acceptodds
Under review as a conference paper at ICLR 2027

MemoryLOD: Source-Grounded Navigation over Multiresolution Memory

Abstract

Long-horizon language agents must retrieve useful details from histories that grow beyond what they can read at once. Compact summaries can lose information needed by later questions, while reading entire histories consumes context and computation. We present MemoryLOD, a memory framework that organizes source files into a navigable graph at multiple levels of detail. Inspired by level of detail in computer graphics, the graph combines coarse views for orientation with fine-grained access to source content. An agent explores a bounded view of this graph, follows local and cross-resolution links, and opens sources to gather evidence for its question. The graph is constructed independently of the query, while navigation and stopping are guided by the evidence the question requires. Current source files remain authoritative: derived representations guide retrieval, and only reopened, verified source content is used as answer evidence. We evaluate MemoryLOD on BEAM, LongMemEval-V2, and UltraDomain, covering conversational memory, multimodal interaction histories, and document question answering. The evaluation examines answer quality and evidence provenance, with diagnostic analyses that distinguish failures to find relevant information from failures to use it.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
language model agents, long-term memory, retrieval-augmented generation, graph navigation, multiresolution representations

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_hMamOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces MemoryLOD, a memory framework for long-horizon language agents. It organizes source records into a graph with multiple levels of detail. The navigation scene presents detailed previews near the current focus and coarse summaries of other regions. Each visit automatically changes the allocation of detail. Routing digests are constructed independently of the question and require no language-model calls. Source openings provide the evidence for answering. A bounded review stage can request additional evidence and revise the answer. The evaluation covers BEAM, LongMemEval-V2, and AMA-Bench. Controlled comparisons support the complete scene policy. Gains over Hybrid RAG are concentrated in LongMemEval-V2. The paper also reports source replay and diagnostic ablations.

Critical strengths

1.The paper proposes a multiresolution navigation framework for agent memory. It improves answer quality over Local Tree across conversational, multimodal, and agent-history tasks. The gains are especially large on LongMemEval-V2. The controlled comparison uses the same model and underlying memory graph. These results identify memory presentation as an important factor in answer quality. This finding provides a useful design principle for agent memory interfaces.

2.The paper evaluates the review stage across several memory settings. Review improves accuracy on both LongMemEval-V2 settings. It reduces AMA-Bench accuracy from 70.00% to 64.40%. These results reveal substantial task variation in the effect of review. The analysis provides an empirical basis for assessing the benefits and limits of review in memory agents.

Critical weaknesses

1.The published comparisons in Figures 4–6 use different models and evaluation settings. The matched comparisons include Local Tree and Hybrid RAG. Local Tree is an internal scene variant. Hybrid RAG uses one reader call. The experiments do not include matched results for established navigation methods such as MemWalker or NaviRAG. The gains over Hybrid RAG on BEAM 100K and AMA-Bench also have paired 95% intervals that include zero. These results provide limited evidence for the method’s advantage over competitive navigation approaches under comparable inference budgets.

2.On AMA-Bench, accuracy decreases from 70.00% in the first draft to 64.40% in the final answer. Review changes the BEAM scores only slightly. The complete post-draft stage accounts for 25.03% of total input and output tokens across 812 executions. The paper provides qualitative examples of review failures. It does not report the frequencies of corrected errors and newly introduced errors. It also does not quantify the main causes of the AMA-Bench regression. This limits the explanation of harmful revisions and the justification for the additional computation.

3.The controlled comparisons use only part of the available test questions. For example, the BEAM 100K comparison includes only 100 of the 400 questions. Each question is evaluated only once. The paper does not report repeated runs or controlled comparisons on the full test sets. The stability of the small performance gains therefore remains unclear.

Questions for authors

1.The paper lacks matched experiments against established navigation methods. Its advantage under comparable inference budgets remains unclear. Can the authors add a comparison with MemWalker or NaviRAG using the same model and test questions? The comparison should use comparable inference budgets and report both answer quality and actual token use.

2.The paper lacks repeated-run validation of the small performance gains. The controlled comparisons also cover only part of the available test questions. Can the authors add independent runs on a fixed subset? These runs should include MemoryLOD and its baselines. Mean scores and variation across runs should be reported.

