TrajLong: Co-Designing Agentic and Long-Context Supervision for Mid-Training

Miao Peng; Qintong Zhang; Nuo Chen; Yuhan Li; Guochen Yan; Xinran Gu; Hongqiu Wu; Hai Wang

cs.CLOct 4, 2026

TrajLong: Co-Designing Agentic and Long-Context Supervision for Mid-Training

Authors: Miao Peng, Qintong Zhang, Nuo Chen, Yuhan Li, Guochen Yan, Xinran Gu, Hongqiu Wu, Hai Wang,

Organizations: The Hong Kong University of Science and Technology (Guangzhou) · Hunyuan Team, Tencent · Peking University

Abstract

LLM agents for coding, search, and workplace tasks increasingly rely on long-context capabilities to effectively aggregate and reason over extended interaction histories. Recent work has incorporated agent trajectories into mid-training stage, drawing on their naturally long and interaction-rich structure. Yet how to organize the information within these trajectories into effective mid-training supervision remains underexplored. In this work, we investigate the relationship between long-context and agent atomic capabilities and introduce TrajLong, a novel framework that compiles trajectories into long-context training tasks with dense supervision, targeting three representative atomic capabilities: evidence grounding, cross-evidence aggregation, and temporal state maintenance. We mid-train Qwen3-14B-Base and Qwen3-30B-A3B-Base with data compiled by TrajLong, followed by supervised fine-tuning. Experiments on 6 long-context and 12 agent benchmarks demonstrate broad performance gains, with controlled ablations showing improvements over raw and masked trajectory baselines. Capability-level analyses further reveal task-dependent associations between long-context and agent atomic capabilities. These findings suggest that the shared capability demands of long-context reasoning and agent execution provide a principled basis for designing mid-training data to develop downstream agent capabilities.

Figures & tables

Panel of Figure 1 — Figure 1: Overview of TrajLong . Left: Full-trajectory and masked mid-training strategies compared with TrajLong compilation. Right: Avg@3 scores of Qwen3-14B variants on seven long-context and agent benchmarks across search, tool-use, and coding tasks.

Figure 2: Correlation between RULER ICL gains and downstream BrowseComp-ZH gains across Qwen3-14B variants mid-trained with different agent trajectories.

Figure 3: The TrajLong data construction pipeline. Search and coding agent trajectories are cleaned and filtered, then compiled into long-context training examples targeting evidence grounding (Retrieval), cross-evidence aggregation (MultidocQA), and temporal state maintenance (State Tracking), providing capability-targeted supervision for long-context agentic mid-training.

Models	LongBench V2		MRCR		GraphWalks		GSM-Inf		CorpusQA
Models	Avg@3	BoN@3	Avg@3	BoN@3	Avg@3	BoN@3	Avg@3	BoN@3	Avg@3	BoN@3
Frontier Models
Gemini-2.5-Flash-Thinking	58.18	67.00	59.45	73.52	41.10	58.41	25.00	42.67	77.41	83.28
QwenLong L1.5-30B-A3B	45.04	54.69	55.12	61.51	39.24	44.16	39.56	64.33	80.14	88.45
GPT-OSS-120B	26.15	38.12	16.66	26.54	30.79	35.48	24.00	42.00	62.31	75.68
Our Models

Table 1: Performance on five long-context benchmarks (Avg@3 and BoN@3). MRCR averages scores equally across 2/4/8-needle settings, and GraphWalks across 0–128K and 128–256K context-length intervals. Subscripts show changes relative to the corresponding Qwen3 model within each group (pp). Bold marks the best score within each group.

Models	Search					Tool Use			Coding
Models	BC	BC-ZH	WideSearch	FSC-V2	SealQA	MCP-Atlas	SkillsBench	WebArena	TB 2.1	NL2Repo	SWE-Verified	SWE-Pro
Frontier Models
Gemini-2.5-Flash-Thinking	12.72	27.89	30.84	34.48	24.02	19.87	7.41	39.43	17.75	12.40	48.60	24.08
QwenLong L1.5-30B-A3B	2.98	11.07	35.27	20.36	16.58	32.73	1.39	27.17	1.87	7.71	31.60	10.94
GPT-OSS-120B	25.62	24.22	36.46	41.09	22.52	39.40	-	31.40	12.36	6.88	53.00	27.13
Our Models

Table 2: Performance on search, tool-use, and coding agent benchmarks (Avg@3). Bold marks the best score within each group. Benchmark details and BoN@3 results are in Appendix B.3 and E.2 .

