ForgetBench: Benchmarking Forgetting Dynamics of Long-Term Parametric Memory in Language Models

# ForgetBench: Benchmarking Forgetting Dynamics of Long-Term Parametric Memory in Language Models
Source: [https://arxiv.org/html/2607.26455](https://arxiv.org/html/2607.26455)
###### Abstract

Large language models \(LLMs\) have demonstrated strong capabilities in knowledge acquisition and reasoning, yet their ability to retain previously acquired knowledge under repeated updates remains insufficiently understood\. Existing evaluation paradigms primarily focus on single\-step reasoning or static knowledge editing, which fail to capture the temporal dynamics of knowledge retention and degradation during continual model modification\. In this work, we propose ForgetBench, a benchmark designed to systematically characterize forgetting behavior in LLMs under continual knowledge editing\. ForgetBench introduces two complementary evaluation paradigms, namely concept\-based QA and scenario\-based QA, to disentangle isolated factual retention from structured relational knowledge preservation\. Building upon a sequential editing framework, we construct temporally ordered knowledge streams and evaluate model behavior across multiple editing stages\. To quantitatively analyze long\-term retention dynamics, we further introduce a unified evaluation framework that models knowledge evolution over time, enabling the measurement of temporal decay, retention strength, and cross\-instance stability\. Extensive experiments across diverse models and editing methods demonstrate that existing approaches fail to strike a balance between long\-term retention and generalization quality\. Our findings highlight the need for more robust memory mechanisms that can effectively acquire, update, and preserve knowledge over time in future LLMs\. Code will be released upon acceptance\.

## 1Introduction

Memory is essential for intelligent systems operating in dynamic environments, where knowledge must be continuously acquired, retained, and updated over time\. Inspired by cognitive studies of memory\(Atkinson and Shiffrin[1968](https://arxiv.org/html/2607.26455#bib.bib36)\), memory is commonly divided into working memory and long\-term memory\. While working memory supports transient information manipulation within a single interaction, long\-term memory enables durable knowledge retention and gradual forgetting across extended periods\. For large language models \(LLMs\), understanding such long\-term memory dynamics is increasingly important as models are expected to adapt continuously after deployment\.

Despite the rapid progress of LLMs, most existing evaluation paradigms\(Yanget al\.[2018](https://arxiv.org/html/2607.26455#bib.bib43); Baiet al\.[2024](https://arxiv.org/html/2607.26455#bib.bib44)\)primarily focus on working memory abilities, such as in\-context reasoning over long inputs or multi\-hop question answering within a single inference process, rather than persistent knowledge retention after updates\. Recent studies have begun to investigate long\-term memory in interactive systems, including multi\-session conversational memory\(Wuet al\.[2024](https://arxiv.org/html/2607.26455#bib.bib54)\)and evolving agent memory evaluation\(Dinget al\.[2026](https://arxiv.org/html/2607.26455#bib.bib56)\)\. However, these benchmarks mainly evaluate externally maintained memory, where retention depends on retrieval mechanisms, memory organization, or interaction histories\. In contrast, our work focuses on intrinsic parametric memory, where knowledge is encoded into model parameters and forgetting emerges from cumulative interference caused by subsequent model updates\. Such a setting is particularly important for continual adaptation, where models repeatedly acquire new knowledge through parameter modification\.

Knowledge editing\(Menget al\.[2022b](https://arxiv.org/html/2607.26455#bib.bib28); Fanget al\.[2024](https://arxiv.org/html/2607.26455#bib.bib29)\)has recently emerged as a promising paradigm for updating factual knowledge in LLMs without full retraining\. Existing editing benchmarks, such as zsRE\(Levyet al\.[2017](https://arxiv.org/html/2607.26455#bib.bib30)\)and UnKE\(Denget al\.[2024](https://arxiv.org/html/2607.26455#bib.bib31)\), typically evaluate whether an individual edit succeeds through metrics including edit accuracy, generalization, and locality\. However, these evaluations treat edits as independent operations and overlook the temporal interaction among consecutive updates\. Consequently, they cannot answer a fundamental question for reliable lifelong adaptation: whether knowledge acquired today can survive future modifications\.

To address this gap, we proposeForgetBench, a benchmark for evaluating long\-term memory dynamics in LLMs under continual knowledge editing\. Different from existing editing evaluations that measure immediate editability, ForgetBench studies the temporal survivability of edited knowledge through sequential editing streams and repeated evaluation\. We construct two complementary settings: \(i\) concept\-based QA, which isolates atomic factual updates for controlled analysis of parameter interference, and \(ii\) scenario\-based QA, which introduces structured relational knowledge through multi\-agent interaction graphs\. By tracking model performance across editing steps, ForgetBench characterizes knowledge retention as forgetting curves, revealing how memories are preserved, degraded, or overwritten over time\.

Our main contributions can be summarized as follows: \(1\) We introduce ForgetBench, a unified benchmark designed to evaluate the long\-term parametric forgetting dynamics of LLMs under continual knowledge editing\. Moving beyond static, single\-step edit evaluations, ForgetBench enables temporal analysis of knowledge acquisition, retention, and forgetting throughout sequential update processes\. \(2\) We design two complementary QA construction paradigms that capture both isolated factual updates and structured relational knowledge\. Based on temporally ordered knowledge streams, we further develop a evaluation framework to quantify retention strength, temporal decay, and cross\-instance stability\. \(3\) Extensive experiments across multiple LLMs and editing methods demonstrate that existing editing approaches face challenges in maintaining both long\-term factual retention and generalization capability\. Our findings provide a systematic analysis of the temporal limitations of current knowledge editing paradigms and motivate future research toward more robust memory mechanisms\.

## 2Related Works

### 2\.1Working & Long\-Term Memory

Memory in cognitive science is commonly divided into short\-term and long\-term components\. Following the Atkinson\-Shiffrin memory model\(Atkinson and Shiffrin[1968](https://arxiv.org/html/2607.26455#bib.bib36)\), short\-term memory supports temporary information manipulation during ongoing processing, whereas long\-term memory enables durable storage, accumulation, and retrieval of knowledge over extended periods\. This distinction has inspired recent efforts to investigate whether language models can maintain information beyond a single inference context\.