2 comments

Thanks for the clear review. Here are answers to your two questions and to the point about the AMA-Bench regression after answer review.

Q1. Matched comparison with MemWalker or NaviRAG, same model and questions, with token use.

We ran NaviRAG, with the same model (Qwen3.8-27B with reasoning), the same 20 BEAM 100K questions, a model-call budget of 25 per question, and native scoring.

MethodScore (fractional credit ×100)Query model callsQuery input / output tokensConstruction tokens
MemoryLOD33.082202.45M / 0.90M0
NaviRAG (adapted port)13.262200.38M / 0.76M123.4M

MemoryLOD wins on 7 questions, loses on 3 and ties on 10. Three of NaviRAG's 13 hierarchy constructions ran past a 4,000-call cap, so six questions count as failures; on the other 14 the scores are 27.62 vs 18.95. Model calls are matched, and output tokens at query time are about the same (0.90M vs 0.76M). MemoryLOD does read more input (2.45M vs 0.38M). That's the trade-off for using zero language-model calls for construction, where NaviRAG used 41.5M input and 81.9M output tokens. This is an adapted, capped port on a small set of questions, not a reproduction of the published results, and we didn't run MemWalker.

Against Hybrid RAG's single reader call, MemoryLOD uses 5.8–10.7x the tokens on the primary settings. Saving tokens isn't the goal. MemoryLOD aims to get as much relevant opened source evidence as possible to the reader within adjustable per-question resource limits (model-call budget, source-opening limit, evidence character budget). We report one setting of those limits, and we don't claim an advantage at equal cost.

Q2. Independent repeated runs on a fixed subset.

We couldn't finish repeated runs during the discussion period. What we do have is a paired 95% interval, resampled over source collections, for every matched difference, and we only claim the differences whose intervals exclude zero. Against Local Tree: BEAM 100K (+11.96 [+6.03, +18.72]), LongMemEval-V2 Medium (+50.00 [+32.61, +65.22]) and AMA-Bench (+8.40 [+1.59, +15.02]). Against Hybrid RAG: LongMemEval-V2 Medium (+45.65 [+28.26, +60.87]). We don't claim anything from the small differences against Hybrid RAG on BEAM 100K and AMA-Bench.

Weakness 2. Frequencies and causes of the AMA-Bench regression.

Looking at all 250 pairs of first draft and final answer, answer review fixes 7 answers and breaks 21 (154 stay correct and 68 stay incorrect), so accuracy goes from 70.00% to 64.40%. We used Codex (gpt-6-astra) to inspect the causes, shown in the table below:

Cause of regression (our coding of the edits)Count
A justified inference is withdrawn because the critic asks for explicit logged support11
Same central answer with added hedging, which the judge no longer credits7
The answer moves from the turn of an action to the turn where its result is observed2
A correct count gets an abstention put in front of it1

All 7 fixes came after the navigator acquired new opened source evidence following the critic's request; 15 of the 21 regressions involved no new evidence. So review helps when it sends the navigator back for evidence, and hurts when it only raises the bar for support. On AMA-Bench the post-draft stage costs 24.77% of input and output tokens and isn't justified by accuracy. We'll add this analysis, report first-draft and final-answer scores side by side, and present answer review as a stage whose effect depends on the task.

Reviewer_hMamOpenReview ↗

Response to author

Thank you for the new NaviRAG experiment and the AMA-Bench analysis. The counts of corrected and newly introduced errors address my concern about answer review. The token costs also help explain its trade-offs.

The NaviRAG comparison still leaves the cost question open. Both methods have the same call limit, but MemoryLOD uses more input tokens at query time. Its lower construction cost is useful, but the overall trade-off needs further evaluation. Repeated runs are also missing. The confidence intervals do not measure variation across runs. The results therefore do not yet establish a stable advantage under comparable costs. For these reasons, I maintain my Weak Reject recommendation.

Reviewer_48YAOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes MemoryLOD, a framework inspired by level of detail (LOD) techniques in computer graphics. It organises source content into a multiresolution graph and allows an agent to navigate memory through dynamically composed scenes. Each scene allocates finer-grained representations near the current focus while keeping coarse summaries of distant regions. The paper evaluates MemoryLOD on BEAM, LongMemEval-V2, and AMA-Bench, and investigates answer quality, structural coverage, source fidelity, navigation behaviour and answer review.

