pith. sign in

arxiv: 2012.14913 · v2 · pith:37XQHZEFnew · submitted 2020-12-29 · 💻 cs.CL

Transformer Feed-Forward Layers Are Key-Value Memories

Mor Geva , Roei Schuster , Jonathan Berant , Omer Levy This is my paper

Pith reviewed 2026-05-13 23:30 UTC · model grok-4.3

classification 💻 cs.CL
keywords transformersfeed-forward layerskey-value memorieslanguage modelsmodel interpretabilityneural network analysis
0
0 comments X

The pith

Transformer feed-forward layers function as key-value memories.

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

The paper shows that feed-forward layers, which hold two-thirds of a transformer's parameters, act as key-value memories. Each key matches particular textual patterns seen during training, while each value produces a distribution over likely next tokens. Lower layers focus on simple surface patterns and upper layers on semantic ones, with the layer output combining multiple such memories. Residual connections then refine the combined result into the final prediction.

Core claim

Feed-forward layers in transformer-based language models operate as key-value memories, where each key correlates with textual patterns in the training examples, and each value induces a distribution over the output vocabulary. The learned patterns are human-interpretable, with lower layers capturing shallow patterns and upper layers learning more semantic ones. The values complement the keys by inducing output distributions that concentrate probability mass on tokens likely to appear immediately after each pattern. The output of a feed-forward layer is a composition of its memories, refined throughout the model via residual connections.

What carries the argument

Key-value memory pairs inside each feed-forward layer, where a key detects an input pattern and a value supplies a next-token distribution.

Load-bearing premise

The correlations between learned keys and input patterns, and between values and output distributions, reflect the actual computation the model performs at inference time.

What would settle it

Alter the weights of one specific key-value pair and measure whether the model's next-token predictions shift only for inputs that match the corresponding pattern.

read the original abstract

Feed-forward layers constitute two-thirds of a transformer model's parameters, yet their role in the network remains under-explored. We show that feed-forward layers in transformer-based language models operate as key-value memories, where each key correlates with textual patterns in the training examples, and each value induces a distribution over the output vocabulary. Our experiments show that the learned patterns are human-interpretable, and that lower layers tend to capture shallow patterns, while upper layers learn more semantic ones. The values complement the keys' input patterns by inducing output distributions that concentrate probability mass on tokens likely to appear immediately after each pattern, particularly in the upper layers. Finally, we demonstrate that the output of a feed-forward layer is a composition of its memories, which is subsequently refined throughout the model's layers via residual connections to produce the final output distribution.

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 manuscript claims that feed-forward layers in transformer language models function as key-value memories: keys correlate with human-interpretable textual patterns from the training data (shallow in lower layers, semantic in upper layers), values induce complementary next-token distributions, and the layer output is a composition of activated memories that is refined via residual connections to produce the final distribution.

Significance. If the interpretation holds, it supplies a concrete mechanistic account of two-thirds of transformer parameters, grounded in empirical probing that reveals interpretable patterns and input-output complementarity. This could support targeted model editing and deeper understanding of how transformers store and retrieve information.

major comments (2)
  1. [§4] §4 (pattern extraction and activation analysis): the reported correlations between keys and n-gram patterns are statistical matches only; without causal interventions such as key ablation, activation patching, or counterfactual input edits, it remains possible that the observed associations are side-effects rather than the operative mechanism in the forward pass W2 · f(W1x).
  2. [§3.2] §3.2 (memory composition claim): the assertion that the FF output is exactly a composition of memories is not fully reconciled with the non-linearity f; the paper should show (via expansion or controlled experiments) that multiple simultaneously activated keys combine linearly in the effective computation rather than through non-linear interactions.
minor comments (2)
  1. [Figure 3] Figure 3 and Table 1: the value-distribution visualizations would be clearer with an explicit random-key baseline to quantify how much the reported concentration exceeds chance.
  2. Notation: the mapping from matrix rows/columns to keys and values is introduced without a compact equation; adding a single-line definition (e.g., key_i = row i of W1) would aid readability.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the insightful comments and the recommendation for major revision. We provide detailed responses to each major comment below, indicating where we will revise the manuscript to address the concerns.

read point-by-point responses

  1. Referee: [§4] §4 (pattern extraction and activation analysis): the reported correlations between keys and n-gram patterns are statistical matches only; without causal interventions such as key ablation, activation patching, or counterfactual input edits, it remains possible that the observed associations are side-effects rather than the operative mechanism in the forward pass W2 · f(W1x).

    Authors: We acknowledge that the primary evidence consists of strong statistical correlations between the keys and specific textual patterns, identified by finding inputs that highly activate each key. These correlations are not merely side-effects, as they directly correspond to the computation in the forward pass where high key activation leads to the associated value contributing to the output. Nevertheless, to provide stronger causal evidence, we will add experiments involving the ablation of specific keys and measure the impact on the model's predictions for inputs containing the corresponding patterns. revision: yes

  2. Referee: [§3.2] §3.2 (memory composition claim): the assertion that the FF output is exactly a composition of memories is not fully reconciled with the non-linearity f; the paper should show (via expansion or controlled experiments) that multiple simultaneously activated keys combine linearly in the effective computation rather than through non-linear interactions.

