The Horizon Gap

Corpus, classification/figure scripts, and manuscript source for:

Mingguang Chen, Licheng Wang, Bo Qu. "The Horizon Gap: Planning, Memory, Execution, Training, and Evaluation for Long-Horizon LLM Agents." 2026. arXiv: submitted, ID pending — this line will be updated once assigned.

One-line summary. Frontier LLMs solve single-step reasoning problems well above human research-contribution level, yet fail in agent loops asked to complete tasks spanning hours rather than seconds. We call this the horizon gap and survey 1,547 arXiv papers (2024-2026) — collected via an eight-thread seed harvest with a disclosed, quantified two-stage bleed filter (26.8% of raw hits excluded as off-topic) plus a targeted supplement into an under-covered theory/safety category — mapping the field's response across six axes: planning & decomposition, memory & context management, execution control & recovery, training for long horizons, evaluation & measurement, and foundations/limits/safety.

Companion to The RSI Survey (arXiv:2607.07663) — that paper covers self-improvement loops; this one covers the horizon a system is asked to sustain while doing so.

Layout

artifacts/
  corpus_v2.csv                    # canonical corpus, 1,547 papers with taxonomy labels
  corpus_seed.csv                  # raw seed harvest (pre-filter)
  corpus_seed_clustered.csv        # seed harvest with KMeans cluster assignments
  corpus_seed_dropped_bleed.csv    # seed rows excluded by the bleed filter (audit trail)
  corpus_seed_excluded.csv         # seed rows excluded for other reasons (audit trail)
  corpus_supplement_dropped.csv    # supplement-harvest rows excluded (audit trail)
  benchmark_durations.csv          # Figure 5 (benchmark-duration landscape) source data
  benchmark_durations_notes.md     # sourcing/methodology notes for benchmark_durations.csv

draft/
  main.md              # manuscript source (10 sections + Data Availability)
  references.bib        # auto-generated bibliography (OpenAlex + arXiv API fallback)
  anchors.bib            # hand-curated seminal works outside the harvest window
  html-header.html, pdf-header.tex   # pandoc build headers
  figures/                # Figures 1, 3, 4, 5, 6 as generated for the paper (Figure 2 uses map_landscape.png)
  scripts/
    seed_harvest.py            # eight-thread arXiv seed harvest
    supplement_harvest.py       # targeted supplement harvest (dedupes against seed arXiv IDs)
    reclassify_corpus.py        # KMeans clustering + keyword rules + bleed filter -> corpus_v2.csv
    explore_cluster.py          # cluster inspection tooling used during taxonomy design
    build_bib.py                 # corpus_v2.csv -> references.bib
    build_figures.py             # Figure 2 (semantic map) and Figure 6 (growth timeline)
    build_concept_figures.py     # Figures 1, 3, 4 (programmatic concept diagrams)
    build_benchmark_figure.py    # Figure 5 (benchmark-duration landscape)
    build_latex.py               # Overleaf/arXiv LaTeX project build

Reproduce the corpus

Order matters — reclassify_corpus.py overwrites corpus_v2.csv from the seed harvest, so supplement_harvest.py must run after it.

cd draft/scripts
python3 seed_harvest.py          # -> artifacts/corpus_seed.csv (arXiv API, eight query threads)
python3 reclassify_corpus.py     # -> artifacts/corpus_v2.csv (SEED ROWS ONLY at this point)
python3 supplement_harvest.py    # appends targeted supplement rows to corpus_v2.csv
python3 build_bib.py             # artifacts/corpus_v2.csv -> draft/references.bib

Reproduce the figures

cd draft/scripts
python3 build_figures.py          # -> map_landscape.png, growth_timeline.png (uses seed rows only)
python3 build_concept_figures.py  # -> figure1/3/4 concept diagrams
python3 build_benchmark_figure.py # -> figure5_benchmark_duration.png (from artifacts/benchmark_durations.csv)

Build the manuscript

cd draft
cat references.bib anchors.bib > combined.bib
pandoc main.md --citeproc --bibliography=combined.bib -s --embed-resources --standalone -o survey_v1.html
pandoc main.md --citeproc --bibliography=combined.bib -o survey_v1.docx
pandoc main.md --citeproc --bibliography=combined.bib --pdf-engine=tectonic \
  -H pdf-header.tex -V geometry:margin=1in -V fontsize=11pt -V colorlinks=true -o survey_v1.pdf
python3 scripts/build_latex.py   # Overleaf/arXiv LaTeX project

License

• Code (draft/scripts/*.py) — MIT, see LICENSE.
• Manuscript, bibliography, figures, and corpus data (draft/main.md, draft/figures/, draft/references.bib, draft/anchors.bib, artifacts/) — CC BY 4.0, see LICENSE-CC-BY.

Citation

The arXiv ID is pending; this entry will be updated once assigned. Cite by title/authors in the meantime:

@article{chen2026horizongap,
  title   = {The Horizon Gap: Planning, Memory, Execution, Training, and Evaluation for Long-Horizon LLM Agents},
  author  = {Chen, Mingguang and Wang, Licheng and Qu, Bo},
  year    = {2026},
  journal = {arXiv preprint (ID pending)}
}

Contact

deepgroundingai@gmail.com · github.com/deepgrounding