Critical strengths
  1. I think the adaptive memory navigation is a bit interesting perspective to some degree.

The LOD inspired observation policy offers a potentially useful way to allocate limited context across local details and global navigation directions. In addition, separating routing representations from source evidence seems also a reasonable design choice for evidence grounded retrieval.

  1. The experiment is established on recent benchmarks with controlled comparisons.

The paper evaluates relatively recent memory benchmarks and provides matched comparisons with Local Tree and Hybrid RAG. The improvements offer useful evidence of the proposed scene policy's effectiveness within the evaluated framework.

Critical weaknesses
  1. I think the research motivation and method advantages over existing hierarchical memory frameworks are not sufficiently established.

The Introduction primarily motivates MemoryLOD through the limitations of compact summaries and full context reading, etc. However, these challenges seem to have been addressed or at least partly addressed by existing hierarchical and adaptive memory retrieval frameworks. Prior works such as MemWalker, ReadAgent, and NaviRAG, etc, support hierarchical navigation and access to detailed evidence, and agent memory systems such as xMemory, GAM, and EverMemOS, etc, also apply structured memory organisation and adaptive retrieval.

In addition, although Section 2 describes differences from several navigation methods, such as NaviRAG and SLoD, it does not explain what important limitations of existing approaches remain unresolved or why the proposed mixed resolution observation policy is necessary/have some advantages in existing memory scenes. Moreover, the paper does not position MemoryLOD against recent hierarchical agent memory frameworks or compare their retrieval frameworks.

  1. I think the experiments do not fully isolate the benefits of the proposed mixed resolution navigation policy.

The experiment shows improvements of MemoryLOD compared to Local Tree, but this baseline is a simplified navigation variant rather than an established adaptive hierarchical retrieval method. The results do not fully tell us the benefits of MemoryLOD relative to existing navigation frameworks.

Furthermore, the paper emphasises full structural coverage, but this metric seems to only guarantee that source units remain represented by nodes at some resolution. It does not establish that relevant information can be effectively discovered. Figure 8 shows that automatic LOD improves local source visibility, but the local only view also achieves the same visibility of 28%. The independent benefits of maintaining coarse representations of distant regions, particularly for cross-branch evidence discovery and multi-hop reasoning, are not sufficiently analysed.

Questions for authors
  1. Research motivation and Method advantages.

Could the author explain what specific limitations of existing hierarchical and adaptive retrieval approaches MemoryLOD tries to address? Especially, how does the proposed observation policy differ in its practical benefits from navigation based methods such as NaviRAG and hierarchical agent memory frameworks such as xMemory, GAM, and EverMemOS, etc? Why is dynamically displaying multiple resolutions within a single scene more effective than existing coarse-to-fine retrieval approaches?

  1. Contribution of mixed resolution navigation.

Could the author further isolate the benefits of mixed resolution scene allocation from hierarchical navigation itself? In particular, beyond structural coverage and local target visibility, does retaining coarse representations of distant regions improve cross-branch evidence discovery, multi-hop reasoning, or navigation efficiency?

  1. Evaluation full benchmark and efficiency.

I understand computed limitaion. But all the benchmark only evaluated by 10% are not enough I think. Therefore, could the author provide a more comprehensive evaluation, at least including one complete benchmark such as AMA-Bench, and existing matched comparisons with hierarchical memory systems? Additionally, how does MemoryLOD compare under equivalent end-to-end inference budgets, considering the costs of memory construction, multiple navigation steps, source openings, and answer review?

Overall, I think the problem is meaningful and appreciate the experimental effort. However, my concerns concern the core framing, method justification, and empirical positioning, rather than presentation alone, and may be difficult to fully address through clarification alone.

2 comments

Thanks for the detailed review. Here are answers to your three questions.

Q1. What limitation does MemoryLOD address, and how does it differ from NaviRAG, xMemory, GAM and EverMemOS?