Early studies on LLM memory have predominantly focused on working memory capabilities, including long\-context reasoning, context engineering\(Asaiet al\.[2023](https://arxiv.org/html/2607.26455#bib.bib38)\), key\-value reuse and compression\(Chevalieret al\.[2023](https://arxiv.org/html/2607.26455#bib.bib39); Liet al\.[2024](https://arxiv.org/html/2607.26455#bib.bib40)\), and chain\-of\-thought reasoning\(Weiet al\.[2022](https://arxiv.org/html/2607.26455#bib.bib41); Xuet al\.[2025a](https://arxiv.org/html/2607.26455#bib.bib42)\)\. These approaches improve the utilization of information temporarily available in the context window, but they do not examine whether newly acquired knowledge can be persistently stored in model parameters\. Similarly, retrieval\-augmented generation \(RAG\) and external memory mechanisms\(Lewiset al\.[2020](https://arxiv.org/html/2607.26455#bib.bib37)\)enhance knowledge access through external retrieval, where memory retention depends on storage, retrieval, or management strategies rather than intrinsic parameter updates\.

Recent benchmarks have extended memory evaluation toward long\-term interactive scenarios\. Existing datasets such as HotpotQA\(Yanget al\.[2018](https://arxiv.org/html/2607.26455#bib.bib43)\), LongBench\(Baiet al\.[2024](https://arxiv.org/html/2607.26455#bib.bib44)\), and LoCoMo\(Maharanaet al\.[2024](https://arxiv.org/html/2607.26455#bib.bib45)\)evaluate reasoning over complex contexts or conversational history, while newer benchmarks and studies further investigate memory consistency in multi\-session\(Wuet al\.[2024](https://arxiv.org/html/2607.26455#bib.bib54); Huet al\.[2026](https://arxiv.org/html/2607.26455#bib.bib55); Dinget al\.[2026](https://arxiv.org/html/2607.26455#bib.bib56)\)and evolving agent environments\(Uddinet al\.[2026](https://arxiv.org/html/2607.26455#bib.bib57); Zhanget al\.[2026](https://arxiv.org/html/2607.26455#bib.bib58)\)\. These studies provide important insights into how LLM\-based systems retrieve and organize information over time\. However, their memory evolution is primarily governed by external memory modules, retrieval policies, or interaction histories, and therefore the observed forgetting reflects failures in memory access or management rather than interference among internal model parameters\.

The forgetting curve provides another perspective for characterizing memory degradation\. Inspired by Ebbinghaus’ forgetting curve\(Ebbinghaus[2013](https://arxiv.org/html/2607.26455#bib.bib46)\), Liu et al\.\(Liuet al\.[2024](https://arxiv.org/html/2607.26455#bib.bib53)\)introduce forgetting curves to analyze memorization in long\-context language models\. However, their formulation defines memory age according to contextual distance, measuring the decay of information presented within an input sequence\. In contrast, ForgetBench focuses on intrinsic parametric memory under continual knowledge editing\. Instead of evaluating sequence\-level decay or externally maintained memory, ForgetBench defines memory age by the number of subsequent editing operations\. By constructing temporally ordered editing streams, it enables the analysis of long\-term stability, forgetting dynamics, and cumulative parameter interference beyond single\-step editing\.

### 2\.2Knowledge Editing

Knowledge editing aims to modify factual knowledge in large language models without full retraining\. Early methods based on causal intervention and direct weight modification\(Mitchellet al\.[2021](https://arxiv.org/html/2607.26455#bib.bib27)\)show that factual associations can be localized within model parameters and explicitly altered\. Building on this insight, scalable approaches such as MEMIT\(Menget al\.[2022b](https://arxiv.org/html/2607.26455#bib.bib28)\)enable large\-scale multi\-fact editing, while AlphaEdit\(Fanget al\.[2024](https://arxiv.org/html/2607.26455#bib.bib29)\)further constrains updates to the null space to mitigate interference with unrelated knowledge\. More recent work explores two main directions: hypernetwork\-based methods\(Tanet al\.[2023](https://arxiv.org/html/2607.26455#bib.bib34); Liet al\.[2025](https://arxiv.org/html/2607.26455#bib.bib32); Guet al\.[2026](https://arxiv.org/html/2607.26455#bib.bib33)\), which learn an auxiliary model to predict parameter updates conditioned on editing requests, and locate\-then\-edit approaches\(Fanget al\.[2024](https://arxiv.org/html/2607.26455#bib.bib29); Panet al\.[2025](https://arxiv.org/html/2607.26455#bib.bib35)\), which first identify knowledge\-critical components and then apply targeted perturbations via optimization\.

Evaluation is commonly conducted on benchmarks such as zsRE\(Levyet al\.[2017](https://arxiv.org/html/2607.26455#bib.bib30)\), CounterFact\(Menget al\.[2022a](https://arxiv.org/html/2607.26455#bib.bib26)\), and UnKE\(Denget al\.[2024](https://arxiv.org/html/2607.26455#bib.bib31)\), measuring edit success, paraphrase generalization, and preservation of non\-target knowledge\. However, these benchmarks are inherently static and treat each edit as an independent operation, which limits their ability to assess long\-term memory dynamics\. In contrast, our benchmark introduces temporally ordered edits, enabling direct measurement of retention, interference, and forgetting curves\. Therefore, existing editing benchmarks evaluate whether models can acquire new knowledge, while ForgetBench further investigates whether such knowledge can survive future updates over extended editing horizons\.

## 3Methodology

### 3\.1Overview of ForgetBench

As illustrated in Fig\.[1](https://arxiv.org/html/2607.26455#S3.F1), ForgetBench is composed of two complementary QA construction paradigms, namelyconcept\-based QAandscenario\-based QA, which differ in both the granularity and structural organization of knowledge\.

![Refer to caption](https://arxiv.org/html/2607.26455v1/x1.png)Figure 1:Overview of ForgetBench\. We construct two categories of QA sequences\. Concept\-based QA \(a\) focuses on isolated concepts, while scenario\-based QA \(b\) evaluates structured knowledge understanding\. Subsequently, sequential editing \(c\) exposes the model to these QA sequences and evaluates forgetting dynamics across editing stages\.The concept\-based setting focuses on atomic factual attributes, where each subject is associated with a single property and independently perturbed over time\. This yields fully disentangled instances for studying localized knowledge updates\. In contrast, the scenario\-based setting constructs QA pairs over structured relational knowledge induced by an interaction graph\. A knowledge graph is first generated through simulated agent\-item interactions, and QA instances are then derived from sampled subgraphs and transformed into natural language contexts, requiring reasoning over interconnected entities\. Given these two complementary constructions, ForgetBench evaluates models under a sequential editing paradigm, where new knowledge is continuously injected over time\. By periodically testing model performance over previously and newly edited instances, we obtain fine\-grained forgetting trajectories that reflect both immediate adaptation and long\-term memory degradation\.

