🤖 AI

Repeated Shared Access Enables Grokking, but Edit Propagation Depends on a Fine-Grained Addressable Memory

This paper demonstrates that while repeated shared access enables models to overcome out-of-distribution generalization barriers ("grokking"), effective factual edit propagation specifically requires a fine-grained, addressable memory substrate that allows the forward computation to write and subsequently reread edited facts, a capability that loop recurrence alone provides only partially.

Original authors: Yanan Niu

Published 2026-06-23

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yanan Niu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are teaching a robot to solve a simple puzzle: "If Alice is Bob's brother, and Bob is Charlie's brother, who is Alice to Charlie?"

The paper investigates how different types of robot brains learn this puzzle and, more importantly, how well they update their answer when you tell them, "Actually, Alice is Bob's sister."

The researchers tested four different robot brain designs to see which ones could learn the puzzle and which ones could easily update their memory without breaking everything else.

The Four Brain Designs

Think of these four designs as different ways a student might study for a test:

The "One-Shot" Reader (Dense): This student reads the textbook once, from start to finish. They are great at memorizing the exact sentences in the book, but if you ask a tricky question that requires combining two facts they haven't seen together before, they get stuck.
The "Re-Reader" (Loop): This student reads the same short chapter over and over again (looping). By re-reading, they can eventually figure out the tricky combination of facts that the "One-Shot" reader missed.
The "Note-Taker" (Dense+Mem): This student reads the book once (like the first student) but has a special notebook where they write down key facts. They can look up facts in this notebook whenever they need them.
The "Re-Reading Note-Taker" (LMC): This student does both: they re-read the chapter and they have a notebook to look up facts.

The Big Discovery: Learning vs. Editing

The paper found a surprising split between learning the puzzle and editing the facts.

1. Learning the Puzzle (Grokking)
To learn the tricky puzzle (where you have to combine two steps), the robot needs repeated access to the information.

The "One-Shot" reader fails.
The "Re-Reader," the "Note-Taker," and the "Re-Reading Note-Taker" all succeed.
The Lesson: It doesn't matter how you get the repeated access (whether by re-reading the same text or looking up a note). As long as the brain gets to see the fact multiple times during the thinking process, it can learn the puzzle.

2. Editing the Facts (The Real Surprise)
Now, imagine you tell the robot: "Change the fact that 'Alice is Bob's brother' to 'Alice is Bob's sister'." You want the robot to update its answer for the whole puzzle automatically.

The "Re-Reader" (Loop) fails here. Even though they learned the puzzle, when you change one fact, the robot doesn't update the final answer correctly. It's like if you crossed out a word in a book you were re-reading; the robot keeps re-reading the old version in its head because it doesn't have a specific place to write the new version.
The "Note-Takers" (Dense+Mem and LMC) succeed perfectly. Because they have a specific, addressable notebook (a memory store), you can go to the exact page and line where "Alice" is written, cross it out, and write "Sister." When the robot re-reads the puzzle, it sees the new note immediately and updates the answer.
The "One-Shot" reader also fails to update, simply because it never figured out the puzzle in the first place.

The Core Metaphor: The Whiteboard vs. The Echo

The paper uses a great distinction between Shared Access and Addressable Memory.

The Echo (Loop): Imagine a room with a microphone and a speaker. If you say "Alice is Bob's brother," the room echoes it back to you. If you want to change it, you have to shout over the echo. The echo is messy; you can't easily isolate just that one sentence to change it without messing up the whole room's sound. This is why the "Re-Reader" brain struggles to edit facts.
The Whiteboard (Memory): Imagine a whiteboard where you write "Alice is Bob's brother." If you want to change it, you walk up to the board, find that specific spot, erase it, and write "Sister." The rest of the room stays quiet. This is why the "Note-Taker" brains are so good at editing.

The "Fine-Grained" Detail

The researchers also found that the quality of the whiteboard matters.

If the whiteboard is huge and has tiny, specific spots for every single fact (Fine-Grained), you can change one fact without accidentally smudging the facts next to it.
If the whiteboard is small and crowded, changing one fact might accidentally erase or change a neighbor's fact.
The paper showed that the "Re-Reading Note-Taker" with a fine-grained whiteboard could change a fact with 100% success and almost zero damage to unrelated facts.

Summary of Findings

To Learn: You just need to see the information multiple times. Whether you do this by looping (re-reading) or using memory (looking up notes) doesn't matter; both work.
To Edit: You need a specific address. You need a place where the fact is stored that you can go to, change, and have the rest of the brain see that change later.
- Looping alone is not enough for editing because it's just a repeating echo, not a specific storage spot.
- Memory (a specific addressable store) is the key to successful editing.

In short: You can teach a robot to solve a puzzle by making it think about the facts over and over. But if you want to fix a mistake in that robot's memory later, you must give it a specific "sticky note" to write the correction on, rather than just making it repeat the old story.

Technical Summary: Repeated Shared Access Enables Grokking, but Edit Propagation Depends on a Fine-Grained Addressable Memory

Problem Statement

This paper investigates the mechanisms underlying factual edit propagation in compositional reasoning tasks. While recent work has established that both looped transformers and memory-augmented models can overcome the "grokking" barrier (delayed generalization to out-of-distribution, or OOD, compositions) where standard dense transformers fail, it remains unclear whether these architectures are functionally equivalent for editing. Specifically, if a model successfully learns a compositional task, does a localized edit to a single atomic fact propagate correctly through downstream compositions? The authors aim to determine whether the ability to propagate edits is a consequence of repeated shared access (a common ingredient for learning) or if it requires a specific addressable memory substrate.

