Unable to Forget: Proactive Interference Reveals Working Memory Limits in LLMs Beyond Context Length

Chupei Wang (University of Virginia); Jiaqiu Vince Sun (New York University)

REVIEW 3 major objections 4 minor 12 references

Unable to Forget: Proactive Interference Reveals Working Memory Limits in LLMs Beyond Context Length

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read LLM retrieval accuracy declines log-linearly toward zero as interfering prior updates accumulate, even when the target is the most recent value and is explicitly cued—a limit independent of context length.

desk verdict Solid, well-controlled benchmark for proactive interference in LLM retrieval, but the headline 'independent of context length' claim rests on a confounded experiment and needs an output-length control before it can stand. read the letter →

arxiv 2506.08184 v3 pith:2EHLVTSH submitted 2025-06-09 cs.CL cs.AIq-bio.NC

Chupei Wang (University of Virginia) , Jiaqiu Vince Sun (New York University) This is my paper

classification cs.CLcs.AIq-bio.NC

keywords proactiveinterferenceworkingmemoryLLMretrievalkey-valuetrackingendurancescorelong-contextevaluationlog-lineardecaypromptengineering

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish that LLMs possess a working-memory-like capacity limit, distinct from context length, which is exposed by proactive interference: when a stream of semantically similar key–value updates precedes a retrieval target, accuracy falls log-linearly toward zero even though the target is always the most recent value and is explicitly cued. It introduces PI-LLM, a synthetic key-update tracking task adapted from the paired-associate forgetting paradigm of human memory research, and reports the same log-linear decay across 35 models spanning 0.6B to over 600B parameters. The decline persists when input length is held constant, and error analysis shows models retrieve earlier overwritten values rather than the target. The paper concludes that resistance to interference is an independent model capability—correlated with parameter count, not context-window size—and that natural-language instructions to forget or focus, and chain-of-thought reasoning, do not remove the bottleneck.

What carries the argument

The load-bearing object is the proactive-interference (PI) retrieval task: an interleaved sequence of key–value updates over a fixed set of keys, with the instruction to return only the current, last value of each key, so every earlier value of the same key is a semantically similar distractor. The paper's quantitative index is the Interference Endurance Score (IES), defined as the area under the curve of retrieval accuracy against log-scaled update count, which ranks models by how slowly they degrade. The argument is carried by the error distribution: erroneous answers are dominated by previously overwritten values of the same key, shifting from adjacent positions to distant ones and eventually to unseen values as interference grows, which is presented as evidence that LLMs lack an unbinding or gating mechanism rather than a search failure.

What would settle it

Run the fixed-length setup while always asking the model to report a single key's final value, varying only the number of prior updates or updated keys; if accuracy no longer falls log-linearly with interference, the claimed isolation of interference from output load is falsified.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that LLM retrieval is constrained by a single, resource-like anti-interference capacity: accuracy declines approximately log-linearly toward zero as the number of updates per key, the number of tracked keys, or the length of each value grows, and this decline is not explained by input length. Because the task always asks for the last-presented value of a repeated key and states the target keys before and after the stream, search difficulty is deliberately minimized, so the failures measure interference directly. Error patterns evolve in three stages—confusion with nearby earlier values, then with distant earlier values, then hallucinated values never present in the input—which the paper interprets as the exhaustion of a working-memory buffer. Parameter size, not advertised context length, predicts the Interference Endurance Score, the area under the accuracy-versus-update-count curve. Explicit prompts to forget outdated updates or focus on recent ones give only marginal gains, chain-of-thought variants do not recover the lost accuracy, and only a mock-QA reset that inserts a hard context boundary substantially helps, though accuracy still declines with interference.

Load-bearing premise

That in the fixed-input-length experiment, varying the number of tracked keys changes only interference load; in fact it also changes how many values the model must output, so the observed decline could partly reflect generation load rather than retrieval interference.

