New Model, Old Habits: Behavior Transfer under Model Swapping in Multi-Session LLM Agents

Youngbin Choi; Sehyun Hwang; Suhyeon Jun; Seoyeon Lee; Sunghyun Choi; MoonJeong Park; Dongwoo Kim

cs.LGOct 14, 2026

New Model, Old Habits: Behavior Transfer under Model Swapping in Multi-Session LLM Agents

Authors: Youngbin Choi, Sehyun Hwang, Suhyeon Jun, Seoyeon Lee, Sunghyun Choi, MoonJeong Park, Dongwoo Kim

Organizations: Graduate School of Artificial Intelligence, POSTECH · Meta · Department of Computer Science and Engineering, POSTECH

Abstract

Multi-session agents can switch between LLMs to balance performance and inference cost. Such switches require transferring a history for continuity, yet their effects on the recipient remain underexplored. Through controlled experiments with multi-session tasks, we uncover behavior transfer: recipients tend to adopt reasoning and action patterns from the models that generated their histories. This effect is most evident with raw-history transfer but is substantially reduced by summarization. Based on these findings, we propose Direction- and Authorship-aware Summarization of History (DASH) that preserves stronger-model behavior while minimizing weak-to-strong transfer via summarization. Experiments with repeated swaps show that DASH outperforms raw-history transfer and naive summarization baselines. These results suggest that effective multi-model orchestration requires considering not only which model to use, but also which behavioral patterns to carry forward.

Figures & tables

Figure 1: An example of behavior transfer in Web Shopping. Both pre-swap histories record the same correct purchases. With its own history, the weak model ( Qwen3.5-4B ) uses short product queries and purchases an incorrect item. With history inherited from the strong model ( Qwen3.5-27B ), the same recipient uses more detailed queries and purchases the correct item.

Figure 2: Swap-induced performance shift versus the no-swap baselines’ performance gap between source $X$ and recipient $Y$ , for all model pairs (left) and the correctness-controlled subset (right). Each point represents a directed swap $X\rightarrow Y$ on one agentic task, with marker shape indicating the model pair and color/fill indicating the task. The horizontal axis measures how much stronger the source is than the recipient when each model runs with its own history ( $\mathrm{PS}_{X}-\mathrm{PS}_{Y}$ ), and the vertical axis measures how much the recipient’s performance changes after inheriting the source’s history, relative to running every session from scratch ( $\mathrm{PS}_{X\rightarrow Y}-\mathrm{PS}_{Y}$ ). Blue ( $x<0$ ) and red ( $x>0$ ) shading indicate weak-to-strong and strong-to-weak swaps, respectively. For example, the purple upward-pointing triangle near $(50,35)$ in the left panel represents a Qwen3.5-27B ( $X$ ) $\rightarrow$ Qwen3.5-4B ( $Y$ ) swap in Web Shopping. Its horizontal position indicates that Qwen3.5-27B outperforms Qwen3.5-4B by nearly $50$ PS when each model uses its own history, and its vertical position shows that, after inheriting Qwen3.5-27B ’s history, Qwen3.5-4B gains approximately $35$ percentage points relative to the baseline in which it handles all sessions from scratch.

Figure 3: Behavior transfer after model swapping. Recipients tend to adopt behavioral patterns from the source model’s history. For each behavioral feature and model pair

(X,Y)

, position

0

and

1

on the horizontal axis represent the baseline behavior feature values of

X

and

Y

, respectively, without model swap. Orange markers show

X

with

Y

’s history, and blue markers show

Y

with

X

’s history. Thus, orange markers near

1

and blue markers near

0

show that the behaviors of the recipient are similar to those of the source model baseline. The plotted positions are

\pi_{Y\rightarrow X}

for orange markers and

1-\pi_{X\rightarrow Y}

for blue markers ( Eq. 7 ). Points are omitted when the two baseline models have nearly identical feature values. Figure C.1 compares behavior before and after compaction, and Table C.2 provides extended results.