    Authors: The non-linearity f is applied element-wise to the pre-activations, meaning each key's activation scalar is computed independently as f(key_i · x). The layer output is then the linear combination sum_i activation_i * value_i. Therefore, the memories combine linearly once activated, with the non-linearity affecting only the activation strength of each memory individually. We will revise §3.2 to include this explicit mathematical expansion and present controlled experiments where we compare the actual FF output to the linear combination of individually computed memory contributions. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical correlations from trained models do not reduce to self-definition or fitted inputs

full rationale

The paper's central claim rests on post-training analysis of existing transformer weights: identifying input patterns that strongly activate specific rows of the first FF matrix (treated as keys) and observing that the corresponding columns of the second matrix induce next-token distributions (treated as values). These are measured correlations on held-out data and activation statistics, not quantities defined in terms of each other or obtained by fitting a parameter whose value is then relabeled as a prediction. No equations are shown to be equivalent by construction, no uniqueness theorem is imported from the authors' prior work to force the interpretation, and the residual composition argument is demonstrated via direct layer-wise ablation rather than assumed. The derivation chain is therefore self-contained against external benchmarks (the trained models themselves).

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The claim rests on empirical observations from trained transformers rather than new free parameters or invented physical entities; standard language-modeling assumptions are used.

axioms (1)
  • domain assumption Transformers are trained via next-token prediction on large corpora
    Invoked implicitly when linking keys to training patterns and values to output distributions.
invented entities (1)
  • key-value memory structure inside feed-forward layers no independent evidence
    purpose: Interpretive lens to explain layer behavior
    This is a conceptual reframing of existing weights, not a new postulated object with independent falsifiable predictions.

pith-pipeline@v0.9.0 · 5435 in / 1148 out tokens · 38726 ms · 2026-05-13T23:30:42.949422+00:00 · methodology

discussion (0)

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

Forward citations

Cited by 60 Pith papers

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

  1. Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 small

    cs.LG 2022-11 conditional novelty 8.0

    GPT-2 small solves indirect object identification via a circuit of 26 attention heads organized into seven functional classes discovered through causal interventions.

  2. Vision-Default, Prior-Override: Causal Mechanisms of Perception-Knowledge Conflict in Vision-Language Models

    cs.CL 2026-06 conditional novelty 7.0

    VLMs default to visual grounding but a sparse circuit of 2.5-4.8% attention heads in later layers mediates prior-knowledge overrides, identified causally via patching and ablation across three model families.

  3. Output Vector Editing for Memorization Mitigation in Large Language Models

    cs.CL 2026-06 unverdicted novelty 7.0

    Output vector editing on MLP neurons suppresses memorization in LLMs up to 87.9% on 6831 sequences in OLMo-7B with a 2.7x gap over zero ablation, ensemble covering 96.5%.

  4. TimeROME-DLM: Temporal Causal Tracing and Low-Rank Inference-Time Knowledge Editing for Masked Diffusion Language Models

    cs.LG 2026-06 unverdicted novelty 7.0

    TimeROME-DLM enables training-free knowledge editing in masked diffusion language models via temporal causal tracing and low-rank residual edit memory applied at inference time.

  5. Why Muon Outperforms Adam: A Curvature Perspective

    cs.LG 2026-06 conditional novelty 7.0

    Muon outperforms Adam by reducing curvature penalty via lower Normalized Directional Sharpness, as shown via Taylor approximation on LLM training and proven on stylized quadratic problems with heterogeneous curvature.

  6. EpiFormer: Learning Antigen-Antibody Interactions for Epitope Prediction via Geometric Deep Learning

    q-bio.QM 2026-06 unverdicted novelty 7.0

    EpiFormer improves epitope prediction F1 score by over 40% via early-fusion cross-attention in GNN layers and sparsity-aware objectives, while recovering known biology as emergent behavior.

  7. Query Lens: Interpreting Sparse Key-Value Features with Indirect Effects

    cs.LG 2026-05 unverdicted novelty 7.0

    Query Lens extends Logit Lens to interpret sparse features via key-value analysis and indirect effects, yielding coherent token signatures where Logit Lens fails, and proposes the Subspace Channel Hypothesis.

  8. Rethinking Visual Neglect: Steering via Context-Preference for MLLM Hallucination Mitigation

    cs.CL 2026-05 unverdicted novelty 7.0

    CAS mitigates object hallucinations in MLLMs by extracting two context preference vectors from designed conflict samples and applying signed residual injection at mid-early MLP layers without retraining or added latency.

  9. ConTact: Contact-First Antibody CDR Design via Explicit Interface Reasoning

    cs.LG 2026-05 unverdicted novelty 7.0

    ConTact decomposes CDR design into surface fingerprint learning, contact prediction, and contact-gated sequence generation using distance-biased attention and weighted loss, reporting 7% RMSD and 10% F1 gains on CHIME...