Editorial extensions

If this is right

Length-based long-context benchmarks are measuring a different axis: extending context appears not to give a model more anti-interference capacity.
Model-size comparisons should account for interference load, since larger parameter counts predict slower log-linear decline while context-window length does not.
Prompt-level fixes such as 'ignore earlier updates' or 'focus on recent text' are insufficient, so improving interference tolerance likely needs architectural changes, for instance an explicit gating or reset mechanism.
Reasoning or chain-of-thought modes do not cure interference-affected retrieval, so allocating more inference compute will not by itself close this gap.
A single interference budget is taxed by three different manipulations—update count, tracked-key count, and value length—so retrieval reliability under distraction can be expected to degrade on any of these axes independently.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A control not run in the paper would query one key at a time while varying interference in a constant-length context; that would separate retrieval interference from the load of generating many answers.
The success of the mock-QA reset hints that any mechanism imposing a hard context boundary, such as a learned forget gate or a memory reset, might reproduce the gain, but it is unknown whether such gating can be selective rather than discarding all history.
The three-stage error shift from nearby values to hallucinations is a behavioral signature that could be tracked mechanistically, for example by probing attention or induction-circuit weights, to test whether old associations remain bound rather than being unbound.
Because the benchmark uses synthetic word categories, an extension to naturalistic streams such as patient vitals, versioned code state, or mutable ledgers would show whether the log-linear law survives realistic semantic overlap.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Two things to know. The core phenomenon is real and cleanly demonstrated: stream key–value updates, query only the final values, and accuracy falls log-linearly with update count across 35 models, with bootstrapped CIs and an error distribution showing the errors really are previously overwritten values. That is a solid, useful result. The second thing is that the headline claim—interference limits retrieval independently of context length—rests on an experiment with a confound, and the strongest version of that claim is not supported.

What's genuinely new: the A-B/A-C proactive-interference probe applied to LLM retrieval, with the target always the last value so search difficulty is minimized, is a real design improvement over Lost-in-the-Middle and the long-context benchmarks. The IES metric is a simple AUC summary and it does discriminate between models. The error-distribution analysis (localized to dispersed to hallucinatory) is descriptive but nice. The 'forget' prompts largely failing, and the mock-QA reset partially working, are useful empirical findings. Code and data are public.

Now the soft spots. Experiment B (Section 3.1.2) is the control meant to show the decline is independent of input length, and it varies T, the number of tracked keys, while holding the update stream fixed. But T is also the number of outputs the model must generate, and accuracy is measured across all T outputs. So the decline could be generation load—format drift, truncation, output-token pressure—rather than retrieval interference. The near-identical declines in Experiments A and B are consistent with a shared output-length confound. There is no control that holds output length constant while varying the number of tracked keys. The value-length experiment (Section 4) does hold the number of outputs fixed, so it shows at least one load dimension degrades retrieval without that confound, but it does not rescue the tracked-keys interpretation. That is a real gap, and it is fixable. The 'log-linear' law is never fitted—no slopes, R², or goodness-of-fit are reported—so it is an eye-fit. The 'single unified bottleneck' is an interpretation: the three manipulations give similar shapes but very different slopes, so 'one resource' is not established. The IES regression on model size and context length is a descriptive fit on the same curves that define IES, so it is a summary, not a predictive test.

Who this is for: people building long-context or retrieval evals. The benchmark is worth having, the update-count result is solid, and the weaknesses are addressable. I would send it to peer review—it deserves referee time—but a serious referee should ask for an output-length control, reported fits, and language that matches the evidence.

Referee Report

3 major / 4 minor

Summary. The paper introduces PI-LLM, a synthetic key-value update task in which models must report the most recent value for each tracked key after a stream of repeated updates. The central empirical finding is a roughly log-linear decline in retrieval accuracy as the number of interfering updates, tracked keys, or value tokens grows, across 35 models. The authors argue that the decline is not explainable by input length because a fixed-input-length variant shows the same pattern, and they report that larger models have higher Interference Endurance Scores (IES) while nominal context length does not predict IES. Prompt-based interventions such as directed forgetting are mostly ineffective, while a mock QA reset partially restores accuracy. The paper concludes that LLM retrieval is limited by a working-memory-like anti-interference capacity distinct from context length.