Figure 4: Behavior transfer is strongest under raw-history transfer, substantially reduced by reasoning pruning, and tends to be reduced further by summarization. The horizontal axis shows the percentage of behavioral features for which the recipient is closer to the source model’s no-swap baseline than to its own (

\pi_{X\rightarrow Y}>0.5

), with both baselines measured under self-generated raw history. For each environment and swap direction, percentages are computed over features with valid normalized behavioral shifts. Feature-level results are provided in Tables C.2 , C.3 and C.4 , and Figure C.4 shows results for additional model pairs.

Figure 5: Mean post-swap performance (

\mathrm{PS}_{X\rightarrow Y}

) across environments with different history transfer strategies. An effective history transfer strategy depends on the model swap direction: summarization achieves the highest mean PS for weak-to-strong swaps, whereas raw-history transfer performs best for strong-to-weak swaps. Scores are averaged equally over Web Shopping, Travel Planning, and Web Search for (a) weak-to-strong and (b) strong-to-weak swaps. No-swap is the recipient model’s baseline performance without a model swap (

\mathrm{PS}_{Y}

). Bold values indicate the highest mean PS within each group. Full results are reported in Table D.1 .

Weak-first ( $\psi_{W}\to\psi_{S}\to\psi_{W}\to\psi_{S}$ )
	History transfer	Qwen3.5-4B Qwen3.5-27B	Qwen3.5-4B DeepSeek	Gemma-4-E4B Gemma-4-31B	Gemma-4-E4B DeepSeek
Web	$\Phi_{\mathrm{raw}}$	23.5	32.5	24.8	27.5
	$\Phi_{\mathrm{DiSH}}$	27.2	39.5	30.5	35.7
	$\Phi_{\mathrm{DASH}}$	32.3	39.8	31.7	35.5
Travel	$\Phi_{\mathrm{raw}}$	29.3	32.2	18.6	27.5
	$\Phi_{\mathrm{DiSH}}$	31.3	47.8	24.8	40.9

Table 1: Performance of history transfer strategies under repeated model swaps on Web Shopping (Web) and Travel Planning (Travel). The proposed DASH transfer strategy achieves the highest PS in most evaluated settings. Values report PS (%) over all sessions following the first swap, for weak-first and strong-first model orders. Each model-pair column lists the weaker model

\psi_{W}

above the stronger model

\psi_{S}

. DeepSeek denotes DeepSeek-V4.1-Flash , and all Gemma models are the instruction-tuned ( -it ) variants.

Figure 6: Direction-aware summarization shifts post-swap behavior toward the strong model, especially under weak-first routing. The horizontal axis shows the percentage of behavioral features whose post-swap means are closer to the strong model’s no-swap baseline than to the weak model’s (

(v_{\mathrm{post}}-v_{W})/(v_{S}-v_{W})>0.5

). Feature means pool sessions after the first swap regardless of the acting model, and percentages are computed over features with valid normalized shifts. Figure C.5 shows results for additional model pairs.

Appendix figures & tables18 assets

Supplementary material from the paper’s appendix.

Appendix

Environment	Tasks	Sessions	$N$	Swap points
Web Shopping	150	900	6	450
Travel Planning	270	1,869	5–8	773
Web Search	221	1,641	4–16	633
Math	40	354	2–16	115
Physics	20	86	2–12	47

Appendix

Table B.1: MemoryArena statistics used in our experiments. Tasks and Sessions count task sequences and individual sessions, respectively.

N

is the number of sessions per sequence. The last column reports the number of evaluated swap positions.

Environment	Max tokens	Reasoning budget	Max steps
Web Shopping	4,096	3,584	20
Travel Planning	32,768	22,938	30
Web Search	15,000 †	1,024	35
Math, Physics	8,192	–	3

Appendix

Table B.2: Generation settings. Max tokens is the limit per model step, except for Web Search, where it is a session-level budget ( † ). Reasoning budgets apply per step, with reasoning mode disabled for Math and Physics.

