Pith. sign in

REVIEW 18 cited by

Not yet reviewed by Pith; the record is open.

This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.

SPECIMEN: schema-true, not a live event

T0 review · schema-true

One-sentence machine reading of the paper's core claim.

pith:XXXXXXXX · record.json · timestamp

arxiv 2203.14680 v3 pith:VBXS3Q2E submitted 2022-03-28 cs.CL

Transformer Feed-Forward Layers Build Predictions by Promoting Concepts in the Vocabulary Space

classification cs.CL
keywords vocabularycomputationconceptsdistributionfeed-forwardlayersmodelsprediction
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved
0 comments
read the original abstract

Transformer-based language models (LMs) are at the core of modern NLP, but their internal prediction construction process is opaque and largely not understood. In this work, we make a substantial step towards unveiling this underlying prediction process, by reverse-engineering the operation of the feed-forward network (FFN) layers, one of the building blocks of transformer models. We view the token representation as a changing distribution over the vocabulary, and the output from each FFN layer as an additive update to that distribution. Then, we analyze the FFN updates in the vocabulary space, showing that each update can be decomposed to sub-updates corresponding to single FFN parameter vectors, each promoting concepts that are often human-interpretable. We then leverage these findings for controlling LM predictions, where we reduce the toxicity of GPT2 by almost 50%, and for improving computation efficiency with a simple early exit rule, saving 20% of computation on average.

discussion (0)

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

Forward citations

Cited by 18 Pith papers

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

  1. The Spectral Lifecycle of Transformer Training: Transient Compression Waves, Persistent Spectral Gradients, and the Q/K--V Asymmetry

    cs.LG 2026-04 unverdicted novelty 8.0

    Transformer weight spectra exhibit transient compression waves that propagate layer-wise, persistent non-monotonic depth gradients in power-law exponents, and Q/K-V asymmetry, with the spectral exponent alpha predicti...

  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. Steering Language Models With Activation Engineering

    cs.CL 2023-08 unverdicted novelty 7.0

    Activation Addition steers language models by adding contrastive activation vectors from prompt pairs to control high-level properties like sentiment and toxicity at inference time without training.

  4. 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.

  5. Mass-Editing Memory in a Transformer

    cs.CL 2022-10 conditional novelty 7.0

    MEMIT scales direct memory editing in transformers from single facts to thousands of associations by optimizing MLP weight updates.

  6. Analysis-by-Proxy: Localization Signals in VLMs Operating as Condition Encoders

    cs.CV 2026-07 conditional novelty 6.0

    A lightweight Q-Former proxy trained on VLM hidden states reveals that localization signals peak in input-dependent intermediate layers, not the final layers used by standard editing pipelines.

  7. RepSelect: Robust LLM Unlearning via Representation Selectivity

    cs.CL 2026-06 unverdicted novelty 6.0

    RepSelect isolates forget-set-specific representations via gradient PCA collapse to achieve 4-50x better post-relearning robustness than baselines across multiple models and forget categories.

  8. 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.

  9. 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.

  10. 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.

  11. Feature Starvation as Geometric Instability in Sparse Autoencoders

    cs.LG 2026-05 unverdicted novelty 6.0

    Adaptive elastic net SAEs (AEN-SAEs) mitigate feature starvation in SAEs by combining ℓ2 structural stability with adaptive ℓ1 reweighting, producing a Lipschitz-continuous sparse coding map that recovers global featu...

  12. 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.

  13. How Do Language Models Compose Functions?

    cs.CL 2025-10 conditional novelty 6.0

    LLMs solve compositional factual recall either by computing intermediates or directly, with mechanism choice correlated to translation geometry in embedding spaces.

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

    cs.CL 2026-05 unverdicted novelty 5.0

    On-policy distillation gains efficiency from early foresight in module allocation and update directions, which the proposed EffOPD method exploits for 3x faster training with comparable performance.

  15. Hessian-Enhanced Token Attribution (HETA): Interpreting Autoregressive LLMs

    cs.CL 2026-04 unverdicted novelty 5.0

    HETA is a new attribution framework for decoder-only LLMs that combines semantic transition vectors, Hessian-based sensitivity scores, and KL divergence to produce more faithful and human-aligned token attributions th...

  16. FocusDiT: Masking Queries in Diffusion Transformers for Fine-grained Image Generation

    cs.CV 2026-06 unverdicted novelty 4.0

    FocusDiT masks non-critical query tokens before they enter the FFN in DiT models, directing capacity toward complex visual details and reporting improved text-to-image results.

  17. GeoMathCode: Understanding Interleaved Math-Code Reasoning for Geometry Problem Solving

    cs.CL 2026-05 unverdicted novelty 4.0

    GeoMathCode interleaves math reasoning with programmatic code outputs for geometry problems in MLLMs and shows that reasoning steps and hierarchical code structures become disentangled in latent space after SFT.

  18. Network Edge Inference for Large Language Models: Principles, Techniques, and Opportunities

    cs.DC 2026-04 unverdicted novelty 3.0

    A survey synthesizing challenges, system architectures, model optimizations, deployment methods, and resource management techniques for large language model inference at the network edge.