Significance. The task design is simple, reproducible, and well suited to isolating proactive interference in modern LLMs, and the coverage of 35 dense, MoE, and CoT models gives the benchmark broad descriptive value. The main update-count experiment controls the number of outputs (46 keys) and the output format, reports bootstrapped confidence intervals, and includes an error-location analysis showing that mistakes are often prior values; these are genuine strengths. Public code and data are promised, which supports reproducibility. If the independence-from-context-length claim survives an output-length control, the paper would establish a practically important evaluation axis. As it stands, the evidence supports an interference effect but not yet the stronger 'beyond context length' conclusion.

major comments (3)

[3.1.2 and 3.2.3] Experiment B varies the number of tracked keys T while holding input length fixed, but accuracy is measured as the fraction of correctly retrieved values across all T keys (Section 2.1), so output length and decoding load grow with T. Experiment A has the same property because the number of updated keys equals the number of queried keys. The nearly identical declines in Figure 9 therefore only rule out input length as the sole cause; they do not rule out generation-load effects such as format drift, response truncation, or attention dilution at decoding time. No control is reported that holds the number of required outputs constant while varying interference, and the error-location analysis in Figures 5 and 19 is performed on Experiment 1, where output length is fixed at 46 keys. Please add such a control or otherwise quantify the fraction of failures attributable to incomplete or malformed outputs in Experiments A and B.
[2.2 and 3.2.3] The 'log-linear decline' is a central quantitative claim, but no regression is reported for any of the curves in Figures 1, 3, or 9: there are no slopes, intercepts, R-squared values, or comparisons against alternative functional forms. Visual inspection of log-scaled axes is not sufficient to establish a universal log-linear law, especially since Figure 4 shows a step-like pattern for sequential ordering. Please fit per-model models such as accuracy ~ a + b log(update_count), and the equivalent for tracked keys and value length, and report goodness of fit and model comparisons.
[2.2.2] The regression of IES on parameter size class and context length is descriptive and may be underpowered: IES is the AUC of the same accuracy curves, the analysis uses N=30 after excluding reasoning and proprietary models, and context lengths cluster at 128k-131k. The current reporting of t and p for two predictors with R-squared = 0.261 does not address collinearity or the limited range of context lengths. I ask for a continuous-size regression where parameter counts are public, a check of variance inflation, and an explicit statement that this is a descriptive fit rather than a claim that context length is causally irrelevant.

minor comments (4)

[2.2.2] The text says IES is computed from the accuracy-versus-update-count function 'see Figure 5', but the IES curves are shown in Figure 16; the figure citation should be corrected.
[Appendix A] The caption of Figure 19 ends with 'See Figure 19 for additional models', which is circular and likely intended to refer to Figure 5 or Figure 18.
[Figure 12] The phrase 'key-value pair length of 6' is ambiguous; it should specify whether this is the value length in words or in tokens.
[References] There is a typo in the Introduction and reference list: 'V odrahalli' should be 'Vodrahalli'.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper reports direct empirical measurements with no derivation chain that reduces to its own inputs.

full rationale

The paper's central claims are empirical measurements: retrieval accuracy is measured on synthetic key-value streams across 35 models, and the observed log-linear decline is presented as a fitted description of those measurements, not as a prediction derived from a fitted parameter. The Interference Endurance Score (IES) is defined as the area under the same accuracy-versus-update-count curves used in the regression against model size and context length (Section 2.2.2); this is a summary statistic of measured data, not a prediction of an independent quantity, so it does not constitute circularity under the specified patterns. The fixed-length experiment (Experiment B, Section 3.1.2) may confound interference load with the number of required outputs, as the skeptic notes, but that is a potential validity threat rather than a circular derivation: the claim that interference is independent of input length is supported by external measurements, and no equation or definition makes the conclusion equivalent to the input. The paper contains no load-bearing self-citations: references to prior human working-memory work (e.g., Oberauer & Vockenberg, 2009; Festini & Reuter-Lorenz, 2014) are used only as analogies and test-design inspiration, not to justify the empirical result. There is no uniqueness theorem imported from the authors' own prior work, and no ansatz is smuggled in via citation. The finding is self-contained against the reported experiments, so the appropriate score is 0.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