In coarse-to-fine methods, the agent or retriever sees one level, or one retrieved subset, at a time, and a language model writes the hierarchy. Once it commits to a branch the other directions drop out of view, and building the hierarchy is expensive. MemoryLOD does three things differently:

  • One scene, several resolutions. At every step each source unit is represented by some scene card, so any other region is a single visit away. NaviRAG instead enters subtrees picked by vector retrieval and walks down them node by node. xMemory, GAM and EverMemOS retrieve from their structures; they don't give the agent a scene to navigate.
  • Zero language-model calls for construction. Routing digests are extracted from the sources with deterministic rules. The four methods above use a language model to build their hierarchies or memory units.
  • Only opened source evidence supports answers. Routing summaries guide navigation and nothing more.

We aren't claiming a mixed-resolution scene always works better, and we don't have matched results against xMemory, GAM or EverMemOS. We'll rewrite the Introduction around this narrower limitation.

Q2. Is mixed-resolution allocation isolated from hierarchical navigation?

You're right that Figure 8 and structural coverage don't show this. Our end-to-end evidence is the Local Tree comparison. Local Tree is hierarchical navigation over the same graph, with the same navigation actions, reader and resource limits; its scene has only the fixed overview and the local tree neighborhood around the current focus.

Setting (questions)MemoryLODLocal TreeQuestions with no source opening (MemoryLOD / Local Tree)Opened source evidence from two or more trajectories or sessions
LongMemEval-V2 Medium (46)67.3917.390.0% / 82.6%58.7% / 4.3%
AMA-Bench (250)64.4056.000.0% / 12.4%single source
BEAM 100K (100)40.4528.490.0% / 38.0%44.0% / 13.0%

Scores are accuracy (%); BEAM 100K is fractional credit ×100, here and below.

On navigation efficiency, we went back through the recorded executions: MemoryLOD gets to its first source opening within about two visits. Local Tree needs three to seven, when it gets there at all. On AMA-Bench the gain comes at the same token cost.

On multi-hop reasoning, the biggest BEAM 100K gaps over Local Tree are on multi-session reasoning (32.50 vs 10.00) and summarization (23.75 vs 2.00), with ten questions per ability.

Q3. Full benchmark, matched hierarchical baseline, end-to-end budget.

On coverage, BEAM 100K already uses the complete 400-question benchmark; the other settings use the fixed 10% selections. We couldn't finish the full AMA-Bench (2,496 questions) during the discussion period, and we commit to reporting it in the camera-ready version.

For a matched hierarchical baseline, we ran a new comparison with NaviRAG: same model, same 20 BEAM 100K questions, and a model-call budget of 25 per question.

MethodScoreQuery input tokensQuery output tokensConstruction input / output tokens
MemoryLOD33.082.45M0.90M0 / 0
NaviRAG (adapted port)13.260.38M0.76M41.5M / 81.9M

Note: three of NaviRAG's 13 constructions ran past a 4,000-call cap, so their six questions count as failures. On the other 14 questions the scores are 27.62 vs 18.95.

On the end-to-end budget, at query time the two methods produce about the same number of output tokens (0.90M vs 0.76M). Per answered question MemoryLOD actually produces fewer (44.8k vs 54.6k), because NaviRAG's six failed questions use no query calls. MemoryLOD does read more input (2.45M vs 0.38M). That's a deliberate trade-off: since MemoryLOD uses zero language-model calls for construction, the navigator reads routing summaries at query time instead, whereas NaviRAG spent 81.9M output tokens on construction before answering anything. Input tokens are also the cheaper side of inference.

Against Hybrid RAG's single reader call, MemoryLOD uses 5.8–11.8x the tokens. Saving tokens isn't the goal. MemoryLOD aims to get as much relevant opened source evidence as possible to the reader within its per-question resource limits (model-call budget, source-opening limit, evidence character budget), and those limits are adjustable. We report one setting, and we don't claim an advantage at equal inference cost.

Thank you for the response and additional analyses. I appreciate the authors' efforts, but I still have concerns about the research motivation, the clarity of the research gap, and the justification of the proposed method's advantages. Therefore, I will maintain my original rating.

