Sequential multi-agent reasoning frameworks such as Chain-of-Agents (CoA) handle long-context queries by decomposing inputs into chunks and processing them sequentially using LLM-based worker agents that read from and update a bounded shared memory. From a probabilistic perspective, CoA aims to approximate the conditional distribution corresponding to a model capable of jointly reasoning over the entire long context. CoA achieves this through a latent-state factorization in which only bounded summaries of previously processed evidence are passed between agents. The resulting bounded-memory approximation introduces a lossy information bottleneck, making the final evidence state inherently dependent on the order in which chunks are processed. In this work, we study the problem of chunk ordering for long-context reasoning. We use the well-known Chow-Liu trees to learn a dependency structure that prioritizes strongly related chunks. Empirically, we show that a breadth-first traversal of the resulting tree yields chunk orderings that reduce information loss across agents and consistently outperform both default document-chunk ordering and semantic score-based ordering in answer relevance and exact-match accuracy across three long-context benchmarks.
Figures & tables
Figure 1: Overview of CoA with chunk order induced by a breadth-first traversal on the Chow–Liu dependency tree ( CL–order ). First, we build a complete graph on chunks, where each edge models mutual information between two chunks using an embedding-based inner-product proxy. The maximum-spanning tree over the chunk graph gives the Chow–Liu tree, which provides a second-order approximation of the global dependency structure between chunks. A breadth-first traversal on the Chow–Liu tree rooted at chunk most similar to query yields the CL–order .
LongQA (Ragas ↑ )
LongQA–MC (EM ↑ )
NarrativeQA (Ragas ↑ )
Model
Memory
Default
Dense
CL–order
Default
Dense
CL–order
Default
Dense
CL–order
Qwen-3-14B
CoA
41.43
42.25
44.12
24.89
26.20
30.26
35.72
38.26
41.23
GPT-4.1-mini
CoA
51.94
47.96
54.35
65.22
67.39
70.29
50.16
50.81
52.39
GPT-4.1
CoA
59.03
58.56
60.68
82.84
84.33
85.07
57.30
55.93
58.08
Table 1: Performance on long-context benchmarks with GPT-4.1 series and Qwen-3 . Here, ↑ indicates that higher is better. For LongQA and NarrativeQA, we use answer relevance metric from Ragas ( Es et al., 2024 ) .
BM25
Qwen-3 Embedding
Model
Memory
Default
Dense
CL–order
Default
Dense
CL–order
Qwen-3-14B
CoA
24.89
26.64
27.51
24.89
26.20
28.82
GPT-4.1-mini
CoA
72.83
75.00
73.90
73.26
75.58
77.91
GPT-4.1
CoA
82.47
83.51
84.54
82.17
83.16
85.14
Table 2: Scoring function and embedding ablation on LongQA–MC .
Figure 2: Comparison of DFS on the complete chunk graph against BFS on the Chow–Liu tree. EM reported on LongQA–MC for the three models.
Appendix figures & tables1 asset
Supplementary material from the paper’s appendix.
Appendix
Hyperparameter
Value
Hyperparameter
Value
Per-chunk token limit
8000
Qwen embedding size
4096
Summary token limit
8000
Generation temperature
0.0
OpenAI embedding size
3072
Nucleus sampling (top_p)
0.95
Appendix
Table 3: Hyperparameters for the long-context QA problem setting with CoA and Chow–Liu Trees.
Sequential memory agents process long documents by reading chunks one after another while maintaining a compact memory state, coupling document traversal to reasoning depth. This coupling introduces sensitivity to evidence placement and ties inference latency linearly to document length. We introduce PARSER, which decouples reading from reasoning. A bank of lightweight subagents each bound to a single chunk read the entire document in parallel, while a lead agent reasons in depth through iterative scatter--gather rounds: at each round it broadcasts a query to all subagents, aggregates the returned evidence, and formulates a deeper follow-up query conditioned on what has been found so far. This decoupled design concentrates all learnable behavior in the lead agent, which is optimized with reinforcement learning, while the subagents remain frozen off-the-shelf models. On multi-hop QA with contexts ranging from 7K to 896K tokens, PARSER with a 4B backbone outperforms the strongest sequential memory baseline by 5.7 points on average and by 12.0 points at 896K tokens. Scaling to a 9B backbone, PARSER surpasses DeepSeek-V4-Pro by 6.3 points. Controlled experiments confirm that PARSER is robust to perturbations in evidence position, order, and distance, conditions that cause large accuracy swings in sequential methods, while reducing inference latency by up to 11x.
Long-context tasks require LLMs to identify and preserve answer-relevant information from large contexts. Chunk-wise memory agents address this issue by sequentially reading document chunks, updating a compact memory, and generating the final answer from the accumulated memory. However, existing RL-based chunk-wise agents either rely on sparse final-answer rewards or use lexical intermediate rewards for memory and retrieval actions. These signals supervise task success or local overlap, but do not directly evaluate whether the final memory supports the ground-truth answer. We propose InfoMem, a reward mechanism for training chunk-wise memory agents that evaluates final-memory utility using answer-conditioned information. InfoMem measures how much the final memory increases the model's per-token log-likelihood of the ground-truth answer. To stabilize RL optimization, InfoMem applies this signal only to successful trajectories and normalizes it before reward composition. Under the same GRPO framework and training budget, InfoMem improves long-context memory-agent performance over comparable memory-agent RL baselines. Analyses show that effective final-memory rewards should operate on successful trajectories, be normalized before reward composition, and be conditioned on the answer rather than the query. Our code is available at https://github.com/GenSouKa1/InfoMem.
Tiancheng Han, Yong Li, Wuzhou Yu +2
1Tongji University · 2Shanghai Innovation Institute · 3Shanghai AI Laboratory
Corpus-level long-context reasoning requires agents to acquire and integrate evidence across document collections that exceed their context windows. Existing agent harnesses enable incremental access to these corpora, but acquired evidence and evolving conclusions often remain entangled in a growing interaction history. This implicit state management can lead to two coupled failure modes: unbounded exploration, where retrieval expands the interaction history without resolving decision-relevant information gaps, and evidence-state drift, where previously acquired evidence or subsequent corrections fail to inform the current conclusion. We present LongHarness, a model-agnostic harness that makes the interaction between acquired evidence and evolving conclusions explicit. Its core mechanism, Evidence-Conclusion Coevolution, uses unresolved conclusions to guide exploration toward missing information, while newly acquired evidence strengthens or revises the conclusions that directed the search. LongHarness couples this process with Evidence Lifecycle Management, which keeps accumulated evidence current and traceable throughout the interaction. Together, these mechanisms focus retrieval on remaining information needs and keep conclusions aligned with current evidence. We evaluate LongHarness on OfficeQA Pro V1 and V2, AA-LCR, and LongBench v2 using GLM-5.3-Flash and Qwen3.8-Flash. Compared with direct reasoning using the same backbone, LongHarness improves GLM-5.3-Flash exact match on OfficeQA Pro from 15.8% to 60.2% on V1 and from 4.4% to 61.1% on V2. Against state-of-the-art agent harnesses, it surpasses Claude Code by 12.0 percentage points on V1 and Codex by 18.9 points on V2. With Qwen3.8-Flash, LongHarness reaches 61.7% on V1, outperforming Claude Code by 8.3 points and Codex by 12.0 points.
Songtao Huang, Zhijie Zhong, Xiaohan He +8
Fudan University · Shanghai Artificial Intelligence Laboratory · Shanghai Jiaotong University +3