Feature	Session-level unit	Definition
Web Shopping
Words per query	Mean words	Mean number of words per search[...] query.
Query copies option text	Fraction of queries	Queries containing at least 80% of the content words of one listed option.
Repeats identical query	0/1	1 if an identical search[...] action is issued more than once in the session.
Plan statement in reasoning	Fraction of calls	first , next, , then , I will , I’ll , let me , plan , step , now I .
Options covered by queries	Fraction of options	Listed options with at least 50% of their content words contained in at least one query.

Appendix

Table B.3: Web Shopping feature definitions for Table C.2 . Units describe measurements within one session. Lexical entries summarize case-insensitive, whole-word matching patterns. An asterisk denotes any suffix, parentheses denote optional text, and slashes separate alternatives. Content words exclude stopwords and very short tokens.

Feature	Session-level unit	Definition
Travel Planning
Budget terms in reasoning	Matches per 1k chars	budget , cost(s) , price(s) , dollar amounts, per person/night , total , cheaper , expensive .
Meal slots filled in plan	Fraction of slots	Breakfast, lunch, and dinner slots with plan entries other than - .
“Format/exactly” in reasoning	Matches per 1k chars	format* , exact(ly) .
No tool call	0/1	1 if the plan is produced without any tool call.
Max reasoning length	Thousand characters	Maximum reasoning length across the session’s model calls, divided by 1,000.

Appendix

Table B.4: Travel Planning and Web Search feature definitions for Table C.2 . Units describe measurements within one session. Pattern notation follows Table B.3 .

Figure C.1: Web Search behavior transfer before and after context compaction, on the common axis of Fig. 3 (same features, markers, colors, and omission rule). Before compaction : each post-swap continuation is measured from the swap until its first context compaction; the session in which that compaction occurs is cut just before it, so the recipient still sees the inherited history in full. This is the measurement shown in Fig. 3 . After compaction : the same continuations measured from their first compaction onward (the remainder of the compacting session and every later session); continuations that never compact are excluded. Once compaction replaces the inherited history with a summary, most markers move back toward their own model’s self behavior.

All paired swap points
	Physics		Math
History used	Self	Inherited	Self	Inherited
History cut ratio (%)	23.0	18.9	13.4	17.8
Continuation cut ratio (%)	15.2	6.3	13.9	14.5
Continuation PS (%)	44.9	52.7	25.3	24.1
Swap effect	${+7.9}$		${-1.3}$

Appendix

Table C.1: Swap effects for DeepSeek-V4.1-Flash . DeepSeek continues with either its own history (Self) or Qwen3.5-4B ’s history (Inherited). A session is cut when its reasoning is truncated at the 8,192-token maximum generation length. History cut ratio measures the percentage of cut sessions in the history available at the swap point. Continuation cut ratio measures this percentage among subsequent sessions solved by DeepSeek under each history. Both ratios are averaged first within and then across tasks. Swap effect is

\mathrm{PS}_{\mathrm{Inherited}}-\mathrm{PS}_{\mathrm{Self}}

. The lower blocks group paired swap points by cut occurrence in DeepSeek ’s own history, regardless of cuts in the inherited history.

Figure C.3: Swap position analysis of model swapping across environments. We report $\mathrm{PS}_{X\rightarrow Y}-\mathrm{PS}_{Y}$ , where $X$ is the source model and $Y$ is the recipient model, using $Y$ ’s self-baseline for the corresponding context and subset.