  10. ConTact: Contact-First Antibody CDR Design via Explicit Interface Reasoning

    cs.LG 2026-05 unverdicted novelty 7.0

    ConTact introduces a contact-then-act architecture with distance-biased cross-attention and contact-weighted loss for antibody CDR design, reporting 5-6% better backbone RMSD and superior contact metrics on CHIMERA-Be...

  11. Uncovering Entity Identity Confusion in Multimodal Knowledge Editing

    cs.CL 2026-05 unverdicted novelty 7.0

    Multimodal knowledge editing causes models to confuse original and edited entity identities in text queries by failing to update image-entity bindings and instead overfitting entity-entity shortcuts.

  12. Sharp Capacity Thresholds in Linear Associative Memory: From Winner-Take-All to Listwise Retrieval

    stat.ML 2026-05 unverdicted novelty 7.0

    Winner-take-all linear memory capacity scales as d² ~ n log n due to extreme values; listwise retrieval via Tail-Average Margin yields d² ~ n with exact asymptotic theory.

  13. Self-Attention as Transport: Limits of Symmetric Spectral Diagnostics

    cs.LG 2026-05 unverdicted novelty 7.0

    Transpose-invariant spectral diagnostics on attention operators are orientation-blind, and a φ-G two-axis diagnostic distinguishes hallucination modes with 0.62-0.84 LC-AUROC and predicted polarity reversal.

  14. How Language Models Process Negation

    cs.CL 2026-05 unverdicted novelty 7.0

    LLMs process negation using both attention-based suppression and constructive representation mechanisms (construction dominant), with late-layer attention shortcuts explaining poor accuracy on negation tasks.

  15. How Language Models Process Negation

    cs.CL 2026-05 unverdicted novelty 7.0

    LLMs implement both attention-based suppression and constructive representations for negation, with construction dominant, despite poor accuracy from late-layer attention shortcuts.

  16. A framework for analyzing concept representations in neural models

    cs.CL 2026-05 unverdicted novelty 7.0

    A new framework shows concept subspaces are not unique, estimator choice affects containment and disentanglement, LEACE works well but generalizes poorly, and HuBERT encodes phone info as contained and disentangled fr...

  17. A Parametric Memory Head for Continual Generative Retrieval

    cs.IR 2026-04 unverdicted novelty 7.0

    A product-key parametric memory head with selective sparse updates mitigates catastrophic forgetting in generative retrieval models during sequential addition of new documents.

  18. One Model to Translate Them All? A Journey to Mount Doom for Multilingual Model Merging

    cs.CL 2026-04 unverdicted novelty 7.0

    Merging fine-tuned models for multilingual translation fails because fine-tuning redistributes language-specific neurons rather than sharpening them, increasing representational divergence in output-generating layers.

  19. Norm Anchors Make Model Edits Last

    cs.LG 2026-01 conditional novelty 7.0

    Norm-Anchor Scaling breaks the norm-feedback loop in sequential LLM editing by anchoring value vectors to original norms, improving long-run performance by 72.2% and extending the editing horizon over 4x.

  20. Improving Dictionary Learning with Gated Sparse Autoencoders

    cs.LG 2024-04 unverdicted novelty 7.0

    Gated SAEs decouple which features to use from how large their activations should be, applying the L1 penalty only to selection and thereby eliminating shrinkage while halving the number of firing features needed for ...

  21. Eliciting Latent Predictions from Transformers with the Tuned Lens

    cs.LG 2023-03 accept novelty 7.0

    Training per-layer affine probes on frozen transformers yields more reliable latent predictions than the logit lens and enables detection of malicious inputs from prediction trajectories.

  22. Robust Harmful Features Under Jailbreak Attacks: Mechanistic Evidence from Attention Head Specialization in Large Language Models

    cs.CR 2026-06 unverdicted novelty 6.0

    Jailbreak attacks suppress Adversarially Compromised Heads in early layers but leave Safety-Aligned Heads robust in mid-layers, enabling competitive detection from persistent activations.

  23. VASAE: Naming SAE Dictionary Directions with Vocabulary-Aligned Anchoring

    cs.CL 2026-06 unverdicted novelty 6.0

    VASAE introduces vocabulary-aligned anchoring to train SAEs that yield features with intrinsic token names, reporting high alignment rates in early layers of GPT-2 and Llama-3.1 without reconstruction loss.

  24. LMs as Task-Specific Knowledge Bases: An Interpretability Analysis

    cs.CL 2026-06 unverdicted novelty 6.0

    LMs store facts in task-specific parameter subsets, shown by inconsistent emergence across tasks during training and distinct localized parameters for the same fact.

  25. Cross-Lingual Exploration for Parametric Knowledge

    cs.CL 2026-06 unverdicted novelty 6.0

    Cross-lingual prompt exploration improves factual recall and consistency in LLMs across 17 languages more efficiently than native-language scaling.