Reviewer_HHjSOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

MemoryLOD is a framework for answering questions over long histories through source-grounded, multiresolution memory navigation. It constructs a query-independent hierarchy with relation links and renders a bounded scene that exposes fine-grained information near the current focus while retaining coarse representations of other regions. The agent separately visits nodes and opens original sources, tracks unresolved evidence requirements, and may perform bounded answer review. Evaluation covers BEAM, LongMemEval-V2, and AMA-Bench, with an additional UltraDomain document-QA study. Validation includes matched Local Tree and Hybrid RAG comparisons, scene and source-opening diagnostics, and source/execution replay.

Critical strengths
  1. A concrete mechanism for allocating detail during memory navigation.

The paper turns the intuition of retaining global orientation while inspecting local details into an explicit scene-construction algorithm with resource constraints and a structural coverage invariant. Separating navigation from source opening also makes the evidence-acquisition process easier to inspect. This is a meaningful system contribution because it addresses how an agent accesses a large memory without requiring all regions to be presented at the same granularity.

  1. Controlled evidence that the proposed scene policy is useful.

The matched Local Tree comparisons retain the backbone and shared answering pipeline while changing the scene policy. MemoryLOD improves over this control across all four primary settings, providing substantive evidence that the proposed presentation mechanism contributes to final answer quality. The source and execution checks strengthen confidence in the reported comparisons by verifying source delivery and consistency between recorded executions and evaluation outputs.

Critical weaknesses
  1. The independent novelty and benefit of the central mechanism are not sufficiently established against close prior work.

ReadAgent already combines global gist with local source text; Corpus2Skill separates hierarchical navigation from source-grounded reading; and G²-Reader tracks evidence gaps and revises retrieval plans. MemoryLOD’s distinctive contribution therefore lies primarily in its automatic, focus-dependent reallocation of detail over a covering hierarchy cut. The Local Tree comparison supports the combined scene policy, but does not fully separate adaptive resolution from global orientation or differences in context volume. Moreover, the scene diagnostics use fixed recorded trajectories and the visibility of the method’s own subsequent opening target. These measurements do not establish improved relevant-evidence recall or recovery from incorrect navigation. This uncertainty concerns the paper’s principal methodological contribution, rather than merely the absence of additional baselines.

  1. The controlled results provide limited evidence of an advantage over hybrid retrieval at comparable inference cost.

The strongest gains over Hybrid RAG occur on LongMemEval-V2, with 46 questions per setting. The paired improvement intervals include zero on BEAM and AMA-Bench. In the document-QA transfer study, MemoryLOD scores 8.005 versus 8.293 for Hybrid RAG, with a paired difference of −0.288 and a reported 95% interval of [−0.470, −0.114] (Appendix C). These results indicate substantial task dependence, although they do not invalidate the positive memory-task results. Additionally, Hybrid RAG uses a single reader call, whereas MemoryLOD performs multiple navigation, reading, and review calls. Consequently, the current comparisons do not establish whether the additional inference cost yields a favorable quality–cost trade-off, which is consequential for a method motivated by constrained memory access.

Questions for authors
  1. Mechanism isolation.

Does Local Tree preserve the same global overview and structural coverage as MemoryLOD, and how do their actual scene-token counts compare? Which reported comparison isolates adaptive detail allocation from differences in visible content or context volume?

  1. Inference-cost comparability.

Using the recorded executions, could you report total input/output tokens and model calls for MemoryLOD, Local Tree, and Hybrid RAG on the matched questions? What conclusions about comparative efficiency are supported by these measurements, given Hybrid RAG’s single-reader configuration?

  1. Scope and source of the gains.

What do the existing error analyses indicate about the much larger improvements on LongMemEval-V2 than on BEAM, AMA-Bench, and UltraDomain? In particular, can these improvements be attributed to better evidence acquisition through the scene policy, as distinct from interface-specific source handling or post-draft answer revision?

1 comment

Thanks for the careful review. Here are answers to your three questions, with some new measurements.

Q1. Mechanism isolation.

Yes. Local Tree always shows the fixed overview, so it has the same structural coverage as MemoryLOD. The rest of its scene is just the local tree neighborhood around the current focus, in place of MemoryLOD's hierarchy cut and source previews. We didn't log scene token counts for Local Tree; total tokens are under Q2. On AMA-Bench the two use almost the same tokens (0.99x), and MemoryLOD is 8.40 points higher [+1.59, +15.02].