### 3\.2Data Generation

Concept\-based QA\.The concept\-based setting constructs independent QA instances over atomic subjects\. Each subject is assigned an attribute, whose pre\-edit value is sampled from a predefined distribution and then replaced by an independently sampled post\-edit value as the editing target\. Each QA instance follows a fixed attribute query templateQQ, with the post\-edit attributeAr​i​g​h​tA^\{right\}serving as the ground\-truth answer and the pre\-edit valueAw​r​o​n​gA^\{wrong\}representing the overwritten memory \(e\.g\.,QQ:“How old is Donald Rodriguez?”,Ar​i​g​h​tA^\{right\}:“22”,Aw​r​o​n​gA^\{wrong\}:“79”\)\. Since all subjects are sampled independently, this setting removes inter\-instance dependencies and provides a controlled environment for analyzing localized knowledge updates\.

Scenario\-based QA\.As specified in Algo\.[1](https://arxiv.org/html/2607.26455#alg1), the scenario\-based setting constructs QA pairs over structured relational knowledge induced by a dynamic interaction graph\. We initialize a set of agents𝒜\\mathcal\{A\}and evolve their interactions through multiple simulation rounds, where active agents participate in collaborative behaviors that generate entities and relational edges\. This process yields a heterogeneous knowledge graph𝒢\\mathcal\{G\}consisting of agents, items, and their interactions\.

Algorithm 1Scenario\-based QA GenerationInput: Simulation roundsKK, subgraph sizeLL, number of subgraphsNN, questions per subgraphtt Output: QA sequence𝒟\\mathcal\{D\}

1:Initialize QA sequence

𝒟←∅\\mathcal\{D\}\\leftarrow\\emptysetand the length

T=N×tT=N\\times t
2:// Phase 1: KG Generation

3:Initialize agent set

𝒜\\mathcal\{A\}, item set

ℐ\\mathcal\{I\}, and edge set

ℰ\\mathcal\{E\}
4:for

k=1k=1to

KKdo

5:Update

𝒜\\mathcal\{A\}by adding and deleting specific agents

6:Select active agents

𝒜a​c​t​i​v​e⊆𝒜\\mathcal\{A\}\_\{active\}\\subseteq\\mathcal\{A\}
7:for all

ai∈𝒜a​c​t​i​v​ea\_\{i\}\\in\\mathcal\{A\}\_\{active\}do

8:Form a collaborative group

𝒜c​o​l⊆𝒜\\mathcal\{A\}\_\{col\}\\subseteq\\mathcal\{A\}to generate item nodes and agent\-item interactions

9:Update

ℐ\\mathcal\{I\}and

ℰ\\mathcal\{E\}
10:endfor

11:endfor

12:Construct knowledge graph

𝒢=\{V=𝒜∪ℐ,E=ℰ\}\\mathcal\{G\}=\\\{V=\\mathcal\{A\}\\cup\\mathcal\{I\},E=\\mathcal\{E\}\\\}
13:// Phase 2: QA Generation

14:Initialize multi\-choice QA generation LLM

ℱ\\mathcal\{F\}
15:Sample

NNdiverse subgraphs

\{𝒢i\}i=1N\\\{\\mathcal\{G\}\_\{i\}\\\}\_\{i=1\}^\{N\}of size

LLfrom

𝒢\\mathcal\{G\}
16:for all

𝒢i\\mathcal\{G\}\_\{i\}do

17:Rephrase

𝒢i\\mathcal\{G\}\_\{i\}into natural\-language context

CiC\_\{i\}
18:for

j=1j=1to

ttdo

19:

\{Qi​j,Ai​jr​i​g​h​t,A1,A2,A3\}←ℱ​\(Ci\),\\\{Q\_\{ij\},A\_\{ij\}^\{right\},A\_\{1\},A\_\{2\},A\_\{3\}\\\}\\leftarrow\\mathcal\{F\}\(C\_\{i\}\),A1A\_\{1\},

A2A\_\{2\},

A3A\_\{3\}are incorrect

20:Sample distractor answer

Ai​jw​r​o​n​g∼\{A1,A2,A3\}A\_\{ij\}^\{wrong\}\\sim\\\{A\_\{1\},A\_\{2\},A\_\{3\}\\\}
21:Add

\(Qi​j,Ai​jr​i​g​h​t,Ai​jw​r​o​n​g\)\(Q\_\{ij\},A\_\{ij\}^\{right\},A\_\{ij\}^\{wrong\}\)into

𝒟\\mathcal\{D\}
22:endfor

23:endfor

24:return

𝒟\\mathcal\{D\}

After graph construction, we sample multiple subgraphs with controlled overlap to maintain structural diversity, and convert each subgraph into a natural language context through rule\-based rephrasing\. Conditioned on these contexts, a large language model is used to generate multiple\-choice QA pairs with one correct answer and several plausible distractors derived from the subgraph\. This formulation converts relational reasoning over graph structures into QA format, enabling evaluation of structured knowledge understanding and interference under continual editing\.

![Refer to caption](https://arxiv.org/html/2607.26455v1/x2.png)Figure 2:An example subgraph used for QA generation\.As illustrated in Fig\.[2](https://arxiv.org/html/2607.26455#S3.F2), a queryQQtests multi\-hop relational reasoning:“Which user followed BillGates and retweeted a post about space exploration? A\. BookLoverHQ B\. FoodieFeest C\. TravelBug101 D\. CulinaryCrafts”\. The correct answerAr​i​g​h​tA^\{right\}\(A\. BookLoverHQ\) satisfies both paths, whileAw​r​o​n​gA^\{wrong\}\(D\. CulinaryCrafts\) acts as a distractor lacking the following link to BillGates\.

## 4Forgetting Quantification

To comprehensively evaluate long\-term memory across different temporal horizons, we examine not only queries within a single sequential editing process, but also variations in the overall QA sequence length\. Specifically, given a maximum sequence lengthTTand evaluation intervalkk, we construct a set of sequence lengths\{tn\}n=0T/k\\\{t\_\{n\}\\\}\_\{n=0\}^\{T/k\}, where eachtnt\_\{n\}corresponds ton⋅kn\\cdot kediting steps\. After the model learns the whole QA sequences, all previously edited knowledge items are jointly evaluated to assess memory retention over time\. We denote by𝒵​\[t,i\]\\mathcal\{Z\}\[t,i\]the correctness of the model on the query corresponding to theii\-th edit, evaluated after thett\-th evaluation round, wherei≤t⋅ki\\leq t\\cdot k\. Each entry𝒵​\[t,i\]∈\{0,1\}\\mathcal\{Z\}\[t,i\]\\in\\\{0,1\\\}indicates whether the knowledge item is correctly recalled\.