The paper introduces no fitted model parameters and no new physical entities; its constructs (IES as AUC, anti-interference capacity) are descriptive summaries of measured curves. The cognitive analogy relies on prior human working-memory literature, listed as domain assumptions. Design choices such as 46 keys, log-spaced update counts, and 400-word categories are experimental settings, not free parameters fitted to support a claim.

assumptions (4)

domain assumption Placing the retrieval target as the last update of each key minimizes search difficulty, so accuracy loss is attributable to interference.
Section 2.1 design premise; underlies the central isolation claim.
domain assumption Human proactive-interference findings (performance plateau, benefit from directed forgetting) are a valid comparative baseline for LLM behavior.
Sections 5 and 9.2 rely on Oberauer and Vockenber 2009 and Festini and Reuter-Lorenz 2014.
domain assumption Resistance to proactive interference reflects working memory capacity in humans; applying the same logic to LLMs yields a 'working memory-like' capacity metric.
Introduction and Section 9.2; this is the interpretive bridge for the IES.
standard math Bootstrap confidence intervals, AUC, and Spearman correlation are standard statistical tools.
Used throughout for confidence intervals, IES computation, and the size regression.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Unable to Forget: Proactive Interference Reveals Working Memory Limits in LLMs Beyond Context Length." pith.science (2026). https://pith.science/paper/2EHLVTSH

@misc{pith2026250608184,
  author       = {Pith},
  title        = {Pith review of: Unable to Forget: Proactive Interference Reveals Working Memory Limits in LLMs Beyond Context Length},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2EHLVTSH}},
  note         = {Machine review of arXiv:2506.08184}
}

read the original abstract

Information retrieval in Large Language Models (LLMs) is increasingly recognized as intertwined with generation capabilities rather than mere lookup. While longer contexts are often assumed to improve retrieval, the effects of intra-context interference remain understudied. To address this, we adapt the proactive interference (PI) paradigm from cognitive science, where earlier information disrupts recall of newer updates. In humans, susceptibility to such interference is inversely linked to working memory capacity. We introduce PI-LLM, an evaluation that sequentially streams semantically related key-value updates and queries only the final values. Although these final values are clearly positioned just before the query, LLM retrieval accuracy declines log-linearly toward zero as interference accumulates; errors arise from retrieving previously overwritten values. Attempts to mitigate interference via prompt engineering (e.g., instructing models to ignore earlier input) yield limited success. These findings reveal a fundamental constraint on LLMs' ability to disentangle interference and flexibly manipulate information, suggesting a working memory bottleneck beyond mere context access. This calls for approaches that strengthen models' ability to suppress irrelevant content during retrieval.

Figures

Figures reproduced from arXiv: 2506.08184 by the authors.

**Figure 1.** Universal log-linear decline in retrieval performance due to interference. Increasing the amount of interfering information preceding a retrieval target within a language model’s input context results in a log-linear decrease in retrieval accuracy across diverse models. The target is positioned after the interfering information and explicitly referenced in the prompt to reduce search difficulty and isolate interfere… view at source ↗

**Figure 2.** Basic input example for the LLM-PI test. A continuous stream of key-value updates is presented to the model, with up to 46 updated keys and 400 distinct values (update count) used in the actual experiments. In this example, three distinct keys (”visual art”, ”tools”, and ”landform”)—color-coded for clarity—each undergo four updates. Numerical prefixes (e.g., ”1*”) denote update order for visualization purposes only… view at source ↗