  26. Gated MLPs as Symmetry-Broken Rank-1 Bilinear Attention

    cs.LG 2026-06 unverdicted novelty 6.0

    Gated MLPs are shown to be symmetry-broken rank-1 bilinear attention mechanisms with query and key factors.

  27. Behavioral and Representational Evidence of Binomial Ordering Preferences in Large Language Models

    cs.CL 2026-06 unverdicted novelty 6.0

    LLMs recover dominant binomial orders from corpora but align less closely with exact preference distributions, with preference strength partially encoded in middle-to-late layers and manipulable via steering.

  28. Want Better Synthetic Data? Steer It: Activation Steering for Low-Resource Language Generation

    cs.CL 2026-06 unverdicted novelty 6.0

    Activation steering on early layers improves diversity of synthetic data for low-resource languages and often boosts downstream classifier performance compared to non-steered prompting.

  29. Variable-Width Transformers

    cs.CL 2026-06 conditional novelty 6.0

    ×-shaped variable-width transformers outperform parameter-matched uniform baselines on language modeling loss with 22% fewer FLOPs and 15% smaller KV cache.

  30. Understanding helpfulness and harmless tension in reward models

    cs.LG 2026-06 unverdicted novelty 6.0

    Mixed-objective reward models underperform single-objective ones because shared neurons support one objective while negatively affecting the other, creating alignment tension.

  31. Redesign Mixture-of-Experts Routers with Manifold Power Iteration

    cs.LG 2026-06 unverdicted novelty 6.0

    Manifold Power Iteration aligns MoE router rows with principal singular directions of experts via a power-then-retract process, with theory showing convergence and experiments on 1B-11B models showing gains.

  32. Substrate Asymmetry in User-Side Memory: A Diagnostic Framework

    cs.CL 2026-06 unverdicted novelty 6.0

    User memory in LLMs factors into three orthogonal axes where parametric adapters and retrieval show opposite strengths, with causal evidence from attention interventions and an alignment tax on RLHF models.

  33. Inside the LLM Word Factory

    cs.CL 2026-06 unverdicted novelty 6.0

    Activation patching localizes English detokenization in Llama2-7B to a two-stage attention-then-MLP process at layer 1 that generalizes to 12 models from 8 families, with depth varying by positional encoding, plus an ...

  34. Expert-Aware Causal Tracing of Factual Recall in Sparse MoE Language Models

    cs.CL 2026-06 unverdicted novelty 6.0

    Expert-aware causal tracing localizes factual recall to specific experts in some MoE models but requires coalitions in others, using CounterFact interventions on subject embeddings.

  35. Multi-component Causal Tracing in Large Language Models

    cs.LG 2026-06 unverdicted novelty 6.0

    A unified multi-component causal tracing method that uses soft interventions and a metric transformation to efficiently select critical LLM components for a target performance metric.

  36. Resonant Context Anchoring: Decoupling Attention Routing and Signal Gain at Inference Time

    cs.CL 2026-06 unverdicted novelty 6.0

    RCA is a training-free module that boosts input context signal strength in the residual stream of LLMs by orthogonal decoupling of attention routing from value magnitude.

  37. TriLens: Per-Layer Logit-Lens Entropy for White-Box Hallucination Detection

    cs.AI 2026-05 unverdicted novelty 6.0

    TriLens detects hallucinations via per-layer entropy trajectories of logit-lens readouts from three internal modules across LLMs and QA benchmarks.

  38. AgForce Enables Antigen-conditioned Generative Antibody Design

    cs.LG 2026-05 unverdicted novelty 6.0

    AgForce improves antigen-conditioned antibody design by using framework dropout, gated bottlenecks, hyperbolic cross attention, MDN sequence head with Potts-like coupling, annealed MCL, and antigen cycle consistency t...

  39. EvoStruct: Bridging Evolutionary and Structural Priors for Antibody CDR Design via Protein Language Model Adaptation

    cs.LG 2026-05 unverdicted novelty 6.0

    EvoStruct integrates evolutionary priors from a protein language model with structural priors from an E(3)-equivariant GNN to raise amino acid recovery by 16% and diversity by 2.3x on CHIMERA-Bench while cutting perpl...

  40. Focus-then-Context: Subject-Centric Progressive Visual Token Reduction for Vision-Language Models

    cs.CV 2026-05 conditional novelty 6.0

    SPpruner reduces visual tokens in VLMs via focus identification followed by context-aware scanning, retaining 22.2% tokens for 2.53x speedup on Qwen2.5-VL with negligible accuracy loss.

  41. 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.

  42. Learning to Foresee: Unveiling the Unlocking Efficiency of On-Policy Distillation

    cs.CL 2026-05 unverdicted novelty 6.0

    On-policy distillation gains efficiency from early foresight in module allocation and low-rank update directions, enabling EffOPD to accelerate training by 3x via adaptive extrapolation without extra modules or tuning.