Figure C.4: Behavior transfer is strongest under raw-history transfer, substantially reduced by reasoning pruning, and tends to be reduced further by summarization. The horizontal axis shows the percentage of behavioral features for which the recipient is closer to the source model’s baseline than to its own (

\pi_{X\rightarrow Y}>0.5

), with both baselines measured under self-generated raw history. For each environment and swap direction, percentages are computed over features with valid normalized behavioral shifts. Feature-level results are provided in Tables C.2 , C.3 and C.4 . 4B, 27B, DS, and G denote Qwen3.5-4B , Qwen3.5-27B , DeepSeek-V4.1-Flash , and Gemma-4-31B-it , respectively.

	Self-history baseline				Normalized shift
Feature	4B	27B	DS	G	$27\mathrm{B}\!\leftarrow\!4\mathrm{B}$	$4\mathrm{B}\!\leftarrow\!27\mathrm{B}$	$\mathrm{DS}\!\leftarrow\!4\mathrm{B}$	$4\mathrm{B}\!\leftarrow\!\mathrm{DS}$	$27\mathrm{B}\!\leftarrow\!\mathrm{G}$	$\mathrm{G}\!\leftarrow\!27\mathrm{B}$
Web Shopping
Words per query	3.82	8.18	9.14	10.87	0.85	0.94	0.73	0.97	0.88	0.87
Query copies option text	0.07	0.39	0.58	0.58	0.92	0.87	0.84	0.90	1.10	1.07
Repeats identical query	0.08	0.48	0.32	0.30	0.74	0.76	0.78	0.96	0.44	0.84
Plan statement in reasoning	0.90	0.46	0.55	0.85	1.00	1.06	1.00	1.59	0.74	0.55

Appendix

Table C.2: Behavioral features under raw-history transfer. Baselines report continuation behavior with self-generated histories. Shift columns report

\pi_{X\rightarrow Y}

, with source

X

and recipient

Y

. Shifts involving G use pair-specific baselines, as detailed in Appendix B.4 . For Web Search, each post-swap continuation (and the corresponding self-history continuation) is measured from the swap up to its first context compaction; the compacting session is cut just before it, and later sessions are excluded. Dashes mark comparisons excluded by the normalization criterion; shifts are also omitted where the two models’ self-history rates differ by at most 0.04, since

\pi

is then dominated by its near-zero denominator. 4B, 27B, DS, and G denote Qwen3.5-4B , Qwen3.5-27B , DeepSeek-V4.1-Flash , and Gemma-4-31B-it , respectively.

	Self-history baseline				Normalized shift
Feature	4B	27B	DS	G	$27\mathrm{B}\!\leftarrow\!4\mathrm{B}$	$4\mathrm{B}\!\leftarrow\!27\mathrm{B}$	$\mathrm{DS}\!\leftarrow\!4\mathrm{B}$	$4\mathrm{B}\!\leftarrow\!\mathrm{DS}$	$27\mathrm{B}\!\leftarrow\!\mathrm{G}$	$\mathrm{G}\!\leftarrow\!27\mathrm{B}$
Web Shopping
Words per query	3.82	8.18	9.14	10.87	0.70	0.44	0.68	0.61	0.32	0.76
Query copies option text	0.07	0.39	0.58	0.58	0.83	0.33	0.81	0.54	0.50	0.98
Repeats identical query	0.08	0.48	0.32	0.30	0.17	0.08	0.13	0.22	0.13	0.16
Plan statement in reasoning	0.90	0.46	0.55	0.85	0.00	0.27	$-0.31$	0.51	0.05	0.40

Appendix

Table C.3: Behavioral features under swap-time reasoning pruning (

\Phi_{\text{prune}}

). Baselines report continuation behavior with self-generated histories, as in Table C.2 . Shift columns report

\pi_{X\rightarrow Y}

for recipient

Y

continuing the history of source

X

transferred by

\Phi_{\text{prune}}

, computed on the same sessions as the transferred-history condition. Dashes mark comparisons excluded by the normalization criterion; shifts are also omitted where the two models’ self-history rates differ by at most 0.04, since

\pi

is then dominated by its near-zero denominator. 4B, 27B, DS, and G denote Qwen3.5-4B , Qwen3.5-27B , DeepSeek-V4.1-Flash , and Gemma-4-31B-it , respectively. For Web Search, each post-swap continuation (and the corresponding self-history continuation) is measured from the swap up to its first context compaction; the compacting session is cut just before it, and later sessions are excluded.

