Trained Agentic Context Management

Bryce Sandlund

cs.LGOct 1, 2026

Trained Agentic Context Management

Authors: Bryce Sandlund

Abstract

We study long context language models. Instead of training long context natively, or designing a long context harness, we train a model over the simplest possible harness: a tool to call itself with any specified prompt and a tool to read tokens in a range from the input context. We finetune Qwen3.6-35B-A3B on a diverse synthetic dataset using this harness. With only 8,000 tokens of context, our small model is as strong as GPT-5.4 with 1M tokens of context on the OOLONG-synth benchmark when document length exceeds 40K tokens.

Figures & tables

Figure 1: Results on RULER and OOLONG-synth. Our finetune runs in our harness with an 8K token per-agent context limit. GPT-5.4 (medium reasoning) and base Qwen3.6-35B-A3B (thinking disabled) receive the full document with the prompt (see Section 4 and Appendix A for more details). Base Qwen and GPT-5.4 are served with 64K and 1M token context limits, respectively. The OOLONG random baseline follows Bertsch et al. (2025, §2.3) .

Panel of Figure 2 — Figure 2: Tree reduce , for order-independent tasks (aggregation, retrieval). Non-leaf nodes delegate half their range to subagents. Leaves read one chunk and compute a leaf op which is aggregated up the tree.

Panel of Figure 3 — Figure 3: Sequential fold , for order-dependent tasks (variable tracking, running state). Each agent reads the next slice, updates the accumulator, and delegates the rest of the document to the next agent.

Panel of Figure 4 — Figure 4: Boundary handling. A leaf/fold agent owns the data that starts in its range. It reads 200 additional tokens at a time until the last item in its range is not cut off (additional calls not pictured in Figures 2 and 3 , for brevity).

RULER (mean of 13 tasks)	10K	20K	40K	80K	160K	320K
GPT-5.4 (full document)	0.954	0.938	0.954	0.954	0.929	0.954
Qwen3.6-35B-A3B (full document)	0.923	0.938	0.938	—	—	—
Finetuned Qwen3.6-35B-A3B (our harness)	0.949	0.949	0.929	0.885	0.868	0.858
OOLONG-synth (mean of 3 families)	10K	20K	40K	80K	160K	320K
GPT-5.4 (full document)	0.718	0.666	0.600	0.539	0.556	0.479
Qwen3.6-35B-A3B (full document)	0.536	0.558	0.388	—	—	—

Table 1: Results by document length (tabular view of Figure 1 ). Our finetune runs in our harness with an 8K per-agent context limit. GPT-5.4 (medium reasoning, 1M token context limit) and base Qwen3.6-35B-A3B (reasoning disabled, 64K token Tinker serving limit) receive the full document in one call (see Appendix A for details). Bold marks the highest score per length.

RULER	10K	20K	40K	80K	160K	320K
NIAH (8)	1.00	1.00	1.00	1.00	1.00	1.00
VT	1.00	1.00	1.00	0.92	0.84	0.64
CWE	1.00	0.94	0.74	0.58	0.24	0.18
FWE	0.93	1.00	0.93	1.00	1.00	0.93
QA-1	0.60	0.60	0.80	0.60	0.60	0.60
QA-2	0.80	0.80	0.60	0.40	0.60	0.80

Table 2: Results by task family . Left: On RULER, our model chooses a lossy aggregation strategy for the common words “CWE” problem type, which degrades when common words are diluted by fillers as document length grows. Right: On OOLONG-synth, leaf classification and merge errors compound at length. Distribution mismatch also arises at 160K+ document length (it sees

\leq

14K doc size in training).

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Benchmark	System prompt task context	Root user message
OOLONG-synth	OOLONG’s dataset description, verbatim (its intro, the number of items, the label set)	OOLONG’s question, verbatim, + “Put your final answer inside \boxed{}.”
RULER	None (empty)	RULER’s task template, verbatim, with {context} replaced by a pointer to the document (below) + “Reply with the answer inside \boxed{}.”

Appendix

Table 3: Benchmark text fields. The task context is placed in the system prompt, so every subagent inherits it without the parent restating it. For RULER, {context} in the template is replaced by: “[The relevant text is in a separate document accessible via the read_chunk tool — see the system prompt for usage.] (Document length:

\langle

\rangle

tokens.)”

Event	Tool result returned to the calling agent (verbatim)
spawn_subagent
Subagent ends with a boxed answer	The content of its last closed \boxed{…}
Subagent exceeds its context budget	agent overflowed context
Subagent stops without a boxed answer	agent did not box an answer
Tree node cap reached	Error: rollout node budget reached. Solve directly.
read_chunk

Appendix

Table 4: Harness responses. An agent’s budget counts its whole rendered conversation (system prompt, tool schema, user message, its own turns and every tool result); each generation is capped at the remaining room. A subagent that answers returns only its boxed content, so its reads and reasoning never enter the parent’s context.

Task	Strategy	$N$	Example question / what the leaf does
Order-independent aggregation over structured records (one record per line)
synth_sum	tree	60	SUM of amt over all records
synth_count	tree	60	how many records have flag=Y
synth_max	tree	60	largest amt
synth_min	tree	60	smallest amt
synth_sumwhere	tree	60	SUM of amt over only the flag=Y records

Appendix

Table 5: Training tasks.

N

is the number of problems per task. Tree : recursive halving, with each parent merging its two children’s partials (for retrieval, the half that finds the answer wins). Fold : one agent reads a slice and hands the rest of the document and its accumulator to the next. Every agent in each tree is a training example.

Long-Context Language Modeling

Long-context language modeling studies how language models can use, retain, and efficiently process information across very long sequences.

See all 182 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
OctoLong: Mid-Training On Cross-Repository Code Contexts Enhances Long-Context Modeling
Aug 5, 2026Indraneil Paul, Falko Helm, Goran Glavaš +1Software Engineering Long-Context Language Modeling
UKP Lab & Center for Applied Cybersecurity ATHENE, TU Darmstadt · CAIDAS, University of Würzburg
ContextPilot: Teaching Agents for Proactive Context Management via Fine-grained RL
Aug 28, 2026Zhuoshi Pan, Qizhi Pei, Junru Lu +4LLM Context Management Long-Horizon LLM Agents
Tsinghua University · Tencent Youtu Lab · Shanghai AI Lab

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

Long-Context Language Modeling

Explore similar work

ACC: Compiling Agent Trajectories for Long-Context Training

OctoLong: Mid-Training On Cross-Repository Code Contexts Enhances Long-Context Modeling

ContextPilot: Teaching Agents for Proactive Context Management via Fine-grained RL