**Figure 3.** Model retrieval accuracy declined approximately log-linearly as the number of updates per key increased. For visualization, models were grouped by estimated parameter size into four tiers—XS, S, M, and L—shown from top to bottom. Larger models (L group) tended to degrade more slowly, while smaller models (XS group) declined the fastest. The x-axis is log-scaled, covering update counts from 3 to 400 (log-spaced). The… view at source ↗

Figures from the paper (17 more)

**Figure 4.** Figure 4: Step-like failure pattern in sequential key–value update tests. Retrieval accuracy remains near-perfect as interfering information is added in strictly sequential order, until a modelspecific threshold is reached—after which performance drops abruptly to near-zero. W…

**Figure 5.** Figure 5: Distribution of model responses across update positions, showing increasing signs of PI as update count increases (left to right). The y-axis lists 11 equal-width bins (Bin 1–Bin 11, green) covering the entire update sequence. The earthy yellow bar indicates the single…

**Figure 6.** Figure 6: Interference Endurance Score (IES, from [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]

**Figure 7.** Figure 7: Comparison of retrieval accuracy between Mixture-ofExperts (MoE) and dense models. Each curve shows retrieval accuracy versus update count for a single model. MoE models are denoted by ”X” markers and labeled with “MoE” in the legend. Across update counts, MoE archite…

**Figure 8.** Figure 8: provides an example input: among 3 distinct keys updated in the sequence, only 2 are tracked (queried) at the end. As my secretary, I need you to carefully read a text stream where the values of multiple keys are being continuously updated. The 2 keys to track include …

**Figure 9.** Figure 9: Varying the number of updated keys (left panels) versus the number of tracked keys (right panels, with updated keys fixed at maximum) yields only minor differences in retrieval accuracy. In all conditions, accuracy declines approximately log-linearly with the number of…

**Figure 10.** Figure 10: Retrieval accuracy as a function of value length, showing a roughly log-linear decline toward near-zero performance. For clarity, models are grouped by parameter size: large models (L; ≥150B parameters) are shown in the upper panel, and smaller models (< 150B) in the…

**Figure 11.** Figure 11: Example input illustrating intervention strategies designed to mitigate proactive interference. Each strategy inserts explicit cues into the update stream, typically near the end (e.g., at the 120th-last update or one-third before the final update). The five strategie…

**Figure 12.** Figure 12: Explicit forgetting and focusing prompts inserted during the update stream (as shown in [PITH_FULL_IMAGE:figures/full_fig_p013_12.png]

**Figure 13.** Figure 13: A selective per-key forgetting prompt induces a distinct pattern of Proactive Interference (PI): instead of enabling successful forgetting, the prompt causes retrieval errors to cluster around the position in the update sequence where the instruction was injected. Th…

**Figure 14.** Figure 14: Chain-of-Thought (CoT) does not improve retrieval performance. Accuracy as a function of update count is shown for four pairs of CoT-enabled models and their corresponding base (nonCoT) versions. In three of the four comparisons, the CoT variant performs worse than …

**Figure 15.** Figure 15: Input example with manipulation of the updated values’s length. In this example, three items from the same category are space-removed, capitalized at the first letter, and concatenated into a single updated value. Bold text indicates the target key-value pairs the mo…

**Figure 16.** Figure 16: Interference Endurance Score (IES) for all models shown in [PITH_FULL_IMAGE:figures/full_fig_p018_16.png]

**Figure 17.** Figure 17: Interference Endurance Scores (IES) from [PITH_FULL_IMAGE:figures/full_fig_p019_17.png]

**Figure 18.** Figure 18: Explicit forgetting and focusing prompts inserted during the update stream (as shown in [PITH_FULL_IMAGE:figures/full_fig_p020_18.png]

**Figure 19.** Figure 19: Distribution of model responses across update positions, showing increasing signs of PI as update count increases (left to right). The y-axis lists 11 equal-width bins (Bin 1–Bin 11, green) covering the entire update sequence. The earthy yellow bar indicates the singl…

