Gradient Bottlenecks, Token Learning Dynamics, and Agentic Self-Evolution: Understanding LLM Training and Deployment

Welcome to today‘s edition of State of AI 👋

This edition covers some fundamental breakthroughs in understanding how language models actually learn and optimize. We‘re diving deep into the mechanics of LLM training, from the surprising discovery that language model output layers destroy 95% of gradient information during backpropagation, to new insights showing that token learning follows concentrated sigmoid patterns hidden beneath smooth scaling laws. Beyond foundation models, we‘re also seeing exciting progress in agentic systems that can self-evolve, make cost-aware decisions, and learn from their own failures in ways that mirror human cognitive processes like sleep and memory consolidation.

Here‘s what caught our attention:

• Failure as a Process: An Anatomy of CLI Coding Agent Trajectories — Treats agent failures as temporal processes rather than static outcomes, discovering that 57.9% of failures are epistemic (misusing available information) and that decisive errors occur 9 steps before they become observable, creating a critical intervention window.

• Lost in Backpropagation: The LM Head is a Gradient Bottleneck — Reveals that the output layer projects V-dimensional gradients through rank-D compression, destroying 95-99% of gradient information and causing ×16 convergence slowdowns independent of model architecture.

• Smooth Scaling Laws Hide Stepwise Token Learning — Decomposes scaling laws into individual token learning events, showing that tokens follow concentrated sigmoid trajectories and that the observed power-law structure emerges from how learning times are distributed across training.

• Language Models Need Sleep: Learning to Self-Modify and Consolidate Memories — Enables continuous post-deployment learning through biologically-inspired sleep phases that consolidate memories and generate synthetic training data, achieving 48.9% accuracy on new knowledge incorporation versus 46.7% for competing approaches.

• SAGEAgent: A Self-Evolving Agent for Cost-Aware Modality Acquisition in Multimodal Survival Prediction — Clinical agent that autonomously decides which diagnostic modalities to acquire, achieving 55% reduction in acquisition burden while maintaining prediction accuracy through dual-memory self-evolution without gradient updates.

• The LLMbda Calculus: AI Agents, Conversations, and Information Flow — First formally verified LLM agent harness with machine-checked security theorems in Lean, preventing prompt injection attacks through provenance-based information-flow control while maintaining utility parity with leading defense systems.

• Resample or Reroute? Budget-Aware Test-Time Model Selection for Large Language Models — Formalizes the runtime decision between generating multiple samples from one model versus switching to another model under fixed cost budgets, achieving 22-31% cost savings with 2-24 point accuracy improvements.

Let‘s get into it 👇

Bi-Weekly AI Research Roundup

Latest research summaries in ML, Robotics, CV, NLP and AI

Contents

1. Failure as a Process: An Anatomy of CLI Coding Agent Trajectories

2. SAGEAgent: A Self-Evolving Agent for Cost-Aware Modality Acquisition in Multimodal Survival Prediction

3. The LLMbda Calculus: AI Agents, Conversations, and Information Flow

4. Scalable Visual Pretraining for Language Intelligence

5. ALICE: Learning a General-Purpose Pathology Foundation Model from Vision, Vision-Language, and Slide-Level Experts

6. Promptable Concept Segmentation from Above: Evaluating SAM 3’s Zero-Shot and One-Shot Capabilities in Remote Sensing

7. Language Models Need Sleep: Learning to Self-Modify and Consolidate Memories

8. Resample or Reroute? Budget-Aware Test-Time Model Selection for Large Language Models

9. All Explanations are Wrong, But Many Are Useful: Exploring the Rashomon Explanation Set with Large Language Models

10. Lost in Backpropagation: The LM Head is a Gradient Bottleneck

11. Smooth Scaling Laws Hide Stepwise Token Learning

12. Neural Collapse Is Forbidden: Information Floors in Language Models

13. PAC-ACT: Post-training Actor-Critic for Action Chunking Transformers

14. B-spline Policy: Accelerating Manipulation Policies via B-spline Action Representations

15. Re$^3$Sim: Generating High-Fidelity Simulation Data via 3D-Photorealistic Real-to-Sim for Robotic Manipulation

Failure as a Process: An Anatomy of CLI Coding Agent Trajectories

Authors: Xiangxin Zhao, Han Li, Shuaiting Li, Tianyi Zhao, Earl T. Barr, Federica Sarro, He Ye

Source and references: https://arxiv.org/abs/2607.09510v1

---

Failure as a Process: An Anatomy of CLI Coding Agent Trajectories

Introduction

This paper presents the first large-scale empirical study treating coding agent failures as temporal processes rather than static outcomes. The researchers analyze 1,794 execution trajectories from seven frontier language models across three coding agent scaffolds to understand how failures emerge, evolve, and become unrecoverable in terminal-based environments.

Key Points

• Process-oriented framework: Introduces a three-timestamp decomposition (t_err, t_lock, t_obs) that captures when decisive errors occur, when they become unrecoverable, and when they become observable—moving beyond simple pass/fail metrics.

• Epistemic dominance: Demonstrates that 57.9% of failures stem from epistemic errors (misusing available information) rather than competence gaps, with false premises being the single largest failure trigger (30.7%).

• Early decisive errors: Reveals that median decisive errors occur at step 7, yet remain hidden until step 16—suggesting outcomes are determined much earlier than trajectory termination.

• Recovery as a process: Shows that 82% of failed trajectories continue executing after recovery becomes impossible, with “repairing the wrong problem” accounting for 39% of all wasted computation.

• Cross-system consistency: Finds that epistemic errors remain dominant across all 21 model-scaffold combinations (44-80%), though overall pass rates vary substantially (19-45%).

Methodology