	Self-history baseline				Normalized shift
Feature	4B	27B	DS	G	$27\mathrm{B}\!\leftarrow\!4\mathrm{B}$	$4\mathrm{B}\!\leftarrow\!27\mathrm{B}$	$\mathrm{DS}\!\leftarrow\!4\mathrm{B}$	$4\mathrm{B}\!\leftarrow\!\mathrm{DS}$	$27\mathrm{B}\!\leftarrow\!\mathrm{G}$	$\mathrm{G}\!\leftarrow\!27\mathrm{B}$
Web Shopping
Words per query	3.82	8.18	9.14	10.87	0.46	$-0.17$	0.40	$-0.08$	0.04	0.00
Query copies option text	0.07	0.39	0.58	0.58	0.53	$-0.16$	0.52	$-0.06$	0.00	$-0.11$
Repeats identical query	0.08	0.48	0.32	0.30	0.30	0.01	0.33	0.02	$-0.06$	0.56
Plan statement in reasoning	0.90	0.46	0.55	0.85	0.01	0.07	0.78	0.09	0.03	0.09

Appendix

Table C.4: Behavioral features under session-wise summarization (

\Phi_{\text{summary}}

). Baselines report continuation behavior with self-generated histories, as in Table C.2 . Shift columns report

\pi_{X\rightarrow Y}

for recipient

Y

continuing the history of source

X

transferred by

\Phi_{\text{summary}}

\pi

\Phi_{\text{summary}}

, both recipients in a pair search more, and more often in parallel, than either model does on its own history, so the search-volume features take large shifts of opposite sign in the two directions; these reflect the summarized context itself rather than transfer from the source model.

Direction	Model / Swap	Subset	Math	Physics
No swap (Self)	Qwen3.5-4B		14.8	25.6
	Qwen3.5-27B		25.2	55.4
	DeepSeek		25.3	44.9
	Gemma		28.4	57.4
Weak $\rightarrow$ strong	Qwen3.5-4B $\rightarrow$ Qwen3.5-27B	All paired	26.4 $(+1.2)$	47.6 $(-7.8)$
	Qwen3.5-4B $\rightarrow$ Qwen3.5-27B	Correctness	26.1 $(+1.0)$	59.5 $(-2.4)$

Appendix

Table C.5: Post-swap performance under raw-history transfer on the formal reasoning environments. Self rows report performance without model swapping. For each directed swap, we report all paired evaluations and the correctness-controlled subset (Correctness), in which the source and recipient histories have matching correctness patterns across all pre-swap sessions. Values in parentheses report

\mathrm{PS}_{X\rightarrow Y}-\mathrm{PS}_{Y}

, where

X

is the source model and

Y

is the recipient model, using

Y

’s self baseline for the corresponding context and subset.

Figure C.5: Direction-aware summarization shifts post-swap behavior toward the strong model, especially under weak-first routing. The horizontal axis shows the percentage of behavioral features whose post-swap means are closer to the strong model’s baseline than to the weak model’s (

(v_{\mathrm{post}}-v_{W})/(v_{S}-v_{W})>0.5

), with both baselines measured under self-generated raw history. Feature means pool sessions after the first swap regardless of the acting model, and percentages are computed over features with valid normalized shifts.