**Figure 20.** Figure 20: The selective per-key forgetting prompt amplifies proactive interference. Keys that received a forgetting instruction prior to the final third of their updates exhibited concentrated errors around the forgetting point, compared to keys without such a prompt. The x-axi…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

12 extracted references · 6 canonical work pages

[1]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...
[2]

In-context learning and induction heads, March 2022

Anthropic . In-context learning and induction heads, March 2022. URL https://www.anthropic.com/news/in-context-learning-and-induction-heads. Publisher: Anthropic

work page 2022
[3]

D., Thomson, N., and Buchanan, M

Baddeley, A. D., Thomson, N., and Buchanan, M. Word length and the structure of short-term memory. Journal of Verbal Learning and Verbal Behavior, 14 0 (6): 0 575--589, December 1975. doi:10.1016/S0022-5371(75)80045-4. URL https://doi.org/10.1016/S0022-5371(75)80045-4

work page doi:10.1016/s0022-5371(75)80045-4 1975
[4]

Festini, S. B. and Reuter-Lorenz, P. A. Cognitive control of familiarity: Directed forgetting reduces proactive interference in working memory. Cognitive, Affective, & Behavioral Neuroscience, 14 0 (1): 0 78--89, March 2014. ISSN 1531-135X. doi:10.3758/s13415-013-0231-1. URL https://doi.org/10.3758/s13415-013-0231-1

work page doi:10.3758/s13415-013-0231-1 2014
[5]

Working Memory Capacity of ChatGPT : An Empirical Study

Gong, D., Wan, X., and Wang, D. Working Memory Capacity of ChatGPT : An Empirical Study . Proceedings of the AAAI Conference on Artificial Intelligence, 38 0 (9): 0 10048--10056, March 2024. ISSN 2374-3468. doi:10.1609/aaai.v38i9.28868. URL https://ojs.aaai.org/index.php/AAAI/article/view/28868. Number: 9

doi:10.1609/aaai.v38i9.28868 2024
[6]

X., and Hasan, S

He, J., Rungta, M., Koleczek, D., Sekhon, A., Wang, F. X., and Hasan, S. Does Prompt Formatting Have Any Impact on LLM Performance ?, November 2024. URL http://arxiv.org/abs/2411.10541. arXiv:2411.10541 [cs] version: 1

arXiv 2024
[7]

Kuo, M., Zhang, J., Ding, A., Wang, Q., DiValentin, L., Bao, Y., Wei, W., Juan, D.-C., Li, H., and Chen, Y. H- CoT : Hijacking the Chain -of- Thought Safety Reasoning Mechanism to Jailbreak Large Reasoning Models , Including OpenAI o1/o3, DeepSeek - R1 , and Gemini 2.0 Flash Thinking , February 2025. URL http://arxiv.org/abs/2502.12893. arXiv:2502.12893 [...

arXiv 2025
[8]

F., Lin, K., Hewitt, J., Paranjape, A., Bevilacqua, M., Petroni, F., and Liang, P

Liu, N. F., Lin, K., Hewitt, J., Paranjape, A., Bevilacqua, M., Petroni, F., and Liang, P. Lost in the Middle : How Language Models Use Long Contexts . Transactions of the Association for Computational Linguistics, 12: 0 157--173, February 2024. ISSN 2307-387X. doi:10.1162/tacl_a_00638. URL https://doi.org/10.1162/tacl_a_00638

doi:10.1162/tacl_a_00638 2024

Show all 12 references

[9]

and Vockenberg, K

Oberauer, K. and Vockenberg, K. Updating of working memory: lingering bindings. Quarterly Journal of Experimental Psychology (2006), 62 0 (5): 0 967--987, May 2009. ISSN 1747-0226. doi:10.1080/17470210802372912

2006 doi
[10]

Introducing GPT -4.1 in the API , April 2025 a

OpenAI . Introducing GPT -4.1 in the API , April 2025 a . URL https://openai.com/index/gpt-4-1/. Publisher: OpenAI

