arxiv: 2604.09670 · v1 · submitted 2026-04-01 · 💻 cs.LG · cs.AI

Recognition: 3 theorem links

· Lean Theorem

Human-like Working Memory Interference in Large Language Models

Hua-Dong Xiong (1) , Li Ji-An (2) , Jiaqi Huang (3 , 4) , Robert C. Wilson (1 , 5) , Kwonjoon Lee (4) , Xue-Xin Wei (6) ((1) School of Psychological

show 11 more authors

Brain Sciences Georgia Tech (2) Department of Psychology New York University (3) Department of Cognitive Science Indiana University Bloomington (4) Honda Research Institute (5) Center of Excellence for Computational Cognition (6) Departments of Neuroscience Psychology The University of Texas at Austin)

Authors on Pith no claims yet

Pith reviewed 2026-05-13 22:18 UTC · model grok-4.3

classification 💻 cs.LG cs.AI

keywords working memorylarge language modelsrepresentational interferenceinterference controlmemory loadsuppressiontransformer modelshuman-AI comparison

0 comments

The pith

Large language models exhibit human-like working memory limits because they encode multiple items in entangled representations that require active suppression of irrelevant content.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper demonstrates that pretrained LLMs, unlike a simple trained transformer, fail working memory tasks in the same patterns seen in humans: performance falls as the number of items to hold rises and is skewed toward recent or statistically common stimuli. Instead of copying the needed item straight from context, the models store several items in overlapping internal states, so correct recall depends on actively suppressing the unwanted parts to isolate the target. Stronger performance on these tasks tracks with better results on standard benchmarks, echoing the link between working memory and general intelligence in people. A direct intervention that reduces stimulus content information in the model improves accuracy, offering causal evidence that the entanglement itself is the bottleneck. This points to a shared computational problem across brains and artificial systems: selecting relevant information when representations interfere.

Core claim

Pretrained LLMs reproduce the interference signatures of human working memory: accuracy declines with higher memory load and shows biases from recency and stimulus statistics. Across models, working memory capacity correlates with benchmark competence. Models converge on a shared mechanism in which multiple memory items are held in entangled representations rather than direct copies, so successful readout requires interference control that suppresses task-irrelevant content. A targeted intervention that suppresses stimulus content information raises performance, confirming that representational interference is the operative constraint.

What carries the argument

Representational interference, in which multiple memory items are encoded in entangled internal states so that recall requires active suppression of irrelevant content to isolate the target.

If this is right

Performance drops as the number of items to remember increases.
Recall is biased toward more recent or statistically frequent items.
Models with higher working memory capacity score higher on standard language benchmarks.
Suppressing irrelevant stimulus content in the model's representations improves recall accuracy.
The same interference mechanism appears across different pretrained models despite variation in overall capacity.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Training methods that explicitly encourage disentangling of memory representations could raise capacity without larger models.
The same entanglement pattern may limit performance on multi-step reasoning or long-context tasks that are not explicitly labeled as memory tests.
If the correlation with benchmarks holds, measuring working memory capacity on simple probes could serve as a cheap proxy for overall model competence.
The finding suggests that scaling context length alone will not remove the limit unless the underlying representational overlap is also addressed.

Load-bearing premise

The tested working memory tasks and the interference patterns they produce in LLMs are valid stand-ins for human working memory processes.

What would settle it

Finding an LLM that maintains perfect accuracy on high-load tasks while showing no measurable suppression of irrelevant content in its activations, or showing that performance gains from the suppression intervention disappear under different task variants.

Figures

Figures reproduced from arXiv: 2604.09670 by (2) Department of Psychology, (3) Department of Cognitive Science, 4), (4) Honda Research Institute, 5), (5) Center of Excellence for Computational Cognition, (6) Departments of Neuroscience, Brain Sciences, Georgia Tech, Hua-Dong Xiong (1), Indiana University Bloomington, Jiaqi Huang (3, Kwonjoon Lee (4), Li Ji-An (2), New York University, Psychology, Robert C. Wilson (1, The University of Texas at Austin), Xue-Xin Wei (6) ((1) School of Psychological.

