---

What's New

[!NOTE] SForge 1.1.0

• E2B backend: --backend e2b runs everything in E2B Sandboxes, no Docker or cluster needed. Official Templates are published under the edgebench namespace on E2B. See the E2B example.
• OpenCode agent: --agent opencode joins Claude Code and Codex as a built-in scaffold.
• Reasoning effort: --effort {low,medium,high,max} sets reasoning effort uniformly across agents.

---

Overview

EdgeBench is a benchmark of 134 real-world tasks for evaluating how autonomous AI agents learn from real-world environments. Instead of measuring one-shot performance, EdgeBench places agents in executable task environments with realistic, multi-level feedback and lets them iterate for 12+ hours per task — tracking the full trajectory of improvement, not just the final score. We publicly release 51 tasks along with the full evaluation framework.

Analyzing ~38,000 hours of agent interaction on all 134 tasks, we find that performance follows a log-sigmoid scaling law as a function of interaction time ($R^2 = 0.998$). See the tech report for details.

Leaderboard

Full Benchmark (134 tasks)

Category Scores @12h (134 tasks)

Model	Scientific & ML	Systems & SE	Optimization	Knowledge	Formal	Games
Claude Opus 4.8	48.5	67.4	36.5	47.0	55.0	39.3
GPT-5.5	44.3	65.0	33.6	45.7	50.0	39.1
GPT-5.4	33.5	54.1	27.9	38.8	40.8	29.0
GLM-5.1	33.8	50.9	26.4	43.5	24.6	29.3
DS-V4-Pro	30.0	43.0	21.5	37.0	14.1	16.9

Open-Source Subset (51 tasks)

Category Scores @12h (51 tasks)

Model	Scientific & ML	Systems & SE	Optimization	Knowledge	Formal	Games
Claude Opus 4.8	38.9	62.0	38.2	38.7	40.9	39.3
GPT-5.5	33.2	60.5	32.3	38.4	49.0	39.1
GPT-5.4	24.6	50.1	29.9	31.6	30.2	29.0
GLM-5.1	26.8	43.6	26.7	31.0	19.9	29.3
DS-V4-Pro	31.1	37.6	24.1	33.2	12.7	16.9

Per-Task Scores by Time Budget (51 tasks)

Each model cell reports scores at @2h / @4h / @6h / @8h / @10h / @12h. Missing valid results are shown as —.