  43. Learning to Foresee: Unveiling the Unlocking Efficiency of On-Policy Distillation

    cs.CL 2026-05 unverdicted novelty 6.0

    On-policy distillation gains efficiency from early foresight in module focus and update directions, enabling EffOPD to accelerate training 3x with comparable performance.

  44. A Geometric Perspective on Next-Token Prediction in Large Language Models: Three Emerging Phases

    cs.LG 2026-05 unverdicted novelty 6.0

    LLMs exhibit three geometric phases in next-token prediction—seeding multiplexing, hoisting overriding, and focal convergence—where predictive subspaces rise, stabilize, and converge across layers.

  45. UniPool: A Globally Shared Expert Pool for Mixture-of-Experts

    cs.LG 2026-05 unverdicted novelty 6.0

    A shared global expert pool in MoE improves validation loss over per-layer experts and allows sublinear expert-parameter growth with depth.

  46. Self-Attention as Transport: Limits of Symmetric Spectral Diagnostics

    cs.LG 2026-05 unverdicted novelty 6.0

    Symmetric spectral diagnostics on attention are structurally blind to flow direction, with asymmetry G as the sole control parameter, yielding a two-axis test that distinguishes bottleneck versus diffuse hallucination...

  47. Logical Consistency as a Bridge: Improving LLM Hallucination Detection via Label Constraint Modeling between Responses and Self-Judgments

    cs.CL 2026-05 unverdicted novelty 6.0

    LaaB improves LLM hallucination detection by mapping self-judgment labels back into neural feature space and using mutual learning under logical consistency constraints between responses and meta-judgments.

  48. The Override Gap: A Magnitude Account of Knowledge Conflict Failure in Hypernetwork-Based Instant LLM Adaptation

    cs.LG 2026-04 conditional novelty 6.0

    Knowledge conflicts in hypernetwork LLM adaptation stem from constant adapter margins losing to frequency-dependent pretrained margins; selective layer boosting and conflict-aware triggering raise deep-conflict accura...

  49. From Signal Degradation to Computation Collapse: Uncovering the Two Failure Modes of LLM Quantization

    cs.CL 2026-04 unverdicted novelty 6.0

    LLM 2-bit quantization fails via either cumulative signal degradation or early computation collapse in key components.

  50. Contrastive Attribution in the Wild: An Interpretability Analysis of LLM Failures on Realistic Benchmarks

    cs.AI 2026-04 conditional novelty 6.0

    Token-level contrastive attribution yields informative signals for some LLM benchmark failures but is not universally applicable across datasets and models.

  51. Representation-Guided Parameter-Efficient LLM Unlearning

    cs.CL 2026-04 unverdicted novelty 6.0

    REGLU guides LoRA-based unlearning via representation subspaces and orthogonal regularization to outperform prior methods on forget-retain trade-off in LLM benchmarks.

  52. BID-LoRA: A Parameter-Efficient Framework for Continual Learning and Unlearning

    cs.LG 2026-04 unverdicted novelty 6.0

    BID-LoRA uses bi-directional low-rank adapters with retain/new/unlearn pathways and escape unlearning to enable continual learning and unlearning while minimizing knowledge leakage and parameter updates.

  53. In-Place Test-Time Training

    cs.LG 2026-04 conditional novelty 6.0

    In-Place TTT adapts LLM MLP projection matrices at test time with a next-token-aligned objective and chunk-wise updates, enabling better long-context performance as a drop-in enhancement.

  54. Automated Attention Pattern Discovery at Scale in Large Language Models

    cs.LG 2026-04 unverdicted novelty 6.0

    AP-MAE reconstructs masked attention patterns in LLMs with high accuracy, generalizes across models, predicts generation correctness at 55-70%, and enables 13.6% accuracy gains via targeted interventions.

  55. The Illusion of Latent Generalization: Bi-directionality and the Reversal Curse

    cs.CL 2026-03 unverdicted novelty 6.0

    Bidirectional objectives mitigate reversal by requiring explicit source-as-target signals and storing directions as distinct representations instead of inducing latent generalization.

  56. Understanding Robustness of Model Editing in Code LLMs

    cs.SE 2025-11 unverdicted novelty 6.0

    A controlled benchmark on 2040 problems reveals poor generalization and high interference in model editing for API updates in code LLMs, with many successes being workarounds rather than true migrations.

  57. AtlasKV: Augmenting LLMs with Billion-Scale Knowledge Graphs in 20GB VRAM

    cs.CL 2025-10 unverdicted novelty 6.0

    AtlasKV integrates billion-scale KGs into LLMs parametrically with sub-linear complexity and low memory by converting triples into key-value representations handled by the model's attention.

  58. How Training Data Shapes the Use of Parametric and In-Context Knowledge in Language Models

    cs.CL 2025-09 unverdicted novelty 6.0

    Balanced parametric and in-context knowledge use in LLMs is an emergent property requiring intra-document repetition, moderate inconsistency, and skewed distributions in training data.