Mid-training Variant	Long-Context			Agent
Mid-training Variant	Oolong-Synth	GraphWalks	CorpusQA	FSC-V2	SkillsBench	SWE-Verified	SWE-Pro
TrajLong - All	58.50 +0.0	32.30 +0.0	63.83 +0.0	43.53 +0.0	11.11 +0.0	54.40 +0.0	36.43 +0.0
SFT only	53.17 -5.33	24.12 -8.18	55.93 -7.90	35.09 -8.44	6.48 -4.63	42.13 -12.27	24.21 -12.22
Search Trajectories	57.33 -1.17	31.58 -0.72	61.80 -2.03	40.05 -3.48	9.72 -1.39	50.27 -4.13	31.37 -5.06
Code Trajectories	55.33 -3.17	32.77 +0.47	61.80 -2.03	39.49 -4.04	8.80 -2.31	52.40 -2.00	34.38 -2.05
Combined Trajectories	56.67 -1.83	31.38 -0.92	61.80 -2.03	40.87 -2.66	9.26 -1.85	53.60 -0.80	35.93 -0.50

Table 3: Ablation Study (Avg@3) of the mid-training data construction on long-context and agent benchmarks. All variants use the same Qwen3-14B backbone, downstream SFT recipe, and evaluation protocol. The trajectory baselines use raw search trajectories, raw code trajectories, or their combination, with an additional masked variant. All mid-training runs use the same token budget.

Figure 4: ICL score gains over Qwen3-14B-Base on long-context (top) and agent (bottom) tasks during mid-training (pp). Training steps are scaled to 40B tokens per run.

Figure 5: Spearman correlations between long-context and agent ICL gains across five Qwen3-14B variants at final mid-training checkpoints, relative to unmasked Raw Trajectories.

Figure 6: Post-SFT Avg@3 gains of Qwen3-14B variants on long-context and agent benchmarks, relative to unmasked Raw Trajectories under the same SFT recipe. Panels use separate color scales.

Setting	SFT Epoch	Long-Context			Agent
Setting	SFT Epoch	Oolong-Synth	GraphWalks	CorpusQA	BC	NL2Repo	SWE-Pro
None	2	53.17 +0.0	24.12 +0.0	55.93 +0.0	12.53 +0.0	6.23 +0.0	24.21 +0.0
ACC (SFT)	2	56.00 +2.83	30.62 +6.50	57.89 +1.96	10.16 -2.37	7.00 +0.77	27.22 +3.01
TrajLong (SFT)	2	56.83 +3.66	28.04 +3.92	59.27 +3.34	13.56 +1.03	9.40 +3.17	31.11 +6.90
	1	55.83 +2.66	31.64 +7.52	64.44 +8.51	11.48 -1.05	14.16 +7.93	34.15 +9.94
	2	58.50 +5.33	32.30 +8.18	63.83 +7.90	11.77 -0.76	15.03 +8.80	36.43 +12.22

Table 4: Training-stage and SFT-epoch comparisons on Qwen3-14B (Avg@3). None denotes SFT on the original data only. Subscripts show changes from None (pp); bold marks column maxima.

Appendix figures & tables17 assets

Supplementary material from the paper’s appendix.

Appendix

Criterion	Assessment	Selection principle
Format validity	Rules	$\bullet$ Nonempty user request and assistant text or actions. $\bullet$ $\geq 5$ non-system messages; $\geq 1$ matched tool-call/response pair. $\bullet$ Remove agent-harness boilerplate and tool-output formatting noise.
Task completeness	Rules + LLM	$\bullet$ Remove dangling terminal calls; reject LLM-labeled incomplete workflows. $\bullet$ Allow completed attempts with explicit external constraints.
Quality	Rules	$\bullet$ Reject $\geq 4$ consecutive calls with identical tools and arguments. $\bullet$ Reject error-marked responses exceeding 70% of all tool responses.
	LLM	$\bullet$ High-value patterns: error recovery and progressive exploration. $\bullet$ Low-value patterns: unproductive loops, dead ends, and irrelevant actions. $\bullet$ Ratios: summed annotated segment lengths / total nodes; reject low-value ratio $>50\%$ . $\bullet$ Clarify user instructions and refine tool-response content.

Appendix

Table 5: Trajectory cleaning criteria.

Domain	Records	Tokens (B)	Sampling (%)	Token mix (%)
Nemotron Interactive	278,880	1.220	37.83	40.01
Nemotron Tool Calling	707,052	3.804	36.94	36.86
Nemotron SWE	5,115	0.309	12.47	10.07
UltraChat	207,865	0.247	7.58	8.08
Nemotron Search	5,968	0.152	5.18	4.97
Total	1,204,880	5.732	100.00	100.00

Appendix

Table 8: SFT corpus statistics and mixture proportions.