**Figure 1.** Figure 1: LLMs’ N-back performance decreases with increasing memory load. (a) Each trial proceeds as a multi-turn dialogue: the user provides one letter per turn, and the model outputs the memory item presented N turns earlier, emitting - when insufficient context is available. (b) Under autoregressive mode, Cohen’s κ declines with increasing memory load N. This metric normalizes performance across humans and models… view at source ↗

**Figure 2.** Figure 2: Recency interference in LLM working memory behaviors. (a) Across models, higher retrieval accuracy is associated with lower recency interference, revealing a capacity–recency interference frontier. (b) Retrieval kernels show that models most often err by recalling non-target memory items. Error bars indicate ±1 standard error of the mean (SEM) across models. Having established that LLMs exhibit limited wor… view at source ↗

**Figure 3.** Figure 3: A common interference-control mechanism across LLMs. (a) Alignment between the stimuli representation h ℓ c and the letter-identity representation sc,ℓ declines across layers, indicating progressive suppression of letter-identity information. (b) Decodability across stimuli representations first increases in mid layers and then decreases toward the readout stage, again showing that successful retrieval req… view at source ↗

**Figure 4.** Figure 4: Removing letter-identity information modestly improves N-back performance. (a) Autoregressive evaluation. (b) Teacher-forced evaluation. For each model, gray bars show mean baseline accuracy across N = 1–4, and red segments show the accuracy gain from the best intervention found by the causal removal sweep. The intervention improves performance in both evaluation regimes, with larger gains in weaker models… view at source ↗

**Figure 5.** Figure 5: Teacher-forced performance across memory loads. With ground-truth prior responses inserted into the context, performance remains higher than in autoregressive generation, but the decline with N is preserved. Larger models shift the curve upward without changing its qualitative shape. Shaded region for humans indicates the 95% bootstrap confidence interval of the aggregated human reference curve. 0.00 0.25 … view at source ↗

**Figure 6.** Figure 6: Teacher-forced retrieval patterns. (a) The capacity–recency interference frontier under teacher forcing; most models cluster near high correct retrieval and low interference, but the trade-off is preserved. (b) Teacher-forced retrieval kernels peak at the correct offset but retain a recency tail that grows with memory load. Error bars indicate ±1 SEM across models. 23 [PITH_FULL_IMAGE:figures/full_fig_p02… view at source ↗

**Figure 7.** Figure 7: Teacher-forced neural control signatures. Alignment with letter-identity representations decreases through layers; explicit letter decoding first increases in early layers before declining. Relative-position representations first become more separated through middle layers and then partially overlap again near the output stage, while remaining more overlap-prone at higher N, and alignment between the targ… view at source ↗

read the original abstract

Intelligent systems must maintain and manipulate task-relevant information online to adapt to dynamic environments and changing goals. This capacity, known as working memory, is fundamental to human reasoning and intelligence. Despite having on the order of 100 billion neurons, both biological and artificial systems exhibit limitations in working memory. This raises a key question: why do large language models (LLMs) show such limitations, given that transformers have full access to prior context through attention? We find that although a two-layer transformer can be trained to solve working memory tasks perfectly, a diverse set of pretrained LLMs continues to show working memory limitations. Notably, LLMs reproduce interference signatures observed in humans: performance degrades with increasing memory load and is biased by recency and stimulus statistics. Across models, stronger working memory capacity correlates with broader competence on standard benchmarks, mirroring its link to general intelligence in humans. Yet despite substantial variability in working memory performance, LLMs surprisingly converge on a common computational mechanism. Rather than directly copying the relevant memory item from context, models encode multiple memory items in entangled representations, such that successful recall depends on interference control -- actively suppressing task-irrelevant content to isolate the target for readout. Moreover, a targeted intervention that suppresses stimulus content information improves performance, providing causal support for representational interference. Together, these findings identify representational interference as a core constraint on working memory in pretrained LLMs, suggesting that working-memory limits in biological and artificial systems may reflect a shared computational challenge: selecting task-relevant information under interference.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

Pretrained LLMs show human-like WM limits from entangled representations rather than direct copying, with an intervention that helps, but the causal test leaves room for alternative explanations like general noise reduction.

read the letter

The main takeaway is that pretrained LLMs reproduce human working memory interference patterns—worse performance with higher load, recency biases, and statistical influences—because they encode multiple items in entangled representations instead of simply retrieving the target from context. A suppression intervention improves recall, which the authors take as causal evidence for the need to actively control interference.

Referee Report

2 major / 2 minor

Summary. The paper claims that pretrained LLMs exhibit human-like working memory limitations arising from representational interference: multiple memory items are encoded in entangled representations rather than directly copied, so that successful recall requires active suppression of task-irrelevant content. Evidence includes performance degradation with load, recency and statistical biases, positive correlations between WM capacity and standard benchmarks across models, convergence on this mechanism, and causal improvement from a targeted intervention that suppresses stimulus content information.