Swap	Model / Swap	Context	Web Shopping	Travel Planning	Web Search
No swap (Self)	Qwen3.5-4B	$\Phi_{\mathrm{raw}}$	12.9	14.4	9.4
	Qwen3.5-27B	$\Phi_{\mathrm{raw}}$	62.1	51.8	15.9
	Gemma	$\Phi_{\mathrm{raw}}$	65.8	24.8	14.1
	DeepSeek	$\Phi_{\mathrm{raw}}$	62.4	57.8	29.3
Weak $\rightarrow$ strong	Qwen3.5-4B $\rightarrow$ Qwen3.5-27B	$\Phi_{\mathrm{raw}}$	28.7 $(-33.4)$	28.5 $(-23.3)$	13.8 $(-2.1)$
		$\Phi_{\mathrm{prune}}$	35.9 $(-26.2)$	26.2 $(-25.6)$	13.7 $(-2.2)$

Appendix

Table D.1: Post-swap performance on the agentic environments in MemoryArena.

\mathcal{M}_{A}\rightarrow\mathcal{M}_{B}

denotes a swap from source model

\mathcal{M}_{A}

to recipient model

\mathcal{M}_{B}

. Values in parentheses report

\mathrm{PS}_{\mathrm{swap}}-\mathrm{PS}_{\mathrm{self}}

. For swap-time pruning and summarization, subsequent sessions retain raw history. All summaries are written by Qwen3.5-27B . DeepSeek and Gemma denote DeepSeek-V4.1-Flash and Gemma-4-31B-it , respectively.

Swap	Model / Swap	Subset	Web Shopping	Travel Planning	Web Search
Weak $\rightarrow$ strong	Qwen3.5-4B $\rightarrow$ Qwen3.5-27B	All paired	28.7 $(-33.4)$	28.5 $(-23.3)$	13.8 $(-2.1)$
		Correctness-controlled	39.7 $(-14.5)$	29.8 $(-12.0)$	12.4 $(-1.4)$
		# tasks (controlled / all)	45 / 150	111 / 270	204 / 221
	Qwen3.5-4B $\rightarrow$ DeepSeek	All paired	47.6 $(-14.8)$	44.4 $(-13.4)$	23.8 $(-5.5)$
		Correctness-controlled	49.3 $(-5.8)$	37.6 $(-0.6)$	21.3 $(-3.9)$
		# tasks (controlled / all)	36 / 150	88 / 270	189 / 221

Appendix

Table D.2: Post-swap performance under raw-history transfer on the agentic environments, for all paired evaluations and the correctness-controlled subset. The controlled subset includes swap points where the source and recipient models have matching success/failure outcomes in every pre-swap session. Values in parentheses report

\mathrm{PS}_{\mathrm{swap}}-\mathrm{PS}_{\mathrm{self}}

, using the recipient’s self score on the corresponding subset. The last row of each model pair reports the number of task sequences in the controlled subset and in all paired evaluations. DeepSeek and Gemma denote DeepSeek-V4.1-Flash and Gemma-4-31B-it , respectively.

Panel of Figure 24 — Figure E.1: Session-wise summarization and the transferred context for a single swap after session $i$ . Each session is summarized independently, with the original request retained verbatim.

Panel of Figure 25 — Figure E.2: Prompt for context compaction in Web Search.

LLM Agents

Large language model agents use language models to plan, use tools, and act across multi-step tasks and environments.

See all 711 papers

Sep 5, 2026

What if LLMs Ate Their Words: Causal History Effects in Multi-Turn Interaction

Multi-turn interaction creates a feedback process in which an LLM's previous responses become context for later behavior. Prior work shows substantial multi-turn degradation and that assistant-generated history can affect later behavior. However, it remains unclear how these effects manifest across models, tasks, turns, and inside a model. We study these gaps across six task families and five models. Degradation from fully specified single-turn input (FULL) to progressively revealed multi-turn interaction (SHARDED) is clearly task- and model-dependent, and stronger one-shot performance does not imply greater interaction robustness. We then retrospectively analyze completed SHARDED conversations by replaying the user messages already observed in each trajectory while editing only assistant-generated history. Replacing prior assistant responses with neutral content (termed neutralization) changes downstream min-max normalized performance by +.027 across 2,973 trajectories. On a prespecified length-controlled subset, short and length-matched neutralization yield nearly identical effects (+.069 versus +.068), showing that simple context shortening is insufficient to explain the effect of history editing. Turn Surgery further intervenes on one assistant turn at a time. Among 237 selected degraded trajectories, 63.7% contain at least one beneficial intervention, while most tested positions remain unchanged; for binary tasks, 48.4% admit a fail-to-success reversal. An open-weight case study links behaviorally consequential history changes to measurable downstream state differences, but finds task-dependent rather than universal internal signatures. Overall, assistant-generated history has active but selective effects on multi-turn performance, motivating selective rather than uniform history management.