Fig\.[3](https://arxiv.org/html/2607.26455#S4.F3)visualizes the evaluation matrix𝒵\\mathcal\{Z\}under both settings, revealing distinct spatiotemporal structures\. In the concept\-based setting, correctness is highly dispersed across both edits and evaluation rounds, reflecting independently introduced factual updates that are uniformly vulnerable to interference\. In contrast, the scenario\-based setting exhibits pronounced vertical coherence, where correctness tends to persist once established and propagates across subsequent rounds, suggesting more stable prediction trajectories induced by contextual dependencies among interactions\.

Overall, our evaluation framework characterizes continual knowledge editing through a unified spatiotemporal perspective over𝒵\\mathcal\{Z\}, where the horizontal axis corresponds to memory age \(i\.e\., forgetting dynamics overΔ=t​k−i\\Delta=tk\-i\), and the vertical axis tracks individual knowledge instances across successive editing rounds\. Based on this formulation, we construct a set of complementary metrics that capture three key dimensions: \(1\) pointwise editing correctness, \(2\) temporal decay along memory age, and \(3\) cross\-instance stability across the editing trajectory\. Together, these metrics provide a unified characterization of both global performance trends and fine\-grained memory evolution dynamics\.

![Refer to caption](https://arxiv.org/html/2607.26455v1/x3.png)Figure 3:The evaluation matrices𝒵\\mathcal\{Z\}of Concept\-based \(upper\) and Scenario\-based \(lower\) tests\.### 4\.1Pointwise Editing Performance

We evaluate the instantaneous editing behavior at the level of individual knowledge updates, focusing on whether newly introduced facts are correctly learned and whether previously acquired knowledge remains intact\. Edit Success \(ES\) measures whether newly edited knowledge is correctly acquired at the time of insertion\. It is defined as the average performance on the diagonal entries of𝒵\\mathcal\{Z\}, Retention \(Ret\) evaluates how well previously edited knowledge is preserved over time\. It is defined as the average performance over all past knowledge\. ES and Ret are computed as:

ES=kT​∑t=1T/k𝒵t,t​k,Ret=kT​∑t=1T/k\(1t​k−1​∑i=1t​k−1𝒵t,i\)\.\\text\{ES\}=\\frac\{k\}\{T\}\\sum\_\{t=1\}^\{T/k\}\\mathcal\{Z\}\_\{t,tk\},\\quad\\text\{Ret\}=\\frac\{k\}\{T\}\\sum\_\{t=1\}^\{T/k\}\(\\frac\{1\}\{tk\-1\}\\sum\_\{i=1\}^\{tk\-1\}\\mathcal\{Z\}\_\{t,i\}\)\.\(1\)
ES captures the immediate effectiveness of the editing operation, and Ret reflects the overall stability of stored knowledge across editing steps, measuring how much previously learned knowledge is retained\.

### 4\.2Temporal Decay Dynamics

To characterize the dynamic decay of long\-term memory over time, we introduce the Forgetting Curve to quantify the average retention of knowledge after undergoing successive editing operations\. Unlike ES or Ret that merely compute the overall average accuracy, the Forgetting Curve explicitly models the relationship between memory age and knowledge retention\. For an arbitrary memory ageΔ\\Delta, we consider all knowledge states satisfyingΔ=t​k−i\\Delta=tk\-i, namely, all knowledge items that have experiencedΔ\\Deltasubsequent edits\. Based on this formulation, the Forgetting Curve is defined as

ℱ​\(Δ\)=𝔼​\[𝒵​\[t,i\]∣t​k−i=Δ\]\.\\mathcal\{F\}\(\\Delta\)=\\mathbb\{E\}\[\\mathcal\{Z\}\[t,i\]\\mid tk\-i=\\Delta\]\.\(2\)
ℱ​\(Δ\)\\mathcal\{F\}\(\\Delta\)yields a trajectory of memory decay across the editing process\. AsΔ\\Deltaincreases,ℱ​\(Δ\)\\mathcal\{F\}\(\\Delta\)typically exhibits a gradual decline, thereby reflecting the forgetting phenomenon of long\-term memory under continual model editing\.

Furthermore, we define the memory half\-life based on the Forgetting Curve, which characterizes the number of editing steps required for memory retention to decay to half of its initial level\. Given the Forgetting Curveℱ​\(Δ\)\\mathcal\{F\}\(\\Delta\), we define the initial retention asℱ​\(0\)\\mathcal\{F\}\(0\)and the half\-life threshold asθ=12​ℱ​\(0\)\\theta=\\frac\{1\}\{2\}\\mathcal\{F\}\(0\)\. The memory half\-lifeΔ1/2\\Delta\_\{1/2\}is then defined as the smallestΔ\\Deltasuch thatℱ​\(Δ\)≤θ\\mathcal\{F\}\(\\Delta\)\\leq\\theta, i\.e\., the first crossing point of the curve with the threshold\.

To improve robustness against local fluctuations inℱ​\(Δ\)\\mathcal\{F\(\\Delta\)\}, we first apply a lightweight smoothing preprocessing step to reduce small\-scale noise\. Since the curve is evaluated at discrete steps, the exact crossing point may lie between two adjacent samples\. We therefore apply linear interpolation: ifℱ​\(Δk−1\)\>θ\\mathcal\{F\}\(\\Delta\_\{k\-1\}\)\>\\thetaandℱ​\(Δk\)≤θ\\mathcal\{F\}\(\\Delta\_\{k\}\)\\leq\\theta, the half\-life is estimated as

Δ1/2=Δk−1\+ℱ​\(Δk−1\)−θℱ​\(Δk−1\)−ℱ​\(Δk\)​\(Δk−Δk−1\)\.\\Delta\_\{1/2\}=\\Delta\_\{k\-1\}\+\\frac\{\\mathcal\{F\}\(\\Delta\_\{k\-1\}\)\-\\theta\}\{\\mathcal\{F\}\(\\Delta\_\{k\-1\}\)\-\\mathcal\{F\}\(\\Delta\_\{k\}\)\}\(\\Delta\_\{k\}\-\\Delta\_\{k\-1\}\)\.\(3\)
Also, we analyze theF​\(Δ\)F\(\\Delta\)from both global retention behavior and local decay dynamics\. We define the Area Under the Forgetting Curve \(AUF\) to measure the average retention level of knowledge throughout its entire lifetime, and the Forgetting Speed \(Speed\) to quantify the local variation rate of the curve:

AUF=𝔼​\[ℱ​\(Δ\)\],Speed=𝔼​\[ℱ​\(Δ\)−ℱ​\(Δ\+1\)\]\.\\text\{AUF\}=\\mathbb\{E\}\[\\mathcal\{F\}\(\\Delta\)\],\\quad\\text\{Speed\}=\\mathbb\{E\}\[\\mathcal\{F\}\(\\Delta\)\-\\mathcal\{F\}\(\\Delta\+1\)\]\.\(4\)
A larger AUF indicates that the model is able to maintain a higher overall retention rate, thereby reflecting superior long\-term memory quality\. Meanwhile, a Speed closer to zero is preferred, as it implies a more stable and approximately monotonic retention curve with minimal temporal fluctuations in memory performance\.

