Tagged articles

mechanistic interpretability

11 articles · Page 1 of 1

Nov 5, 2026 · Artificial Intelligence

Neural Networks Are Surprisingly Interpretable: New Research Challenges Black Box Assumptions

Former OpenAI researcher Nick Cammarata's paper 'Neural Networks Are Already Clean' reveals that neural network interpretability is far higher than previously believed: Qwen3-32B shows 90% of neurons have a single primary function, while modern vision architectures like ConvNeXt achieve 99-100% neuron interpretability without scaling model size or data, suggesting earlier chaos stemmed from architectural and tool limitations.

ConvNeXtLLM interpretabilityNick Cammarata

0 likes · 5 min read

Neural Networks Are Surprisingly Interpretable: New Research Challenges Black Box Assumptions

Machine Learning Algorithms & Natural Language Processing

Nov 3, 2026 · Artificial Intelligence

MIT's GPT-6 Astra Edits Qwen Weights Directly, Boosting Math Skills 6x Without Training

MIT researchers demonstrate GPT-6 Astra directly modifying Qwen3.5-0.8B's weights, improving MATH-500 scores from 7.9% to 48.1% (6x) by altering only 0.016% of parameters without gradient descent, and even writing neural network weights from scratch for games, outperforming PPO baselines 20x.

GPT-6 AstraLarge Language ModelsMIT research

0 likes · 9 min read

MIT's GPT-6 Astra Edits Qwen Weights Directly, Boosting Math Skills 6x Without Training

PaperAgent

Nov 2, 2026 · Artificial Intelligence

MIT's GPT-6 Astra Rewrites Qwen Weights Directly, Achieving 6x Math Improvement

MIT CSAIL's weight-space programming lets GPT-6 Astra directly edit neural weights without training, fixing Qwen3.5-0.8B's reasoning loops by modifying 0.016% of parameters to boost math 6x and instruction following 3x, while also handwriting game networks that surpass PPO, though limited to unblocking existing capabilities rather than creating new ones.

GPT-6 AstraLarge Language ModelsMIT CSAIL

0 likes · 10 min read

MIT's GPT-6 Astra Rewrites Qwen Weights Directly, Achieving 6x Math Improvement

Data Party THU

Oct 6, 2026 · Industry Insights

Is CS Academia Dead? AI PhDs Face Existential Crisis as Foundation Models Reshape Research

UT Austin PhD student Cao Jin questions whether computer science academia is obsolete as foundation models absorb specialized research, arguing that incremental SOTA-chasing papers are becoming obsolete and researchers must either join industry to build models or pivot to fundamental science like mechanistic interpretability.

AI ResearchFoundation ModelsPhD career

0 likes · 9 min read

Is CS Academia Dead? AI PhDs Face Existential Crisis as Foundation Models Reshape Research

Data Party THU

Oct 1, 2026 · Artificial Intelligence

Stanford & Tsinghua Find Two Neuron Types Encoding Reward in LLMs

Researchers from Stanford and Tsinghua identify sparse value and dopamine neurons in LLMs that encode state value and temporal difference errors, forming a causal reward subsystem enabling confidence estimation and process reward modeling for improved reasoning.

LLM interpretabilityconfidence estimationdopamine neurons

0 likes · 7 min read

Stanford & Tsinghua Find Two Neuron Types Encoding Reward in LLMs

Machine Heart

Sep 5, 2026 · Artificial Intelligence

How Mechanist Lets AI Discover Its Own Cognitive Mechanisms

Mechanist automates mechanistic interpretability research by generating hypotheses, running causal interventions on language models, discovering distinct attention heads for belief-state reasoning, and enabling targeted steering that improves both reasoning and biological sequence generation.

AI Self-ResearchAttention HeadsBelief State Reasoning

0 likes · 12 min read

How Mechanist Lets AI Discover Its Own Cognitive Mechanisms

Machine Heart

Jun 30, 2026 · Artificial Intelligence

Is There Really a Unique Mechanism in LLMs? Rethinking Functional Anisotropy

A recent ICML 2026 paper disproves the long‑held assumption that each task in a large language model is supported by a single, unique circuit, showing through overlap‑aware sheaf repulsion that many structurally dissimilar, sparse sheafs can achieve identical performance across multiple benchmarks, and proposing a distributive dense circuit hypothesis to explain this non‑uniqueness.

Large Language Modelscircuit discoverydistributed dense circuit

0 likes · 15 min read

Is There Really a Unique Mechanism in LLMs? Rethinking Functional Anisotropy

Machine Heart

Jun 28, 2026 · Artificial Intelligence

Which Training Data Shapes Large‑Model Abilities? Introducing Mechanistic Data Attribution (MDA)

The paper presents Mechanistic Data Attribution, a framework that traces the origins of specific internal mechanisms such as induction heads to particular training samples, revealing that repetitive "garbage" data—not high‑quality text—drives their emergence, and validates this causal link through deletion and augmentation experiments while enabling scalable data‑driven model improvement.

Causal InterventionData AugmentationInduction Heads

0 likes · 12 min read

Which Training Data Shapes Large‑Model Abilities? Introducing Mechanistic Data Attribution (MDA)

Machine Learning Algorithms & Natural Language Processing

Jun 18, 2026 · Artificial Intelligence

Why LLMs Miss Simple Addition: Geometric Mechanism Behind Arithmetic Errors

A recent ICML 2026 paper from Nanjing University reveals that large language models encode correct arithmetic information in structured geometric manifolds, yet errors arise from noisy quantization at decision boundaries, and proposes probing, Iso‑Raw‑Sum Trajectory, and a dual‑stream consistency check to diagnose and correct these mistakes.

Arithmetic ErrorsDual-Stream ConsistencyIso-Raw-Sum Trajectory

0 likes · 11 min read

Why LLMs Miss Simple Addition: Geometric Mechanism Behind Arithmetic Errors

Machine Heart

Jun 17, 2026 · Artificial Intelligence

Why Large Language Models Miss Simple Addition: Iso‑Raw‑Sum Trajectories Reveal the Geometry of Errors

Despite excelling at complex reasoning, LLMs often err on multi‑digit addition; probing shows correct answers reside in hidden states, and the authors reveal a structured geometric manifold—digit basins, carry fibers, and Iso‑Raw‑Sum trajectories—explaining how errors arise via noisy quantization at decision boundaries.

Arithmetic ErrorsGeometric AnalysisLLM

0 likes · 12 min read

Why Large Language Models Miss Simple Addition: Iso‑Raw‑Sum Trajectories Reveal the Geometry of Errors

Network Intelligence Research Center (NIRC)

Nov 11, 2025 · Artificial Intelligence

What Is Mechanistic Interpretability and Why It Matters for Large Language Models

The article defines mechanistic interpretability as reverse‑engineering LLMs to reveal how they represent knowledge and make decisions, explains its importance for transparency, risk mitigation, and model improvement, and surveys key techniques such as causal tracing, zero‑making, noise‑making, and logit‑lens methods with illustrative examples.

Large Language ModelsModel Editingcausal tracing

0 likes · 8 min read

What Is Mechanistic Interpretability and Why It Matters for Large Language Models