  59. RepIt: Steering Language Models with Concept-Specific Refusal Vectors

    cs.AI 2025-09 unverdicted novelty 6.0

    RepIt creates semantic backdoors in frontier language models by steering refusal vectors for specific concepts, allowing targeted unsafe responses while preserving safe scores on standard benchmarks.

  60. Provable Knowledge Acquisition and Extraction in One-Layer Transformers

    cs.LG 2025-07 unverdicted novelty 6.0

    In a stylized one-layer transformer, pre-training encodes factual knowledge via relation-specific feature directions and attention patterns; fine-tuning extracts it through a relation-covering mechanism that succeeds ...

Reference graph

Works this paper leans on

300 extracted references · 300 canonical work pages · cited by 65 Pith papers

  1. [1]

    Bastani and Y

    O. Bastani and Y. Ioannou and L. Lampropoulos and D. Vytiniotis and A. Nori and A. Criminisi , booktitle =. Measuring neural net robustness with constraints , year =

    work page

  2. [2]

    J. Z. Kolter and E. Wong , journal =. Provable defenses against adversarial examples via the convex outer adversarial polytope (published at

    work page

  3. [3]

    Wong and J

    E. Wong and J. Z. Kolter , booktitle =. Provable defenses against adversarial examples via the convex outer adversarial polytope , year =

    work page

  4. [4]

    Dvijotham and R

    K. Dvijotham and R. Stanforth and S. Gowal and T. Mann and P. Kohli , journal =. A Dual Approach to Scalable Verification of Deep Networks , year =

    work page

  5. [5]

    Hein and M

    M. Hein and M. Andriushchenko , booktitle =. Formal guarantees on the robustness of a classifier against adversarial manipulation , year =

    work page

  6. [6]

    A. A. Ahmadi and A. Majumdar , journal =

    work page

  7. [7]

    Dalvi and A

    N. Dalvi and A. Dasgupta and R. Kumar and V. Rastogi , booktitle =. Aggregating crowdsourced binary ratings , year =

    work page

  8. [8]

    Joglekar and H

    M. Joglekar and H. Garcia-Molina and A. Parameswaran , booktitle =. Comprehensive and reliable crowd assessment algorithms , year =

    work page

  9. [9]

    Zhang and X

    Y. Zhang and X. Chen and D. Zhou and M. I. Jordan , journal =. Spectral methods meet EM: A provably optimal algorithm for crowdsourcing , volume =

    work page

  10. [10]

    Balsubramani and Y

    A. Balsubramani and Y. Freund , booktitle =. Scalable semi-supervised aggregation of classifiers , year =

    work page

  11. [11]

    Craven and J

    M. Craven and J. Kumlien and others , booktitle =. Constructing biological knowledge bases by extracting information from text sources , year =

    work page

  12. [12]

    Socratic Learning: Augmenting Generative Models to Incorporate Latent Subsets in Training Data

    P. Varma and B. He and D. Iter and P. Xu and R. Yu and C. D. Sa and C. R. arXiv preprint arXiv:1610.08123 , title =

    work page Pith review arXiv

  13. [13]

    Shin and S

    J. Shin and S. Wu and F. Wang and C. D. Sa and C. Zhang and C. R. Incremental knowledge base construction using. Very Large Data Bases (VLDB) , number =

    work page

  14. [14]

    Roth and D

    B. Roth and D. Klakow , booktitle =. Combining Generative and Discriminative Model Scores for Distant Supervision , year =

    work page

  15. [15]

    Takamatsu and I

    S. Takamatsu and I. Sato and H. Nakagawa , booktitle =. Reducing wrong labels in distant supervision for relation extraction , year =

    work page

  16. [16]

    C. D. Sa and A. Ratner and C. R. Deepdive: declarative knowledge base construction , volume =. ACM SIGMOD Record , number =

    work page

  17. [17]

    Wu and L

    S. Wu and L. Hsiao and X. Cheng and B. Hancock and T. Rekatsinas and P. Levis and C. R. Proceedings of SIGMOD 2018 , title =

    work page 2018

  18. [18]

    Alfonseca and K

    E. Alfonseca and K. Filippova and J. Delort and G. Garrido , booktitle =. Pattern learning for relation extraction with a hierarchical topic model , year =

    work page

  19. [19]

    Bunescu and R

    R. Bunescu and R. Mooney , booktitle =. Learning to extract relations from the web using minimal supervision , year =

    work page

  20. [20]

    Parkash and D

    A. Parkash and D. Parikh , booktitle =. Attributes for classifier feedback , year =

    work page

  21. [21]

    Druck and B

    G. Druck and B. Settles and A. McCallum , booktitle =. Active learning by labeling features , year =

    work page

  22. [22]

    Raghavan and O

    H. Raghavan and O. Madani and R. Jones , booktitle =. InterActive Feature Selection , volume =

    work page

  23. [23]

    G. S. Mann and A. McCallum , journal =. Generalized expectation criteria for semi-supervised learning with weakly labeled data , volume =

    work page

  24. [24]

    MacCartney , howpublished =

    B. MacCartney , howpublished =. SippyCup , year =

    work page

  25. [25]