### 4\.3Cross\-Instance Stability

While the Forgetting Curve captures the retention dynamics over memory ageΔ\\Delta, it does not explicitly characterize the temporal consistency of each knowledge trajectory across successive editing steps\. To complement this perspective, we introduce the Temporal Consistency \(TC\) metric to quantify the stability of individual memory states throughout the editing process\. Specifically, for a fixed knowledge indexii, we consider its temporal sequence𝒵⋅,i=\{𝒵​\[t,i\]\}t≥i/k\\mathcal\{Z\}\_\{\\cdot,i\}=\\\{\\mathcal\{Z\}\[t,i\]\\\}\_\{t\\geq i/k\}defined over all evaluation rounds where the corresponding memory state is observed\. We first define the transition rate of each knowledge trajectory asdi=1\|𝒵⋅,i\|−1​∑t\|𝒵​\[t\+1,i\]−𝒵​\[t,i\]\|d\_\{i\}=\\frac\{1\}\{\|\\mathcal\{Z\}\_\{\\cdot,i\}\|\-1\}\\sum\_\{t\}\|\\mathcal\{Z\}\[t\+1,i\]\-\\mathcal\{Z\}\[t,i\]\|\. The TC score is then computed as:

TC=∑iwi​\(1−di\)∑iwi,\\text\{TC\}=\\frac\{\\sum\_\{i\}w\_\{i\}\(1\-d\_\{i\}\)\}\{\\sum\_\{i\}w\_\{i\}\},\(5\)wherewi=\|𝒵⋅,i\|w\_\{i\}=\|\\mathcal\{Z\}\_\{\\cdot,i\}\|accounts for heterogeneous temporal coverage induced by the triangular structure of𝒵\\mathcal\{Z\}\. A higher TC indicates that individual knowledge states undergo fewer transitions across successive editing rounds, reflecting more stable memory evolution and reduced fluctuation under continual editing\.

Importantly, TC complements rather than replaces retention metrics, as temporal stability and memory preservation represent distinct properties\. A trajectory may remain stable \(high TC\) even after the target knowledge is lost\. Thus, they must be interpreted jointly: retention measures the quantity of preserved knowledge, while TC characterizes evolutionary smoothness\. This joint perspective reveals whether editing methods suffer from abrupt memory disruptions or exhibit stable, albeit degraded, states\.

![Refer to caption](https://arxiv.org/html/2607.26455v1/x4.png)Figure 4:Forgetting curves of Concept\-based and Scenario\-based tests\. Curves are smoothed via mean aggregation with a bin size of 10 for visual clarity\. Parametric editing displays non\-biological forgetting dynamics: concept\-based editing often exhibits extreme dynamics, ranging from rapid memory collapse to non\-decay, whereas scenario\-based editing displays artificial self\-recovery driven by shared contexts\.
### 4\.4Generalization and Fluency

We introduce Gen \(Generalization\) and Flu \(Fluency\) to evaluate the quality and stability of continuous memory updates\. Specifically, Gen measures whether edited knowledge generalizes to rephrased queries \(the accuracy difference between paraphrased and original queries\), ensuring the model avoids mere rote memorization\. Flu monitors language generation stability by calculating the ratio of valid lexical tokens, directly detecting text degeneration or model collapse caused by cumulative parameter modifications\.

## 5Experiments

### 5\.1Experimental Setup

Benchmark Details\.In the dataset generation process, profiles for concept\-based QA are randomly sampled from a name pool containing 2,500 distinct profiles\. For scenario\-based QA generation, we initialize a population of 20 agents, all instantiated using Qwen2\.5\-32B\. These agents engaged in more than 50 rounds of interaction, resulting in an interaction graph comprising 56,095 nodes and 96,082 edges\. We randomly sample 2,500 subgraphs from the interaction graph, each containing about 20 nodes with a maximum pairwise overlap of 20%\. These subgraphs are then provided to a dedicated QA\-generation agent for question synthesis\. Following this process, we constructed a benchmark consisting of 6,431 question instances in total\. We provide more details in the supplementary material\.

Table 1:Concept\-based evaluation results on ForgetBench\. The best result within each model is underlined for every metric, while the overall best result is highlighted in bold face\. Existing editing methods suffer a stark trade\-off between retention and generalization \(e\.g\., AlphaEdit\), forcing models into rote memorization rather than associative knowledge integration\.Baselines and Evaluation Protocols\.We select four representative knowledge editing approaches, including MEMIT\(Menget al\.[2022b](https://arxiv.org/html/2607.26455#bib.bib28)\), WISE\(Wanget al\.[2024](https://arxiv.org/html/2607.26455#bib.bib48)\), AlphaEdit\(Fanget al\.[2024](https://arxiv.org/html/2607.26455#bib.bib29)\), and UltraEdit\(Guet al\.[2025](https://arxiv.org/html/2607.26455#bib.bib47)\)\. We evaluate the long\-term memory capabilities across four mainstream large language models, including Llama\-3 8B/Llama\-3\.1 8B\(Grattafioriet al\.[2024](https://arxiv.org/html/2607.26455#bib.bib49)\), Qwen\-2\.5 7B\(Yanget al\.[2024](https://arxiv.org/html/2607.26455#bib.bib51)\), and DeepSeek\-R1 7B\(Guoet al\.[2025](https://arxiv.org/html/2607.26455#bib.bib50)\), covering varying architectures, pre\-training scales, and reasoning capabilities\. The implementation details and hyperparameter settings follow the default configurations provided by EasyEdit2\(Xuet al\.[2025b](https://arxiv.org/html/2607.26455#bib.bib59)\)\. We adopt our proposed evaluation metrics, including basic statistical measures, namely ES and Ret, which measures the acquisition and retention of knowledge, and metrics associated with forgetting dynamics, including the half\-lifeΔ1/2\\Delta\_\{1/2\}, forgetting speed and AUF\.

Preliminary experiments show that, most language models and knowledge editing methods suffer substantial degradation in both ES and Ret once the sequence length exceeds 500 in either evaluation scenario\. Consequently, we uniformly set the maximum sequence lengthTTto 500 throughout all experiments\. Furthermore, we set the sequence intervalkkto 10, which provides a sufficiently fine\-grained assessment of forgetting behavior while avoiding excessive computational cost\. All experiments are conducted on 4 NVIDIA GeForce RTX 4090 GPUs, each with 24 GB of memory\.