Benchmark	Task	Description	Format	Length setting	Size
MRCR-ICL	Co-reference resolution	Reproduce an ordinally specified response among repeated requests, with the required prefix.	TC	$\leq$ 256K	1,711
RULER-ICL	RULER-MultiKey	Retrieve the queried key’s value among key–value pairs and distractors.	TC	4K–128K	300
	Multivalue	Retrieve all values associated with the queried key.	TC	4K–128K	300
	QA1	Answer a SQuAD question using a relevant passage among distractors.	TC	4K–128K	300
	QA2	Combine supporting documents to answer a HotpotQA question.	TC	4K–128K	300
	CWE	Identify the 10 most frequent words in a long numbered list.	TC	4K–128K	300

Appendix

Table 9: Long ICL benchmark overview. MCQ denotes multiple-choice question answering; TC denotes text completion. Length denotes nominal settings or subset labels, except for BrowseCompLong-ICL, which reports the observed complete-prompt token range; S/M/L denotes short, medium, and long subsets. Size counts evaluated queries, excluding ICL demonstrations. MRCR includes the 2-, 4-, and 8-needle settings. RULER covers six selected subtasks, each evaluated at 4K, 8K, 16K, 32K, 64K, and 128K. BrowseCompLong-ICL uses three randomly selected demonstrations; its complete prompts span approximately 20K–243K tokens in the verified run.

Benchmark	Description	Format	Size
Bug Localization	Locate the faulty line using an issue description and numbered code.	TC	114
Bug Fixing	Select the correct replacement for a specified faulty line.	MCQ	114
Code Summary	Identify the summary that matches a code snippet’s behavior.	MCQ	235
Planning Preference	Choose the next action from the task goal and interaction history.	MCQ	181
Multi-hop Replanning	Select a feasible continuation or task status after tool failure.	MCQ	210
Tool Selection	Match a user request to a tool using candidate descriptions.	MCQ	103

Appendix

Table 10: Agent ICL benchmark overview. MCQ denotes multiple-choice question answering; TC denotes text completion.

Parameter	Configuration
Maximum sequence length	262,144 tokens
RoPE base	16,000,000 (no additional scaling)
Optimizer	AdamW
Adam $\beta_{1}$ , $\beta_{2}$ ; $\epsilon$	$(0.9,0.95)$ ; $10^{-8}$
Peak / minimum learning rate	$4\times 10^{-5}$ / $4\times 10^{-6}$
Learning-rate schedule / warmup	Cosine / 10%

Appendix

Table 11: Mid-training configuration. Global batch size counts independent packed sequences, excluding sequence-parallel replicas.

Parameter	Configuration
Maximum sequence length	131,072 tokens
RoPE base	16,000,000 (no additional scaling)
Optimizer	AdamW
Adam $\beta_{1}$ , $\beta_{2}$ ; $\epsilon$	$(0.9,0.95)$ ; $10^{-8}$
Peak / minimum learning rate	$2\times 10^{-5}$ / $2\times 10^{-6}$
Learning-rate schedule / warmup	Cosine / 3%

Appendix

Table 12: Downstream SFT configuration. The positional configuration is retained from mid-training, while training sequences are capped at 128K tokens.

Question category	Reasoning requirement
Cross-document or version comparison	Identify shared conclusions, differences, or changes across related documents or versions.
Multi-hop fact composition	Use information from one source to locate and interpret the next source in a chain.
Quantitative reasoning	Combine numerical facts to compute a derived quantity with a meaningful interpretation.
Aggregation	Count, rank, deduplicate, or identify extrema across related records.
Relational reasoning	Connect evidence through calls, data flow, or relationships between definitions and their use.
Conditional reasoning	Infer an outcome under stated conditions or a hypothetical change.

Appendix

Table 13: Question and answer categories used to construct MultidocQA data. Question categories guide the reasoning task, while answer categories specify response requirements. Only compatible combinations are used.

Snapshot subtype	Question target
After edit	Recover the complete file immediately after its $K$ -th successful Edit.
Before edit	Recover the complete file immediately before its $K$ -th successful Edit, incorporating preceding updates.
Final state	Recover the complete file at the end of its reconstructible state history.
Suffix rollback	Recover the file after undoing the last $R$ successful Edits in a consecutive suffix, without crossing a Write or Delete boundary.
Provenance subtype	Question target
Final-state last writer	Identify the Edit or Write that last wrote a target fragment still present in the final state.