Jinnan Li, Zheren Fu, Yue Wang +3

School of Artificial Intelligence Jilin University Changchun, China · School of Cyber Science and Technology, University of Science and Technology of China Hefei, China · School of Information and Library Science, University of North Carolina at Chapel Hill Chapel Hill, North Carolina, USA

LLM Context Management Causal Interventions in Language Models Multi-Turn Dialogue Evaluation

Ask PDF

May 30, 2026

Post-Training Recipe, More Than Model Family, Shapes Multi-Agent LLM Conversational Behavior

Multi-LLM systems use multiple language models to deliberate, judge each other's outputs, or coordinate as agents. Their value depends on the models producing measurably different conversational behaviors when given the same input. Prior offline studies recommend drawing one model per family for behavioral diversity, because LLMs prefer outputs from their own family when rating one another in isolation. Whether the same family label predicts behavior in interactive multi-LLM systems, the setting that real deployed systems use, has not been tested. We study this with a 940,000-chain 11-checkpoint corpus and a 1.6M-chain same-base Llama factorial. On our validated headline metric, hedging, a reasoning-distilled Llama checkpoint shifts by 18% depending on which same-base partner it replies to, more than any cross-family hedging gap in the controlled subset. Qwen, closed-API, and runtime checks suggest the pattern is not isolated, while repair and challenge analyses remain exploratory because their surface-cue detectors are weaker. Overall, the results identify post-training recipe as a first-class axis for multi-LLM panel composition and show that model family alone is an incomplete proxy for conversational diversity.

Luyang Zhang, Jialu Wang, Fei Xue +1

Carnegie Mellon University · University of California, Santa Cruz · Independent Researcher

Language Model Post-Training Multi-LLM Collaboration

Ask PDF

Oct 14, 2026

XferAgent: Cross-Agent Experience Transfer for Self-Improving LLM Agents

Large language model (LLM) agents can reuse experience from prior task executions to improve performance on subsequent tasks. Although existing methods enable cross-agent experience sharing, differences in agents' execution histories and task requirements make shared experiences difficult to reuse effectively. We propose XferAgent, a framework for cross-agent experience transfer that organizes reusable experience in local and shared experience graphs. XferAgent reconciles compatible contributions from heterogeneous agents while avoiding unreliable merges. For each new task, it selects shared experiences and integrates them with the target agent's local experience to support task execution. Experiments across diverse task domains show improvements over representative memory-based methods and effective experience transfer between heterogeneous agent frameworks. Cumulative task success also increases over successive task passes as agents update and reuse their experiences.

Weihao Li, Jun Bai, Shijie Chen +3

McCormick School of Engineering, Northwestern University · School of Computer Science, McGill University · Mila – Quebec AI Institute +1

LLM Agent Self-Improvement Memory-Augmented Language Model Agents Graph-Based Agent Memory Transfer Learning Continual Learning for LLM Agents

Ask PDF

Abstract

LLM Agents

Explore similar work

What if LLMs Ate Their Words: Causal History Effects in Multi-Turn Interaction

Post-Training Recipe, More Than Model Family, Shapes Multi-Agent LLM Conversational Behavior

XferAgent: Cross-Agent Experience Transfer for Self-Improving LLM Agents