### 5\.2Experimental Results

Evaluation on the Concept\-based Test\.Tab\.[1](https://arxiv.org/html/2607.26455#S5.T1)reports the concept\-based evaluation results, exposing a profound, systemic retention\-generalization trade\-off in parametric knowledge editing\. While AlphaEdit consistently dominates retention\-oriented metrics, these memory preservation gains are bought at a catastrophic cost to generalization \(Gen\)\. This trade\-off is most severe in high\-capacity models: on DeepSeek\-R1, AlphaEdit’s peak Ret \(0\.979\) directly corresponds to a devastating Gen score of \-0\.930, and on Qwen\-2\.5, its strong Ret \(0\.885\) is accompanied by a Gen of \-0\.559\. This negative correlation reveals a critical limitation of null\-space projection: imposing rigid orthogonal constraints to shield historical parameters from interference severely restricts the update directions\.Consequently, the model is forced into a state of rote memorization that fails to integrate with its broader associative structures, preventing it from generalizing the edited knowledge to paraphrased queries\. This rigid compartmentalization stands in stark contrast to robust, associative human memory, which naturally balances retention and generalization\.

Table 2:Scenario\-based evaluation results on ForgetBench\. Structured environments provide semantic scaffolding that significantly mitigates the retention\-generalization trade\-off and improves observable recall across all methods\.Other baselines fail to navigate this trade\-off, collapsing on either side of the spectrum\. Without explicit interference control, MEMIT suffers complete parametric amnesia \(near\-zero Ret on Llama\-3 variants\), although it maintains high TC due to static, un\-updated prediction states\. Conversely, WISE acts as a conservative baseline; it preserves generalization \(Gen≈\\approx0\) and generation fluency \(Flu\) via its routing mechanism but fails to write updates into parameters, yielding weak long\-term retention\. UltraEdit similarly yields minimal retention, proving that single\-step optimization capacity does not scale to lifelong memory stability\.

The forgetting curves in the upper panel of Fig\.[4](https://arxiv.org/html/2607.26455#S4.F4)confirm these dynamics while revealing a sharp divergence from the classic Ebbinghaus model\(Ebbinghaus[2013](https://arxiv.org/html/2607.26455#bib.bib46)\)\. While the Ebbinghaus curve is characterized by a smooth, monotonic exponential decay, concept\-based LLM editing exhibits an anomalous, non\-biological dichotomy: MEMIT suffers instantaneous collapse to zero retention, whereas AlphaEdit maintains a flat, non\-decaying trajectory\. Across backbones, Qwen\-2\.5 and DeepSeek\-R1 show stronger intrinsic memory robustness than Llama\-3, showing that base model significantly influences resistance to parametric interference\.

Evaluation on the Scenario\-based Test\.Tab\.[2](https://arxiv.org/html/2607.26455#S5.T2)reports the results under the scenario\-based setting\. The most prominent distinction from the concept\-based setting is a systematic surge in both ES and Ret across all baselines\. For instance, UltraEdit on Llama\-3 8B climbs from near\-collapse in the concept\-based test to a competitive performance here\.

Our scenario\-based test shows that structured environments provide critical semantic scaffolding, enabling LLMs to retrieve memories that would otherwise be parameterally forgotten\. This contextual buffering significantly mitigates the retention\-generalization trade\-off\. The relational density narrows the performance gap between AlphaEdit and UltraEdit, while also alleviating AlphaEdit’s catastrophic generalization drop \(e\.g\., its Gen score on DeepSeek\-R1 improves from \-0\.930 to \-0\.327\)\. However, ForgetBench exposes a crucial caveat: while contextual redundancy improves observable recall, it does not guarantee intrinsic parameter consolidation\. As shown by AlphaEdit’s superior TC, stable lifelong memory still hinges on weight evolution rather than superficial contextual cues\. Conversely, WISE achieves near\-perfect fluency \(Flu: 0\.999\) and generalization \(Gen≈0\\approx 0\) but limited actual retention, proving that fluent generation does not equate to genuine memory preservation\.

The forgetting curves in the lower panel of Fig\.[4](https://arxiv.org/html/2607.26455#S4.F4)exhibit a much smoother decay gradient than the concept\-based test, superficially resembling the Ebbinghaus curve\. However, unlike human monotonic decay, they display highly erratic fluctuations and negative Speed values, indicating an artificial self\-recovery phenomenon where later edits to adjacent entities reinforce older associative retrieval pathways\.

Ultimately, these two tests provide complementary views: the concept\-based test isolates intrinsic parametric memory, while the scenario\-based test measures associative schema\-driven robustness\. This contrast highlights that while human memory integrates factual precision with associative flexibility\(McClellandet al\.[1995](https://arxiv.org/html/2607.26455#bib.bib60)\), current LLM editing paradigms are forced to trade one for the other\.

## 6Conclusion

In this work, we introduced ForgetBench, a benchmark designed to evaluate long\-term memory retention dynamics in LLMs under continual knowledge editing\. Moving beyond static, single\-step evaluations, ForgetBench models knowledge retention as a temporal process and provides a systematic framework for analyzing forgetting trajectories over sequential updates\. Our experiments across diverse models and editing methods reveal a consistent trade\-off between long\-term factual retention and query generalization, showing that transient editing success does not necessarily translate into stable knowledge preservation\. These findings highlight the importance of developing more robust parameter\-update mechanisms that can better maintain previously acquired knowledge while adapting to new information\.

Limitations and Future Works\.First, ForgetBench relies on synthetic or semi\-synthetic datasets, which may not fully capture the complexity of real\-world knowledge evolution\. Future work could integrate continuously evolving corpora to enhance realism\. Second, our benchmark focuses exclusively on sequential knowledge editing\. Extending evaluations to alternative memory update mechanisms, such as continual pretraining or reinforcement learning\-based adaptation, would provide a more comprehensive understanding of long\-term memory dynamics in LLMs\.