You're right that the controlled comparisons only test the joint effect of structural coverage and adaptive detail. The scene policy works as a whole: focus-dependent detail allocation decides which scene cards are shown, so the two can't be varied independently. What can be separated is the amount shown, which is why we report token counts, and where the detail goes, which our new comparison changes. Root-cut refines the hierarchy cut from the root instead of around the focus; everything else stays the same, and both keep full structural coverage.

SettingQuestionsFocus-cut (paper)Root-cutFocus wins / losses / ties
LongMemEval-V2 Medium103/101/102 / 0 / 8
BEAM 100K200.33080.44184 / 5 / 11

It's a mixed result on few questions. So we'll claim the benefit of the mixed-resolution scene as a whole, which the Local Tree comparison tests, and not of refining around the focus by itself. We'll also make clear that the scene study at fixed recorded navigation states is descriptive only.

Q2. Inference cost on matched questions.

Per question: model calls / input tokens / output tokens (output includes reasoning; all recorded attempts counted).

Setting (questions)MemoryLODLocal TreeHybrid RAGMemoryLOD ÷ Hybrid RAG
BEAM 100K (100)9.9 / 109.6k / 37.3k10.8 / 67.4k / 25.3k1 / 10.6k / 3.1k10.7x
LongMemEval-V2 Small (46)11.0 / 296.0k / 29.3k9.7 / 81.9k / 18.9k1 / 53.9k / 2.4k5.8x
LongMemEval-V2 Medium (46)12.4 / 420.2k / 37.5k9.4 / 83.7k / 19.7k1 / 53.6k / 1.6k8.3x
AMA-Bench (250)8.5 / 79.8k / 30.7k10.5 / 79.0k / 32.8k1 / 12.2k / 3.3k7.2x
UltraDomain (43)10.1 / 101.4k / 28.2k10.9 / 58.3k / 16.0k1 / 7.4k / 3.6k11.8x
  • MemoryLOD isn't trying to reduce input tokens. It aims to get as much relevant opened source evidence as possible to the reader within fixed resource limits, while Hybrid RAG stops after one retrieval and a single reader call. We'll say this plainly in the paper.
  • The cost is adjustable. It's set by the per-question resource limits (model-call budget, source-opening limit, evidence character budget, scene card budget) and by whether answer review runs. The table shows one setting; we haven't measured answer quality at smaller limits.
  • What the comparison supports. At this setting MemoryLOD costs much more than Hybrid RAG, and that only pays off where retrieval alone falls short: +37 and +46 points on LongMemEval-V2, no clear gain on BEAM 100K or AMA-Bench, and a loss on UltraDomain. Against Local Tree, under the same resource limits, the AMA-Bench gain comes at the same token cost.

Q3. Source of the LongMemEval-V2 gains.

The gains don't come from answer review. MemoryLOD's first draft is already well ahead of Hybrid RAG:

SettingHybrid RAGMemoryLOD first draftDifference [95% interval]MemoryLOD final answer
LongMemEval-V2 Small36.9663.04+26.09 [+18.18, +33.33]73.91
LongMemEval-V2 Medium21.7458.70+36.96 [+17.39, +54.35]67.39

They don't come from interface-specific source handling either. All three methods use the same source opening with the same evidence character and image limits, and Local Tree, which differs from MemoryLOD only in the scene, scores 17.39% on Medium.

The recorded executions point to evidence acquisition. Local Tree finishes without a single source opening on 82.6% of Medium questions; MemoryLOD never does. When Local Tree does open a source (8 questions) it scores 75.0%; on the other 38 it scores 5.3%.

On BEAM 100K and AMA-Bench, Hybrid RAG's retrieval already finds the evidence, so there's less to gain. We agree the advantage depends on the task, and the UltraDomain result stays in the paper.

On novelty. We agree with how you describe ReadAgent, Corpus2Skill and G²-Reader, and we'll state our contribution more narrowly: one bounded scene that shows several resolutions at once while preserving structural coverage, with routing digests built using zero language-model calls.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.