Appendix

Table 14: Question subtypes and answer formats used to construct State Tracking data.

Frontier Models
Models	Oolong-Synth
Models	Avg@3	BoN@3
Gemini-2.5-Flash-Thinking	78.17	91.00
QwenLong L1.5-30B-A3B	75.50	89.00
GPT-OSS-120B	68.00	84.50
Our Models

Appendix

Table 15: SFT performance on Oolong-Synth. Subscripts show gains over the corresponding Qwen3 baseline (pp). The best scores for each backbone LLM are bolded .

Models	Search					Tool Use			Coding
Models	BC	BC-ZH	WideSearch	FSC-V2	SealQA	MCP-Atlas	SkillsBench	WebArena	TB 2.1	NL2Repo	SWE-Verified	SWE-Pro
Frontier Models
Gemini-2.5-Flash-Thinking	12.72	43.42	44.09	46.44	38.75	47.47	16.67	51.98	34.83	12.40	48.60	24.08
QwenLong L1.5-30B-A3B	6.48	21.35	45.14	32.51	28.02	62.77	4.17	36.12	4.49	15.09	43.60	19.56
GPT-OSS-120B	37.91	38.41	51.01	55.75	34.23	69.83	-	43.23	22.47	6.88	67.00	40.63
Our Models

Appendix

Table 16: Overall SFT performance (BoN@3, %) across diverse search, tool-use, and coding agentic benchmarks. BC-ZH, FSC-V2, and TB 2.1 denote BrowseComp-ZH, FinSearchComp V2, and Terminal-Bench 2.1, respectively. MCP-Atlas reports unthresholded claim coverage; scoring and three-trial aggregation are defined in Appendices B.3 and C.2 . Empty cells indicate unavailable or invalid evaluation results.

Figure 9: General ICL performance after mid-training on ten evaluations. Outer panels show task scores on individually zoomed axes; the center shows their unweighted mean.

Figure 10: Long-context and Agent ICL accuracy gains over Raw Trajectories (pp). Dashed lines separate the Raw reference row.

Figure 11: Correlations between long-context ICL gains and downstream agent gains after SFT. Each of the five panels contains six matched training variants. Axes show gains over Qwen3-14B-Base and SFT without mid-training, respectively (pp). Dashed lines are descriptive linear fits;

\rho

denotes Spearman correlation.

Figure 12: Long-to-Agent correlations after SFT (BoN@3) across five Qwen3-14B variants: Retrieval, MultidocQA, State Tracking, All, and Raw Trajectories. Cells show Spearman $\rho$ with average ranks for ties on a shared $[-1,1]$ scale.

TrajLong / Retrieval
User	Run today’s SEO workflow for example-store.example. Review the saved history and strategy, rotate the target markets across India, Sri Lanka, and Nepal, and avoid duplicate publications.
… [1 assistant turn omitted; $\approx$ 2.9k tokens] …
Assistant	Answer: Today is Monday, so I will check the site’s indexing status and recent articles, then prepare the weekly SEO report.
	Tool Call: web_search({ "query": "site:example-store.example", "max_results": 10 })
Observation	[example-store.example (search results)]
	Result 1: finding your goldmine: the ultimate guide to successful dollar store site selection …

Appendix

Table 17: A TrajLong (Retrieval) example.

TrajLong / MultidocQA
You are given a set of evidence items below, each marked with a unique ID, together with its title and version when available. Read the evidence carefully, identify the items relevant to each question, and answer the question or questions that follow. Evidence IDs are provided to help locate and distinguish evidence items. Use them to keep track of the relevant evidence while solving each question. Follow each question’s own output requirements.
… [103 additional evidence items omitted from this compact display] …
[EVIDENCE id=file-d4ba5ce340e9 \| title=AuthRepository.kt \| version=1] … class AuthRepository(…) { … private val api: AuthApi by lazy { ApiClient.createApi(AuthApi::class.java) } suspend fun loginWithPassword(…): Result<LoginResponse> = safeCall { val response = api.loginWithPassword(LoginRequest(phone, password)); … } suspend fun loginWithSms(…): Result<LoginResponse> = safeCall { val response = api.loginWithSms(SmsLoginRequest(phone, code)); … } … }
[EVIDENCE id=file-ce5214a29b14 \| title=AuthApi.kt \| version=1] … interface AuthApi { @POST("/api/v1/auth/tenant/login") suspend fun loginWithPassword(…) … @POST("/api/v1/auth/sms/send") suspend fun sendSmsCode(…) … @POST("/api/v1/auth/sms/login") suspend fun loginWithSms(…) … … }
[EVIDENCE id=file-f879713c7014 \| title=AppNavGraph.kt \| version=1] … LoginScreen(onLoginSuccess = { val user = authState.user val homeRoute = when { user?.isAdmin() == true -> Screen.AdminHome.route else -> Screen.WorkerHome.route } navController.navigate(homeRoute) { popUpTo(Screen.Login.route) { inclusive = true } } }) …
[EVIDENCE id=file-8c0dc44e8e97 \| title=NotificationApi.kt \| version=1] … @GET("/api/v1/notifications") suspend fun getNotifications( @Query("page") page: Int = 1, @Query("limit") limit: Int = 20, @Query("is_read") isRead: Boolean? = null): PaginatedResponse<NotificationResponse> …