## References

- A\. Asai, Z\. Wu, Y\. Wang, A\. Sil, and H\. Hajishirzi \(2023\)Self\-rag: learning to retrieve, generate, and critique through self\-reflection\.InThe Twelfth International Conference on Learning Representations,Cited by:[§2\.1](https://arxiv.org/html/2607.26455#S2.SS1.p2.1)\.
- R\. C\. Atkinson and R\. M\. Shiffrin \(1968\)Human memory: a proposed system and its control processes\.InPsychology of Learning and Motivation,Vol\.2,pp\. 89–195\.Cited by:[§1](https://arxiv.org/html/2607.26455#S1.p1.1),[§2\.1](https://arxiv.org/html/2607.26455#S2.SS1.p1.1)\.
- Y\. Bai, X\. Lv, J\. Zhang, H\. Lyu, J\. Tang, Z\. Huang, Z\. Du, X\. Liu, A\. Zeng, L\. Hou,et al\.\(2024\)Longbench: a bilingual, multitask benchmark for long context understanding\.InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics \(volume 1: Long papers\),pp\. 3119–3137\.Cited by:[§1](https://arxiv.org/html/2607.26455#S1.p2.1),[§2\.1](https://arxiv.org/html/2607.26455#S2.SS1.p3.1)\.
- A\. Chevalier, A\. Wettig, A\. Ajith, and D\. Chen \(2023\)Adapting language models to compress contexts\.InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing,pp\. 3829–3846\.Cited by:[§2\.1](https://arxiv.org/html/2607.26455#S2.SS1.p2.1)\.
- J\. Deng, Z\. Wei, L\. Pang, H\. Ding, H\. Shen, and X\. Cheng \(2024\)Everything is editable: extend knowledge editing to unstructured data in large language models\.arXiv preprint arXiv:2405\.15349\.Cited by:[§1](https://arxiv.org/html/2607.26455#S1.p3.1),[§2\.2](https://arxiv.org/html/2607.26455#S2.SS2.p2.1)\.
- Y\. Ding, W\. Xia, Y\. Zhao, J\. Su, J\. Yang, Z\. Zhang, K\. Wang, and W\. Yang \(2026\)MemGround: long\-term memory evaluation kit for large language models in gamified scenarios\.arXiv preprint arXiv:2604\.14158\.Cited by:[§1](https://arxiv.org/html/2607.26455#S1.p2.1),[§2\.1](https://arxiv.org/html/2607.26455#S2.SS1.p3.1)\.
- H\. Ebbinghaus \(2013\)Memory: a contribution to experimental psychology\.Annals of Neurosciences20\(4\),pp\. 155\.Cited by:[§2\.1](https://arxiv.org/html/2607.26455#S2.SS1.p4.1),[§5\.2](https://arxiv.org/html/2607.26455#S5.SS2.p3.1)\.
- J\. Fang, H\. Jiang, K\. Wang, Y\. Ma, S\. Jie, X\. Wang, X\. He, and T\. Chua \(2024\)Alphaedit: null\-space constrained knowledge editing for language models\.arXiv preprint arXiv:2410\.02355\.Cited by:[§1](https://arxiv.org/html/2607.26455#S1.p3.1),[§2\.2](https://arxiv.org/html/2607.26455#S2.SS2.p1.1),[§5\.1](https://arxiv.org/html/2607.26455#S5.SS1.p2.1)\.
- A\. Grattafiori, A\. Dubey, A\. Jauhri, A\. Pandey, A\. Kadian, A\. Al\-Dahle, A\. Letman, A\. Mathur, A\. Schelten, A\. Vaughan,et al\.\(2024\)The llama 3 herd of models\.arXiv preprint arXiv:2407\.21783\.Cited by:[§5\.1](https://arxiv.org/html/2607.26455#S5.SS1.p2.1)\.
- X\. Gu, G\. Chen, Y\. Yang, J\. Han, and A\. Zhang \(2026\)Hierarchical orthogonal residual spread for precise massive editing in large language models\.arXiv preprint arXiv:2601\.11441\.Cited by:[§2\.2](https://arxiv.org/html/2607.26455#S2.SS2.p1.1)\.
- X\. Gu, Z\. Huang, J\. Gu, and K\. Zhang \(2025\)UltraEdit: training\-, subject\-, and memory\-free lifelong editing in language models\.arXiv preprint arXiv:2505\.14679\.Cited by:[§5\.1](https://arxiv.org/html/2607.26455#S5.SS1.p2.1)\.
- D\. Guo, D\. Yang, H\. Zhang, J\. Song, P\. Wang, Q\. Zhu, R\. Xu, R\. Zhang, S\. Ma, X\. Bi,et al\.\(2025\)Deepseek\-r1: incentivizing reasoning capability in llms via reinforcement learning\.arXiv preprint arXiv:2501\.12948\.Cited by:[§5\.1](https://arxiv.org/html/2607.26455#S5.SS1.p2.1)\.
- C\. Hu, T\. Li, X\. Gao, H\. Chen, D\. Xu, Y\. Bai, T\. Lin, X\. Zhao, X\. Li, J\. An,et al\.\(2026\)EverMemBench: benchmarking long\-term interactive memory in large language models\.arXiv preprint arXiv:2602\.01313\.Cited by:[§2\.1](https://arxiv.org/html/2607.26455#S2.SS1.p3.1)\.
- O\. Levy, M\. Seo, E\. Choi, and L\. Zettlemoyer \(2017\)Zero\-shot relation extraction via reading comprehension\.InProceedings of the 21st Conference on Computational Natural Language Learning \(CoNLL 2017\),pp\. 333–342\.Cited by:[§1](https://arxiv.org/html/2607.26455#S1.p3.1),[§2\.2](https://arxiv.org/html/2607.26455#S2.SS2.p2.1)\.
- P\. Lewis, E\. Perez, A\. Piktus, F\. Petroni, V\. Karpukhin, N\. Goyal, H\. Küttler, M\. Lewis, W\. Yih, T\. Rocktäschel,et al\.\(2020\)Retrieval\-augmented generation for knowledge\-intensive nlp tasks\.Advances in Neural Information Processing Systems33,pp\. 9459–9474\.Cited by:[§2\.1](https://arxiv.org/html/2607.26455#S2.SS1.p2.1)\.
- Y\. Li, Y\. Huang, B\. Yang, B\. Venkitesh, A\. Locatelli, H\. Ye, T\. Cai, P\. Lewis, and D\. Chen \(2024\)Snapkv: llm knows what you are looking for before generation\.Advances in Neural Information Processing Systems37,pp\. 22947–22970\.Cited by:[§2\.1](https://arxiv.org/html/2607.26455#S2.SS1.p2.1)\.
- Z\. Li, H\. Jiang, H\. Chen, B\. Bi, Z\. Zhou, F\. Sun, J\. Fang, and X\. Wang \(2025\)Reinforced lifelong editing for language models\.arXiv preprint arXiv:2502\.05759\.Cited by:[§2\.2](https://arxiv.org/html/2607.26455#S2.SS2.p1.1)\.