    D. H. Younger , journal =. Recognition and parsing of context-free languages in time n3 , volume =

    work page

  26. [26]

    A. J. Ratner and C. M. D. Sa and S. Wu and D. Selsam and C. R. Data programming: Creating large training sets, quickly , year =. Advances in Neural Information Processing Systems (NIPS) , pages =

    work page

  27. [27]

    B. S. H. and H. Bryan and R. Alexander and R. Christopher , booktitle =. Learning the Structure of Generative Models without Labeled Data , year =

    work page

  28. [28]

    Corney and D

    D. Corney and D. Albakour and M. Martinez-Alvarez and S. Moussa , booktitle =. What do a million news articles look like? , year =

    work page

  29. [29]

    Wei and Y

    C. Wei and Y. Peng and R. Leaman and A. P. Davis and C. J. Mattingly and J. Li and T. C. Wiegers and Z. Lu , booktitle =. Overview of the BioCreative

    work page

  30. [30]

    A. J. Ratner and S. H. Bach and H. Ehrenberg and J. Fries and S. Wu and C. R. Very Large Data Bases (VLDB) , title =

    work page

  31. [31]

    Srivastava and I

    S. Srivastava and I. Labutov and T. Mitchell , booktitle =. Joint concept learning and semantic parsing from natural language explanations , year =

    work page

  32. [32]

    Ling and S

    H. Ling and S. Fidler , booktitle =. Teaching Machines to Describe Images via Natural Language Feedback , year =

    work page

  33. [33]

    Li and A

    J. Li and A. H. Miller and S. Chopra and M. Ranzato and J. Weston , journal =. Learning Through Dialogue Interactions , year =

    work page

  34. [34]

    Andreas and D

    J. Andreas and D. Klein and S. Levine , journal =. Learning with Latent Language , year =

    work page

  35. [35]

    J. E. Weston , booktitle =. Dialog-based language learning , year =

    work page

  36. [36]

    L. V. Ahn and R. Liu and M. Blum , booktitle =. Peekaboom: a game for locating objects in images , year =

    work page

  37. [37]

    Krening and B

    S. Krening and B. Harrison and K. M. Feigh and C. L. Isbell and M. Riedl and A. Thomaz , journal =. Learning from explanations using sentiment and advice in

    work page

  38. [38]

    Guidotti and A

    R. Guidotti and A. Monreale and F. Turini and D. Pedreschi and F. Giannotti , journal =. A Survey Of Methods For Explaining Black Box Models , year =

    work page

  39. [39]

    Yessenalina and Y

    A. Yessenalina and Y. Choi and C. Cardie , booktitle =. Automatically generating annotator rationales to improve sentiment classification , year =

    work page

  40. [40]

    Arora and E

    S. Arora and E. Nyberg , booktitle =. Interactive annotation learning with indirect feature voting , year =

    work page

  41. [41]

    Grechkin and H

    M. Grechkin and H. Poon and B. Howe , journal =. EZLearn: Exploiting Organic Supervision in Large-Scale Data Annotation , year =

    work page

  42. [42]

    Ratinov and D

    L. Ratinov and D. Roth and D. Downey and M. Anderson , booktitle =. Local and Global Algorithms for Disambiguation to

    work page

  43. [43]

    Kalyanpur and B

    A. Kalyanpur and B. K. Boguraev and S. Patwardhan and J. W. Murdock and A. Lally and C. A. Welty and J. M. Prager and B. Coppola and A. Fokoue-Nkoutche and L. Zhang and Y. Pan and Z. M. Qui , journal =. Structured data and inference in DeepQA , volume =

    work page

  44. [44]

    Lee and P

    K. Lee and P. H. Seo and J. Choi and S. Koo and G. G. Lee , journal =. Conversational knowledge teaching agent that uses a knowledge base , year =

    work page

  45. [45]

    Han and J

    S. Han and J. Bang and S. Ryu and G. G. Lee , journal =. Exploiting knowledge base to generate responses for natural language dialog listening agents , year =

    work page

  46. [46]

    Ellis and J

    J. Ellis and J. Getman and H. Simpson and K. Griffitt and H. T. Dang and R. Grishman and H. Ji and C. DePrince and T. Riese and N. Kuster , journal =

    work page

  47. [47]

    J. A. Aslam and V. Pavlu and E. Yilmaz , booktitle =. A statistical method for system evaluation using incomplete judgments , year =

    work page

  48. [48]

    Buckley and D

    C. Buckley and D. Dimmick and I. Soboroff and E. Voorhees , booktitle =. Bias and the limits of pooling for large collections , year =

    work page

  49. [49]

    Buckley and E

    C. Buckley and E. M. Voorhees , booktitle =. Retrieval evaluation with incomplete information , year =

    work page

  50. [50]

    Sakai and N

    T. Sakai and N. Kando , booktitle =. On information retrieval metrics designed for evaluation with incomplete relevance assessments , year =

    work page

  51. [51]