Methodology

The study employs a controlled synthetic Knowledge Graph (KG) QA setting (based on Wang et al. [26]) involving atomic facts and 2-hop compositions. The authors utilize a 2×2 experimental grid crossing two factors:

Backbone Type: Dense (single-pass) vs. Looped (repeated recomputation with tied weights).
Memory Access: None vs. Shared Memory (explicit store queried between iterations).

This yields four architectures:

Dense: Standard transformer, no memory, no loop.
Loop: Looped transformer, no external memory.
Dense+Mem: Dense backbone with shared memory (no loop).
LMC (Loop-Memory Coupling): Looped backbone with shared memory.

Experimental Protocol:

Study 1 (Learning): Measures "grokking" (the onset of OOD accuracy > 0.1) to identify which architectures cross the generalization barrier. Auxiliary controls include a single-shared-FFN memory and a partially shared loop to isolate the necessity of sparse routing and full recurrence.
Study 2 (Editing): Uses a common edit-propagation protocol. A successful direct edit (flipping a single-hop prediction) is applied to each substrate. Propagation is measured by whether the 2-hop composition correctly updates to reflect the new bridge fact.
Localization: Causal patching and interchange interventions are used to verify that facts are stored in specific, dominant memory sites and reused during composition.
Diagnostics: A "HOLDANS" (Hold-Answer-Subspace) interchange diagnostic is used to measure the timing of fact injection and the size of the reuse window.

Key Contributions

1. Repeated Shared Access as the Enabling Principle for Grokking

The paper confirms that repeated shared access is the common ingredient enabling OOD generalization (grokking), regardless of the specific implementation.

Both Loop (repeated recomputation) and LMC/Dense+Mem (repeated memory rereading) successfully cross the OOD barrier.
Dense models (no sharing) fail to generalize.
Auxiliary controls show that sparse MoE routing is not strictly necessary for grokking (a single shared FFN suffices), nor is full weight tying (partial sharing works). The critical factor is the ability to reuse a fact-bearing substrate across iterations.

2. Localization of Facts to Addressable Sites

In the LMC architecture, the study demonstrates that atomic facts are localized to dominant memory experts and reused during composition.

Single-Store Regime: For 199/199 tested atoms, the dominant expert at the single-hop step (step0) is identical to the dominant expert used when the fact serves as a bridge in a 2-hop query (step1).
Causal Control: Suppressing the dominant expert collapses recall, while suppressing others does not. Bridge-substitution interventions confirm that the composition relies on iterative single-hop lookups rather than direct 2-hop memorization.

3. Addressability Determines Edit Propagation

The central finding is a dissociation between learning competence and editing affordance. While repeated access enables learning, addressable memory is required for robust edit propagation.

Memory-Bearing Substrates (LMC, Dense+Mem): Exhibit strong edit propagation (0.71–0.96). A localized edit to a single fact successfully updates the 2-hop composition.
Memory-Free Substrates (Loop, Dense): Exhibit weak or non-existent propagation.
- Loop: Shows intermediate propagation (0.04–0.30). It shares computation but lacks an explicit addressable site to target for editing.
- Dense: Near-zero propagation (0.00–0.03).
Statistical Separation: The split between memory-bearing and memory-free cells is statistically significant (Mann–Whitney $p=0.008$ ), while the two memory-bearing cells are statistically indistinguishable ( $p=0.55$ ).

4. Precision of LMC Editing

Editing in LMC is highly precise.

One-Row Edits: Modifying only the single largest activated value row ( $m=1$ ) at the dominant memory site achieves 100% direct success and mean 0.989 propagation on internal probes.
Locality: This sparse edit minimizes collateral damage to unrelated facts (~0.1% movement). However, "co-resident" facts (those sharing the same expert) can leak in proportion to hidden-state similarity, supporting a superposition view of locality.

Results Summary

Architecture	Backbone	Memory	Grokking (OOD)	Edit Propagation (Strong)
Dense	Dense	None	No	0.00 – 0.03
Loop	Looped	None	Yes	0.04 – 0.30 (Intermediate)
Dense+Mem	Dense	Shared	Yes	0.71 – 0.96 (Strong)
LMC	Looped	Shared	Yes	0.78 – 0.92 (Strong)

Injection Window: The "HOLDANS" diagnostic suggests a coarse ordering: LMC injects facts early into an addressable site with a wide reuse window; Loop injects late with a narrow window; Dense has no reliable window. However, this diagnostic is a substrate-level interpretation and does not predict propagation at the seed level (e.g., within Loop, high HOLDANS scores do not correlate with high propagation).

Significance and Claims

The paper claims to provide a controlled mechanism study (a "model organism" for edit propagation) rather than a new architecture proposal. Its significance lies in:

Dissociating Learning from Editing: It demonstrates that the mechanisms enabling a model to learn compositional tasks (repeated shared access) are distinct from those enabling a model to edit those tasks (fine-grained addressability).
Refining the Role of Loops: Looped transformers are shown to be an intermediate, partial case for editing. They share computation (enabling learning) but lack the explicit addressable site required for clean, targeted editing.
The Necessity of Addressability: Successful edit propagation depends on whether the substrate exposes a fine-grained addressable memory that the forward computation can write to and later reread.
Granularity Dependence: The paper notes that the editing advantage of memory scales with the fineness of the addressable store. While the primary results use fine-grained routing ( $N=128$ ), a coarse-grained test ( $N=13$ ) showed that the advantage persists but attenuates, suggesting that "addressability" is the core factor, though its efficacy is modulated by routing granularity.

The authors explicitly state that these findings are derived from a controlled synthetic setting and do not claim to generalize directly to natural-language pretraining or retrieval-augmented systems without further validation. The work establishes a mechanistic precondition: for edit propagation to occur, the model must store facts in an addressable, in-loop memory that is reread by normal computation.