- X\. Liu, R\. Zhao, P\. Huang, C\. Xiao, B\. Li, J\. Wang, T\. Xiao, and J\. Zhu \(2024\)Forgetting curve: a reliable method for evaluating memorization capability for long\-context models\.InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing,pp\. 4667–4682\.Cited by:[§2\.1](https://arxiv.org/html/2607.26455#S2.SS1.p4.1)\.
- A\. Maharana, D\. Lee, S\. Tulyakov, M\. Bansal, F\. Barbieri, and Y\. Fang \(2024\)Evaluating very long\-term conversational memory of llm agents\.InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),pp\. 13851–13870\.Cited by:[§2\.1](https://arxiv.org/html/2607.26455#S2.SS1.p3.1)\.
- J\. L\. McClelland, B\. L\. McNaughton, and R\. C\. O’Reilly \(1995\)Why there are complementary learning systems in the hippocampus and neocortex: insights from the successes and failures of connectionist models of learning and memory\.\.Psychological review102\(3\),pp\. 419\.Cited by:[§5\.2](https://arxiv.org/html/2607.26455#S5.SS2.p7.1)\.
- K\. Meng, D\. Bau, A\. Andonian, and Y\. Belinkov \(2022a\)Locating and editing factual associations in gpt\.Advances in Neural Information Processing Systems35,pp\. 17359–17372\.Cited by:[§2\.2](https://arxiv.org/html/2607.26455#S2.SS2.p2.1)\.
- K\. Meng, A\. S\. Sharma, A\. Andonian, Y\. Belinkov, and D\. Bau \(2022b\)Mass\-editing memory in a transformer\.arXiv preprint arXiv:2210\.07229\.Cited by:[§1](https://arxiv.org/html/2607.26455#S1.p3.1),[§2\.2](https://arxiv.org/html/2607.26455#S2.SS2.p1.1),[§5\.1](https://arxiv.org/html/2607.26455#S5.SS1.p2.1)\.
- E\. Mitchell, C\. Lin, A\. Bosselut, C\. Finn, and C\. D\. Manning \(2021\)Fast model editing at scale\.arXiv preprint arXiv:2110\.11309\.Cited by:[§2\.2](https://arxiv.org/html/2607.26455#S2.SS2.p1.1)\.
- H\. Pan, X\. Wang, Y\. Cao, Z\. Shi, X\. Yang, J\. Li, and M\. Wang \(2025\)Precise localization of memories: a fine\-grained neuron\-level knowledge editing technique for llms\.arXiv preprint arXiv:2503\.01090\.Cited by:[§2\.2](https://arxiv.org/html/2607.26455#S2.SS2.p1.1)\.
- C\. Tan, G\. Zhang, and J\. Fu \(2023\)Massive editing for large language models via meta learning\.arXiv preprint arXiv:2311\.04661\.Cited by:[§2\.2](https://arxiv.org/html/2607.26455#S2.SS2.p1.1)\.
- M\. N\. Uddin, K\. Shubham, E\. Blanco, C\. Baral, and G\. Wang \(2026\)From recall to forgetting: benchmarking long\-term memory for personalized agents\.InFindings of the Association for Computational Linguistics: ACL 2026,pp\. 26814–26841\.Cited by:[§2\.1](https://arxiv.org/html/2607.26455#S2.SS1.p3.1)\.
- P\. Wang, Z\. Li, N\. Zhang, Z\. Xu, Y\. Yao, Y\. Jiang, P\. Xie, F\. Huang, and H\. Chen \(2024\)Wise: rethinking the knowledge memory for lifelong model editing of large language models\.Advances in Neural Information Processing Systems37,pp\. 53764–53797\.Cited by:[§5\.1](https://arxiv.org/html/2607.26455#S5.SS1.p2.1)\.
- J\. Wei, X\. Wang, D\. Schuurmans, M\. Bosma, F\. Xia, E\. Chi, Q\. V\. Le, D\. Zhou,et al\.\(2022\)Chain\-of\-thought prompting elicits reasoning in large language models\.Advances in Neural Information Processing Systems35,pp\. 24824–24837\.Cited by:[§2\.1](https://arxiv.org/html/2607.26455#S2.SS1.p2.1)\.
- D\. Wu, H\. Wang, W\. Yu, Y\. Zhang, K\. Chang, and D\. Yu \(2024\)Longmemeval: benchmarking chat assistants on long\-term interactive memory\.arXiv preprint arXiv:2410\.10813\.Cited by:[§1](https://arxiv.org/html/2607.26455#S1.p2.1),[§2\.1](https://arxiv.org/html/2607.26455#S2.SS1.p3.1)\.
- Y\. Xu, X\. Guo, Z\. Zeng, and C\. Miao \(2025a\)Softcot: soft chain\-of\-thought for efficient reasoning with llms\.InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),pp\. 23336–23351\.Cited by:[§2\.1](https://arxiv.org/html/2607.26455#S2.SS1.p2.1)\.
- Z\. Xu, S\. Wang, K\. Xu, H\. Xu, M\. Wang, X\. Deng, Y\. Yao, G\. Zheng, H\. Chen, and N\. Zhang \(2025b\)Easyedit2: an easy\-to\-use steering framework for editing large language models\.InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing: System Demonstrations,pp\. 522–535\.Cited by:[§5\.1](https://arxiv.org/html/2607.26455#S5.SS1.p2.1)\.
- A\. Yang, B\. Yang, B\. Hui, B\. Zheng, B\. Yu, C\. Zhou, C\. Li, C\. Li, D\. Liu, F\. Huang,et al\.\(2024\)Qwen2 technical report\.arXiv preprint arXiv:2407\.10671\.Cited by:[§5\.1](https://arxiv.org/html/2607.26455#S5.SS1.p2.1)\.
- Z\. Yang, P\. Qi, S\. Zhang, Y\. Bengio, W\. Cohen, R\. Salakhutdinov, and C\. D\. Manning \(2018\)HotpotQA: a dataset for diverse, explainable multi\-hop question answering\.InProceedings of the 2018 Conference on Empirical Methods in Natural Language Processing,pp\. 2369–2380\.Cited by:[§1](https://arxiv.org/html/2607.26455#S1.p2.1),[§2\.1](https://arxiv.org/html/2607.26455#S2.SS1.p3.1)\.
- H\. Zhang, Z\. Tang, X\. Yu, X\. Liu, Y\. Gong, H\. Huang, Y\. Lu, W\. Deng, F\. Sun, Q\. Zhang,et al\.\(2026\)Beyond static dialogues: benchmarking realistic, heterogeneous, and evolving long\-term memory\.arXiv preprint arXiv:2605\.31086\.Cited by:[§2\.1](https://arxiv.org/html/2607.26455#S2.SS1.p3.1)\.