    G. V. Cormack and C. R. Palmer and C. L. A. Clarke , booktitle =. Efficient Construction of Large Test Collections , year =

    work page

  52. [52]

    Yilmaz and E

    E. Yilmaz and E. Kanoulas and J. A. Aslam , booktitle =. A simple and efficient sampling method for estimating

    work page

  53. [53]

    Vannella and D

    D. Vannella and D. Jurgens and D. Scarfini and D. Toscani and R. Navigli , booktitle =. Validating and Extending Semantic Knowledge Bases using Video Games with a Purpose , year =

    work page

  54. [54]

    Pavlick and H

    E. Pavlick and H. Ji and X. Pan and C. Callison-Burch , booktitle =. The Gun Violence Database: A new task and data set for

    work page

  55. [55]

    W. E. Webber , school =. Measurement in Information Retrieval Evaluation , year =

    work page

  56. [56]

    Zobel , booktitle =

    J. Zobel , booktitle =. How reliable are the results of large-scale information retrieval experiments? , year =

    work page

  57. [57]

    E. M. Voorhees and D. Harman , booktitle =. Overview of the Eight Text REtreival Conference (

    work page

  58. [58]

    Adel and B

    H. Adel and B. Roth and H. Sch\". Human Language Technology and North American Association for Computational Linguistics (HLT/NAACL) , title =

    work page

  59. [59]

    A. B. Owen , publisher =. Monte Carlo theory, methods and examples , year =

    work page

  60. [60]

    K. S. Jones and C. V. Rijsbergen , journal =. Report on the Need for and Provision of an ``Ideal test collection , year =

    work page

  61. [61]

    D. K. Harman , journal =. The first text retrieval conference (TREC-1) Rockville, MD, U.S.A., 4-6 November, 1992 , volume =

    work page 1992

  62. [62]

    Ji and R

    H. Ji and R. Grishman and H. Text Analytics Conference , title =

    work page

  63. [63]

    R. L. Burden and J. D. Faires , publisher =. Numerical Analysis (3rd ed.) , year =

    work page

  64. [64]

    Liu and S

    A. Liu and S. Soderland and J. Bragg and C. H. Lin and X. Ling and D. S. Weld , booktitle =. Effective Crowd Annotation for Relation Extraction , year =

    work page

  65. [65]

    H. T. Dang , journal =. Cold Start Knowledge Base Population at

    work page

  66. [66]

    Ellis and J

    J. Ellis and J. Getman and D. Fore and N. Kuster and Z. Song and A. Bies and S. Strassel , journal =. Overview of linguistic resources for the

    work page

  67. [67]

    Ellis and X

    J. Ellis and X. Li and K. Griffitt and S. M. Strassel , journal =. Linguistic Resources for 2012 Knowledge Base Population Evaluations , year =

    work page 2012

  68. [68]

    Plank , journal =

    B. Plank , journal =. What to do about non-standard (or non-canonical) language in

    work page

  69. [69]

    Novikova and O

    J. Novikova and O. Du. Empirical Methods in Natural Language Processing (EMNLP) , title =

    work page

  70. [70]

    Lin and M

    C. Lin and M. Rey , booktitle =. Looking for a Few Good Metrics:

    work page

  71. [71]

    Cohan and N

    A. Cohan and N. Goharian , booktitle =. Revisiting Summarization Evaluation for Scientific Articles , year =

    work page

  72. [72]

    Lavie and M

    A. Lavie and M. Denkowski , journal =. The Meteor Metric for Automatic Evaluation of Machine Translation , volume =

    work page

  73. [73]

    Denkowski and A

    M. Denkowski and A. Lavie , booktitle =. Meteor Universal: Language Specific Translation Evaluation for Any Target Language , year =

    work page

  74. [74]

    Vedantam and C

    R. Vedantam and C. L. Zitnick and D. Parikh , booktitle =

    work page

  75. [75]

    G. A. Miller and J. G. Beebe-Center , journal =. Some Psychological Methods for Evaluating the Quality of Translations , volume =

    work page

  76. [76]

    J. H. Lau and A. Clark and S. Lappin , journal =. Grammaticality, Acceptability, and Probability: A Probabilistic View of Linguistic Knowledge , volume =

    work page

  77. [77]

    See and P

    A. See and P. J. Liu and C. D. Manning , booktitle =. Get To The Point: Summarization with Pointer-Generator Networks , year =

    work page

  78. [78]

    Paulus and C

    R. Paulus and C. Xiong and R. Socher , booktitle =. A Deep Reinforced Model for Abstractive Summarization , year =

    work page

  79. [79]

    Lin and M

    T. Lin and M. Maire and S. Belongie and J. Hays and P. Perona and D. Ramanan and P. Doll. Microsoft. European Conference on Computer Vision (ECCV) , pages =

    work page

  80. [80]

    J. M. Conroy and H. T. Dang , booktitle =. Mind the Gap : Dangers of Divorcing Evaluations of Summary Content from Linguistic Quality , year =

    work page

Showing first 80 references.