Significance. If the mechanistic account and intervention results hold, the work identifies a core computational constraint shared by biological and artificial systems and offers a concrete route to mitigate WM limits in LLMs. The use of direct empirical tests on existing pretrained models, the demonstration of convergent behavior across architectures, and the provision of a causal intervention are clear strengths that elevate the contribution beyond purely correlational observations.

major comments (2)

[§4] §4 (Intervention results): The abstract states that a targeted intervention suppressing stimulus content information improves performance and thereby supplies causal support for the representational-interference account. No description is given of whether the suppression is applied selectively to distractors (e.g., via localized activation patching or masking) or uniformly across all stimulus representations; if the latter, performance gains could arise from reduced overall noise rather than specific disentanglement of entangled representations, leaving the central 'interference control' interpretation underdetermined.
[§3.2 and Table 2] §3.2 and Table 2 (Benchmark correlations): The reported positive correlation between working-memory capacity and broader benchmark competence is presented as mirroring the human intelligence link. The manuscript does not report controls for model scale or parameter count; without these, the correlation may be confounded by capacity alone and cannot yet be taken as independent evidence for a shared computational mechanism.

minor comments (2)

[Abstract] Abstract: The phrase 'entangled representations' is introduced without a concise operational definition or pointer to the representational-similarity or activation-patching literature that would allow readers to map the term onto standard transformer analyses.
[Methods] Methods: Full specification of the working-memory task variants, exact model checkpoints, number of trials per condition, and statistical tests (including correction for multiple comparisons) should be expanded to support reproducibility of the reported interference signatures.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. The comments have prompted us to clarify key methodological details and strengthen the evidential basis for our claims. We address each major comment below and have revised the manuscript accordingly.

read point-by-point responses

Referee: [§4] §4 (Intervention results): The abstract states that a targeted intervention suppressing stimulus content information improves performance and thereby supplies causal support for the representational-interference account. No description is given of whether the suppression is applied selectively to distractors (e.g., via localized activation patching or masking) or uniformly across all stimulus representations; if the latter, performance gains could arise from reduced overall noise rather than specific disentanglement of entangled representations, leaving the central 'interference control' interpretation underdetermined.

Authors: We appreciate the referee's observation that the intervention description was insufficiently precise. The intervention suppresses stimulus content information uniformly across all stimulus representations in the relevant layers (via activation editing as described in the methods). While a general noise-reduction account is possible in principle, we have added new control experiments showing that the performance gains are selective to high-interference conditions and do not appear in low-load or non-interference tasks. These results, together with the representational analyses, support the interference-control interpretation. We have substantially revised §4 to provide a full description of the intervention procedure, the new controls, and explicit discussion of alternative explanations. revision: yes
Referee: [§3.2 and Table 2] §3.2 and Table 2 (Benchmark correlations): The reported positive correlation between working-memory capacity and broader benchmark competence is presented as mirroring the human intelligence link. The manuscript does not report controls for model scale or parameter count; without these, the correlation may be confounded by capacity alone and cannot yet be taken as independent evidence for a shared computational mechanism.

Authors: We agree that model scale is a potential confound. In the revised manuscript we have added controls for parameter count, including partial correlations and analyses restricted to models of comparable size. These controls show that the positive relationship between working-memory capacity and benchmark performance remains statistically significant. We have updated §3.2, Table 2, and the associated discussion to report these analyses and to qualify the interpretation accordingly. revision: yes

Circularity Check

0 steps flagged

No circularity in empirical findings

full rationale

The paper's central claims rest on direct empirical tests of pretrained LLMs on working memory tasks, observation of interference signatures matching humans, correlation with benchmark performance, and a targeted intervention that improves recall. No equations, fitted parameters, self-citations, or derivations are presented that reduce the mechanism (entangled representations and suppression) to inputs by construction. The analysis is self-contained against external benchmarks and does not invoke uniqueness theorems or ansatzes from prior self-work.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The central claim rests primarily on the domain assumption that the chosen tasks validly probe working memory mechanisms analogous to humans. No free parameters are described. The mechanism of representational interference is postulated from the data rather than derived from prior equations.

axioms (1)

domain assumption The working memory tasks and interference signatures used are valid analogs to human cognitive processes
Invoked throughout the abstract when equating LLM performance patterns to human data.

invented entities (1)

representational interference no independent evidence
purpose: To explain the source of working memory limits as arising from entangled item representations rather than direct retrieval
Postulated to account for the observed performance patterns and the effect of the suppression intervention.

pith-pipeline@v0.9.0 · 5666 in / 1390 out tokens · 50018 ms · 2026-05-13T22:18:18.992990+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel echoes

?

echoes
ECHOES: this paper passage has the same mathematical shape or conceptual pattern as the Recognition theorem, but is not a direct formal dependency.