2025
[11]

multiple-needle-in-a-haystack

OpenAI . OpenAI MRCR : Long -context “multiple-needle-in-a-haystack” benchmark, 2025 b . URL https://huggingface.co/datasets/openai/mrcr

2025
[12]

Michelangelo: Long Context Evaluations Beyond Haystacks via Latent Structure Queries , September 2024

Vodrahalli, K., Ontanon, S., Tripuraneni, N., Xu, K., Jain, S., Shivanna, R., Hui, J., Dikkala, N., Kazemi, M., Fatemi, B., Anil, R., Dyer, E., Shakeri, S., Vij, R., Mehta, H., Ramasesh, V., Le, Q., Chi, E., Lu, Y., Firat, O., Lazaridou, A., Lespiau, J.-B., Attaluri, N., and O...

2024 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

[2] [2]

In-context learning and induction heads, March 2022

Anthropic . In-context learning and induction heads, March 2022. URL https://www.anthropic.com/news/in-context-learning-and-induction-heads. Publisher: Anthropic

work page 2022

[3] [3]

D., Thomson, N., and Buchanan, M

Baddeley, A. D., Thomson, N., and Buchanan, M. Word length and the structure of short-term memory. Journal of Verbal Learning and Verbal Behavior, 14 0 (6): 0 575--589, December 1975. doi:10.1016/S0022-5371(75)80045-4. URL https://doi.org/10.1016/S0022-5371(75)80045-4

work page doi:10.1016/s0022-5371(75)80045-4 1975

[4] [4]

Festini, S. B. and Reuter-Lorenz, P. A. Cognitive control of familiarity: Directed forgetting reduces proactive interference in working memory. Cognitive, Affective, & Behavioral Neuroscience, 14 0 (1): 0 78--89, March 2014. ISSN 1531-135X. doi:10.3758/s13415-013-0231-1. URL https://doi.org/10.3758/s13415-013-0231-1

work page doi:10.3758/s13415-013-0231-1 2014

[5] [5]

Working Memory Capacity of ChatGPT : An Empirical Study

Gong, D., Wan, X., and Wang, D. Working Memory Capacity of ChatGPT : An Empirical Study . Proceedings of the AAAI Conference on Artificial Intelligence, 38 0 (9): 0 10048--10056, March 2024. ISSN 2374-3468. doi:10.1609/aaai.v38i9.28868. URL https://ojs.aaai.org/index.php/AAAI/article/view/28868. Number: 9

doi:10.1609/aaai.v38i9.28868 2024

[6] [6]

X., and Hasan, S

He, J., Rungta, M., Koleczek, D., Sekhon, A., Wang, F. X., and Hasan, S. Does Prompt Formatting Have Any Impact on LLM Performance ?, November 2024. URL http://arxiv.org/abs/2411.10541. arXiv:2411.10541 [cs] version: 1

arXiv 2024

[7] [7]

Kuo, M., Zhang, J., Ding, A., Wang, Q., DiValentin, L., Bao, Y., Wei, W., Juan, D.-C., Li, H., and Chen, Y. H- CoT : Hijacking the Chain -of- Thought Safety Reasoning Mechanism to Jailbreak Large Reasoning Models , Including OpenAI o1/o3, DeepSeek - R1 , and Gemini 2.0 Flash Thinking , February 2025. URL http://arxiv.org/abs/2502.12893. arXiv:2502.12893 [...

arXiv 2025

[8] [8]

F., Lin, K., Hewitt, J., Paranjape, A., Bevilacqua, M., Petroni, F., and Liang, P

Liu, N. F., Lin, K., Hewitt, J., Paranjape, A., Bevilacqua, M., Petroni, F., and Liang, P. Lost in the Middle : How Language Models Use Long Contexts . Transactions of the Association for Computational Linguistics, 12: 0 157--173, February 2024. ISSN 2307-387X. doi:10.1162/tacl_a_00638. URL https://doi.org/10.1162/tacl_a_00638

doi:10.1162/tacl_a_00638 2024