Why Fine-Tuned LLMs (SFT & LoRA) Fail to Reason

2026-07-12 · Source: Discover AI · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, extended

Summary

The article explores why fine-tuned Large Language Models (LLMs), particularly those using SFT & LoRA, struggle to reason with newly memorized facts, a problem called the "knowing-using gap." This gap manifests as an accuracy disparity and a temporal lag: LLMs quickly memorize facts (100% accuracy) but fail to apply them in multi-hop reasoning, often plateauing at ~20% performance despite extensive training. The core issue is not storage but a "knowledge circuit misalignment hypothesis." Memorized facts are stored in early or late layers, while reasoning circuits are in mid-layers. The vanishing gradient problem during fine-tuning prevents new knowledge from permeating these critical mid-layer reasoning circuits. Researchers from Hong Kong University of Science and Technology, in a paper published July 9, 2026, propose "self-patching," a diagnostic technique. This method injects activation vectors from memorization runs into mid-layers during generalization, recovering 55-75% of lost reasoning. This indicates a need for architectural improvements to internal communication, not just model scaling.

Key takeaway

For Machine Learning Engineers fine-tuning LLMs for complex tasks, recognize that current supervised fine-tuning methods may lead to a "knowing-using gap." Your models might memorize facts perfectly but fail to apply them in multi-hop reasoning due to internal architectural limits and vanishing gradients. Consider diagnostic tools like activation patching to pinpoint these failures. Prioritize research into new transformer architectures that improve internal knowledge propagation, rather than solely scaling model size.

Key insights

Fine-tuned LLMs memorize facts but fail to reason with them due to internal communication breakdowns, not storage.

Principles

• Knowing a fact is not equivalent to using it for reasoning in LLMs.
• LLM reasoning circuits are bottlenecked in mid-layers.
• Vanishing gradients hinder knowledge propagation to reasoning circuits.

Method

Self-patching involves caching activation vectors from memorization runs and injecting them into specific mid-layers during generalization runs to diagnose reasoning failures.

In practice

• Use activation patching to diagnose LLM reasoning failures.
• Focus future LLM architecture on internal communication, not just scale.

Topics

• Large Language Models
• Fine-tuning
• AI Reasoning
• Transformer Architecture
• Activation Patching
• Knowing-Using Gap

Best for: Research Scientist, AI Engineer, NLP Engineer, AI Scientist, Machine Learning Engineer, AI Student

Related on AIssential

• Towards Mechanistically Understanding Why Memorized Knowledge Fails to Generalize in Large Language Model Finetuning — LLMs' "Knowing--Using Gap" stems from memorized knowledge being misrouted internally, preventing generalization to reasoning tasks.
• Fine-Tuning LLMs, Part 2: SFT, LoRA, DPO, and the RL Methods That Power Reasoning Models — Fine-tuning and reinforcement learning are complementary stages for LLM specialization, not competing choices.
• BREAKING: LLM “reasoning” continues to be deeply flawed — LLMs, even those marketed for reasoning, continue to exhibit fundamental and widespread reasoning failures.
• GheGheGhe at SemEval-2026 Task 11: Decoupling Logic from Belief with Bias-Targeted Fine-Tuning and Neuro-Symbolic Syllogistic Reasoning — Decoupling LLM processing from logical reasoning via neuro-symbolic methods significantly enhances accuracy in semantic evaluation tasks.
• TUDUM: A Turkish-Thinking Reasoning Pipeline for Qwen3.5-27B — Adapting LLMs for deep linguistic reasoning requires explicit fine-tuning of internal thought processes, not just output translation.
• LLM Watch Weekly: When Scale Isn't Enough — Scaling alone cannot overcome data biases or fundamental reasoning gaps in LLMs and VLMs.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Discover AI.