Anticipating Safety Issues in E2E Conversational AI: Framework and Tooling. ArXiv abs/2107.03451 (2021). https://arxiv.org/abs/2107.03451 Emily Dinan, Angela Fan, Ledell Wu, Jason Weston, Douwe Kiela, and Adina Williams. 2020. Multi-Dimensional Gender Bias Classification. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Proce...

  7. [7]

    The tactics & tropes of the Internet Research Agency. https://digitalcommons.unl.edu/cgi/viewcontent.cgi?article= 1003&context=senatedocs Michael Diskin, Alexey Bukhtiyarov, Max Ryabinin, Lucile Saulnier, Quentin Lhoest, Anton Sinitsin, Dmitry Popov, Dmitry Pyrkin, Maxim Kashirin, Alexander Borzunov, et al. 2021. Distributed Deep Learning in Open Collabor...

  8. [8]

    The First Mathematical Reasoning in General Artificial Intelligence Workshop, ICLR 2021 (2021)

    Training a First-Order Theorem Prover from Synthetic Data. The First Mathematical Reasoning in General Artificial Intelligence Workshop, ICLR 2021 (2021). https://mathai-iclr.github.io/papers/papers/MATHAI_18_paper.pdf Jaime F. Fisac, Neil F. Lugovoy, Vicenç Rúbies Royo, S. Ghosh, and C. Tomlin. 2019. Bridging Hamilton-Jacobi Safety Analysis and Reinforce...

  9. [9]

    Intrinsic bias metrics do not correlate with application bias

    Multimodal neurons in artificial neural networks. Distill 6, 3 (2021), e30. Seraphina Goldfarb-Tarrant, Rebecca Marchant, Ricardo Muñoz Sánchez, Mugdha Pandya, and Adam Lopez. 2021. Intrinsic Bias Metrics Do Not Correlate with Application Bias. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th Internat...

  10. [10]

    nutrition label

    Learning agile and dynamic motor skills for legged robots. Science Robotics 4, 26 (2019). Janet Shibley Hyde, Rebecca S. Bigler, Daphna Joel, Charlotte Chucky Tate, and Sari M. van Anders. 2019. The Future of Sex and Gender in Psychology: Five Challenges to the Gender Binary. American Psychologist 74 (2019), 171–193. H. Iida, Dung Thai, Varun Manjunatha, ...

  11. [11]

    Cognitive Science , author =

    Unmasking Clever Hans predictors and assessing what machines really learn. Nature communications 10, 1 (2019), 1–8. Jill H. Larkin and Herbert A. Simon. 1987. Why a Diagram is (Sometimes) Worth Ten Thousand Words. Cognitive Science 11, 1 (1987), 65–100. https://doi.org/10.1111/j.1551-6708.1987.tb00863.x arXiv:https://onlinelibrary.wiley.com/doi/pdf/10.111...

  12. [12]

    How Can We Accelerate Progress Towards Human-like Linguistic Generalization?

    Latent predictor networks for code generation. arXiv preprint arXiv:1603.06744 (2016). Tal Linzen. 2020. How Can We Accelerate Progress Towards Human-like Linguistic Generalization?. InProceedings of the 58th Annual Meeting of the Association for Computational Linguistics . Association for Computational Linguistics, Online, 5210–5217. https://doi.org/10.1...

  13. [13]

    Everyone wants to do the model work, not the data work

    Human Comprehension of Fairness in Machine Learning. In Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society (New York, NY, USA) (AIES ’20). Association for Computing Machinery, New York, NY, USA, 152. https://doi.org/10.1145/3375627.3375819 Hassan Sajjad, Fahim Dalvi, Nadir Durrani, and Preslav Nakov. 2020. On the Effect of Dropping Layers o...

  14. [14]

    The woman worked as a babysitter: On biases in language generation

    Time-contrastive networks: Self-supervised learning from video. In 2018 IEEE international conference on robotics and automation (ICRA). IEEE, 1134–1141. Ali Shafahi, Parsa Saadatpanah, Chen Zhu, Amin Ghiasi, Christoph Studer, David Jacobs, and Tom Goldstein. 2019. Adver- sarially robust transfer learning. arXiv preprint arXiv:1905.08232 (2019). Neal A Sh...

  15. [15]

    https://www.cambridge.org/core/journals/european-review/article/abs/improving-ratings-audit-in-the-british- university-system/FC2EE640C0C44E3DB87C29FB666E9AAB Yolande Strengers, Lizhen Qu, Qiongkai Xu, and Jarrod Knibbe. 2020. Adhering, Steering, and Queering: Treatment of Gender in Natural Language Generation. In Proceedings of the 2020 CHI Conference on...

  16. [16]

    Jiachen T

    Multilingual is not enough: BERT for Finnish. arXiv preprint arXiv:1912.07076 (2019). Rob Voigt, Nicholas P Camp, Vinodkumar Prabhakaran, William L Hamilton, Rebecca C Hetey, Camilla M Griffiths, David Jurgens, Dan Jurafsky, and Jennifer L Eberhardt. 2017. Language from police body camera footage shows racial disparities in officer respect. Proceedings of...

  17. [17]

    Clinical Pharmacology & Therapeutics 92, 4 (2012), 414–417

    Pharmacogenomics knowledge for personalized medicine. Clinical Pharmacology & Therapeutics 92, 4 (2012), 414–417. Jenna Wiens, Suchi Saria, Mark Sendak, Marzyeh Ghassemi, Vincent X Liu, Finale Doshi-Velez, Kenneth Jung, Katherine Heller, David Kale, Mohammed Saeed, et al. 2019. Do no harm: a roadmap for responsible machine learning for health care. Nature...

  18. [18]

    arXiv preprint arXiv:2010.11934 (2020)

    mt5: A massively multilingual pre-trained text-to-text transformer. arXiv preprint arXiv:2010.11934 (2020). Eugene Yang, Sean MacAvaney, David D Lewis, and Ophir Frieder. 2021. Goldilocks: Just-Right Tuning of BERT for Technology-Assisted Review. arXiv preprint arXiv:2105.01044 (2021). Mengjiao Yang and Ofir Nachum. 2021. Representation Matters: Offline P...