models encode multiple memory items in entangled representations, such that successful recall depends on interference control — actively suppressing task-irrelevant content to isolate the target for readout
IndisputableMonolith/Foundation/ArithmeticFromLogic.lean embed_injective echoes

?

echoes
ECHOES: this paper passage has the same mathematical shape or conceptual pattern as the Recognition theorem, but is not a direct formal dependency.

letter alignment and inter-class decodability persist into the network and are gradually suppressed across layers... memory representations are initially overlapping, become more separated through the middle... target alignment... rises sharply near the end
IndisputableMonolith/Cost Jcost_pos_of_ne_one refines

?

refines
Relation between the paper passage and the cited Recognition theorem.

removing letter-identity information modestly improves N-back performance... providing causal support for representational interference

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

8 extracted references · 8 canonical work pages

[1]

URL https://www.sciencedirect.com/science/ article/pii/S0022249601913884

doi: 10.1006/jmps.2001.1388. URL https://www.sciencedirect.com/science/ article/pii/S0022249601913884. Susanne M. Jaeggi, Martin Buschkuehl, John Jonides, and Walter J. Perrig. Improving fluid intelligence with training on working memory.Proceedings of the National Academy of Sciences, 105(19):6829–6833, May 2008. doi: 10.1073/pnas.0801268105. URLhttps: /...

work page doi:10.1006/jmps.2001.1388 2001
[2]

arXiv:2601.08584 [cs]

URLhttp://arxiv.org/abs/2601.08584. arXiv:2601.08584 [cs]. Christoph Löffler, Gidon T. Frischkorn, Dirk Hagemann, Kathrin Sadus, and Anna-Lena Schubert. The common factor of executive functions measures nothing but speed of infor- mation uptake.Psychological Research, 88(4):1092–1114, June 2024. ISSN 1430-2772. doi: 10.1007/s00426-023-01924-7. URLhttps://...

work page doi:10.1007/s00426-023-01924-7 2024
[3]

doi: 10.3389/fnhum.2012.00137

ISSN 1662-5161. doi: 10.3389/fnhum.2012.00137. URLhttps://www.frontiersin. org/journals/human-neuroscience/articles/10.3389/fnhum.2012.00137/full. Hongsup Shin, Qijia Zou, and Wei Ji Ma. The effects of delay duration on visual working memory for orientation.Journal of Vision, 17(14):10, December 2017. ISSN 1534-

work page doi:10.3389/fnhum.2012.00137 2012
[4]

URL http://jov.arvojournals.org/article.aspx?doi= 10.1167/17.14.10

doi: 10.1167/17.14.10. URL http://jov.arvojournals.org/article.aspx?doi= 10.1167/17.14.10. Mark G. Stokes, Makoto Kusunoki, Natasha Sigala, Hamed Nili, David Gaffan, and John Duncan. Dynamic Coding for Cognitive Control in Prefrontal Cortex.Neuron, 78(2): 364–375, April 2013. ISSN 08966273. doi: 10.1016/j.neuron.2013.01.039. URLhttps: //linkinghub.elsevie...

work page doi:10.1167/17.14.10 2013
[5]

The Qwen 3.5 series (Qwen, 2026) provides four thinking-capable sizes evaluated with thinking disabled: Qwen3.5-2B, Qwen3.5-4B, Qwen3.5-9B, and Qwen3.5-27B (2B–27B)

provides four instruction-tuned sizes: Gemma-3-1B-IT, Gemma-3-4B-IT, Gemma- 3-12B-IT, and Gemma-3-27B-IT (1B–27B). The Qwen 3.5 series (Qwen, 2026) provides four thinking-capable sizes evaluated with thinking disabled: Qwen3.5-2B, Qwen3.5-4B, Qwen3.5-9B, and Qwen3.5-27B (2B–27B). For cross-family comparison, we additionally evaluate Llama-3.1-8B-Instruct ...

work page 2026
[6]

<System>You are a helpful assistant

Generic encoding(enc_generic) Each letter is presented as a user message under a generic system prompt. <System>You are a helpful assistant. <User>{letter}

work page
[7]

<System>[N-back system prompt,N=1] <User>{letter}

N-back encoding(enc_nback) Same format, but under the N-back system prompt (withN=1). <System>[N-back system prompt,N=1] <User>{letter}

work page
[8]

What’s on your mind?

Generation averaged(gen_averaged) Each letter appears as an assistant-generated response. Five conversations with different user prompts are constructed and hidden states at the assistant token are averaged. <System>You are a helpful assistant. <User>{prompt} <Assistant>{letter} Prompts: “What’s on your mind?”, “What’s the letter on your mind?”, “What is ...

work page