Appendix

Table 18: A TrajLong (MultidocQA) example.

TrajLong / State Tracking
The following is a historical agent trajectory enclosed between [BEGIN_OF_TRAJECTORY] and [END_OF_TRAJECTORY].
[BEGIN_OF_TRAJECTORY]
System	You are an AI coding assistant.
… [20 assistant turns omitted; $\approx$ 28.0k tokens] …
User	Can we adjust the font size? It looks a little too small.
Assistant	Tool Call: Read({"file_path": "…/src/renderer/index.css"})

Appendix

Table 19: A TrajLong (State Tracking) example.

LLM Agent Training

Large language model agent training teaches models to plan, use tools, and learn from multi-step interaction with environments.

See all 588 papers

ACC: Compiling Agent Trajectories for Long-Context Training
May 21, 2026Qisheng Su, Zhen Fang, Shiting Huang +8Supervised Fine-Tuning Long-Context Modeling
MoE Key Lab of BIPC, University of Science and Technology of China · Shanghai Innovation Institute · Shanghai AI Laboratory
KLong: Training LLM Agent for Extremely Long-horizon Tasks
Feb 19, 2026Yue LiuLLM Agent Training Long-Horizon Agent Tasks
NUS
MidTool: Mid-training Data Synthesis for Agentic Tool Use
Aug 20, 2026Fengqing Jiang, Yite Wang, Boyi Liu +5LLM Tool Use LLM Agent Training
University of Washington · Snowflake · University of North Carolina at Chapel Hill

May 21, 2026J/K move · Enter open · S save

ACC: Compiling Agent Trajectories for Long-Context Training

Qisheng Su, Zhen Fang, Shiting Huang, Yu Zeng+7

MoE Key Lab of BIPC, University of Science and Technology of China · Shanghai Innovation Institute · Shanghai AI Laboratory

Supervised Fine-Tuning Long-Context Modeling Long-Context Inference

Open paper AskPDF arXiv

Recent development of agents has renewed demand for long-context reasoning capacity of LLMs. However, training LLMs for this capacity requires costly long-document curation or heuristic context synthesis. We observe that agents produce massive trajectories when solving problems, invoking tools and receiving environment observations across many turns. The evidence needed to answer the original question is thus scattered throughout these turns, requiring integration of distant context segments. Nevertheless, standard agent SFT masks tool responses and only trains turn-level tool selection, creating a supervision blind spot where these scattered signals go unused. We propose Agent Context Compilation (ACC), which converts trajectories from search, software engineering, and database querying agents into long-context QA pairs that combine the original question with tool responses and environment observations gathered across multiple turns, training the model to answer directly without tool use. This makes the dependencies between the question and the evidence explicit, enabling direct supervision of long-context reasoning over distant segments without additional annotation. ACC is a simple but effective approach that can be combined with any existing long-context extension or training method, providing scalable supervised fine-tuning data. We validate ACC on long-range dependency modeling tasks through MRCR and GraphWalks, challenging benchmarks requiring cross-turn coreference resolution and graph traversal over extended contexts. Training Qwen3-30B-A3B with ACC achieves 68.3 on MRCR (+18.1) and 77.5 on GraphWalks (+7.6), results comparable to Qwen3-235B-A22B, while preserving general capabilities on GPQA, MMLU-Pro, AIME, and IFEval. Further mechanism analysis reveals that the ACC-trained model exhibits task-adaptive attention restructuring and expert specialization.

Abstract

LLM Agent Training

Explore similar work

ACC: Compiling Agent Trajectories for Long-Context Training

KLong: Training LLM Agent for Extremely Long-horizon Tasks

MidTool: Mid-training Data Synthesis for Agentic Tool Use