Task	Category	Opus 4.8	GPT-5.5	GPT-5.4	GLM-5.1	DS-V4-Pro
bipedalwalker_locomotion_rl	Scientific & ML	16.7/20.8/22.4/23.3/23.3/23.3	14.7/14.9/15.2/15.2/16.0/21.0	13.9/13.9/13.9/14.5/14.5/17.5	13.9/20.3/21.5/22.5/22.5/22.5	8.9/14.8/17.6/20.4/20.4/20.6
borden_source_inversion	Scientific & ML	7.5/19.8/26.2/28.5/38.5/48.4	20.1/27.0/29.4/37.8/38.1/38.5	7.2/7.3/7.6/7.9/8.0/8.0	7.0/10.3/12.0/12.3/12.5/15.1	7.0/11.6/15.1/26.6/36.7/38.2
dabic_gravity_inversion	Scientific & ML	9.5/15.2/15.7/17.4/17.5/17.5	15.9/16.2/16.7/17.0/17.2/17.3	14.6/14.6/15.5/15.5/15.0/15.0	9.2/13.7/16.0/16.5/16.5/17.1	—/12.7/12.7/12.7/13.0/13.8
graph_node_classification	Scientific & ML	59.4/62.7/65.0/65.6/66.5/66.6	54.7/55.1/55.1/55.3/55.9/56.0	54.9/56.2/56.5/56.9/57.5/57.6	49.4/52.3/52.3/52.3/52.3/52.3	46.0/48.2/49.2/51.3/51.7/51.8
ann_vector_search_qps	Systems & SE	26.2/57.0/58.6/58.7/59.4/59.7	22.3/34.3/35.1/36.0/40.0/40.7	27.5/30.2/44.5/45.2/49.7/50.2	6.7/24.4/25.6/25.6/26.1/38.3	9.4/19.6/22.4/22.8/23.8/23.8
arc_compiler_runtime	Systems & SE	49.3/52.0/52.0/52.0/52.0/52.0	55.5/56.5/60.9/70.3/71.0/72.4	45.1/46.5/49.8/49.8/50.0/50.0	47.7/48.0/48.4/48.7/48.7/48.7	40.3/41.7/44.2/44.2/44.2/44.2
exchange_core_throughput	Systems & SE	40.7/57.0/58.5/58.9/59.7/59.7	15.4/37.2/39.9/44.3/51.3/53.2	14.3/40.8/41.0/45.2/46.4/47.3	29.2/43.7/46.5/48.6/50.3/52.6	32.9/33.8/45.0/47.7/48.4/48.6
ffmpeg_swscale_reimplementation	Systems & SE	9.9/17.6/19.8/20.9/21.1/21.1	8.8/14.3/15.1/15.3/15.3/15.3	5.4/8.5/9.4/11.6/13.3/13.9	0.3/0.3/0.4/2.2/2.2/2.2	0.1/1.9/2.0/3.8/3.8/3.8
git_rewrite_in_zig	Systems & SE	22.0/22.8/22.8/22.8/23.1/23.1	16.1/16.9/17.7/18.2/18.2/18.4	9.6/13.8/14.0/14.2/14.2/15.4	12.0/20.2/23.3/23.4/23.4/23.5	8.5/13.5/16.0/17.6/17.8/17.9
integer_compression_codec	Systems & SE	69.4/69.7/74.8/74.9/75.2/75.3	61.1/67.6/73.9/73.9/74.3/74.4	38.6/40.9/41.2/42.2/42.2/42.3	23.5/27.3/28.5/28.7/28.9/28.9	15.9/16.0/16.2/16.2/16.2/16.2
juliet_vulnerability_analyzer	Systems & SE	71.9/74.9/75.4/75.6/75.6/75.6	81.0/83.2/85.4/86.8/87.4/89.8	52.9/66.1/74.3/76.0/76.8/77.2	59.3/60.7/62.8/63.5/63.5/63.5	46.8/63.1/66.1/66.2/66.2/66.2
rust_multicrate_reconstruction	Systems & SE	—/—/—/—/—/—	27.5/42.6/53.1/54.9/57.8/57.8	16.7/19.9/21.3/21.4/21.4/21.4	24.8/24.8/25.2/25.2/37.5/38.5	20.5/21.7/22.7/23.1/23.5/23.6
schemathesis_config_modernization	Systems & SE	82.5/85.0/86.1/87.4/87.4/87.7	79.1/82.2/82.9/83.2/83.6/84.0	67.2/68.8/68.8/71.7/71.7/71.9	58.3/59.7/60.4/61.2/61.7/61.7	54.3/54.3/55.3/55.3/55.3/55.6
schemathesis_datagen_pipeline	Systems & SE	68.0/70.2/70.2/70.2/70.2/70.2	54.6/54.6/56.7/56.7/56.7/56.7	56.6/56.6/56.6/56.6/56.6/56.6	62.1/64.2/64.2/67.0/67.0/67.0	47.9/50.1/52.3/52.3/52.3/52.3
schemathesis_reporting_observability	Systems & SE	73.9/75.6/76.2/76.2/76.2/76.2	76.6/76.6/76.6/76.6/77.1/77.1	70.0/73.7/74.7/75.7/76.2/76.2	61.9/61.9/61.9/61.9/61.9/61.9	59.4/62.4/63.0/63.0/65.0/65.0
vliw_kernel_optimization	Systems & SE	74.0/76.0/77.7/79.5/79.6/80.9	71.6/75.7/77.1/79.5/83.1/85.6	75.7/77.0/77.2/78.7/79.1/79.1	5.6/9.5/27.5/35.0/35.9/35.9	0.2/24.9/28.1/33.0/33.9/34.1
ad_placement_optimization	Optimization	65.2/66.1/66.9/67.1/67.4/67.7	44.0/53.3/59.5/61.6/62.9/62.9	41.8/42.4/43.1/47.7/47.9/48.1	48.7/52.7/53.3/56.5/58.5/58.8	25.5/28.5/35.2/35.8/36.2/36.2
apple_incremental_game	Optimization	42.7/44.9/45.9/48.6/49.9/50.6	26.6/29.8/30.6/32.7/33.1/33.6	28.3/30.3/32.0/33.3/33.9/34.9	19.0/19.0/19.1/19.1/19.1/19.1	19.6/19.7/19.7/19.7/19.7/19.7
equivalence_class_divide_and_conquer	Optimization	11.2/15.3/17.0/20.1/20.8/21.3	11.8/15.5/15.8/21.3/22.2/22.4	14.5/17.0/18.3/18.7/20.2/20.3	3.8/4.2/10.0/8.0/10.6/10.6	0.7/1.8/3.2/3.2/3.4/3.4
grid_turing_robot	Optimization	34.7/37.1/37.3/39.6/40.3/40.3	40.4/41.6/41.9/42.0/42.1/42.2	26.8/26.8/27.2/28.9/28.9/28.9	20.0/21.0/24.6/24.6/24.6/25.7	23.7/24.1/24.1/24.1/24.2/24.2
jagua_nesting_optimization	Optimization	11.2/17.8/24.5/31.4/41.0/44.2	15.9/19.4/20.0/20.6/21.3/21.6	22.4/23.0/23.9/24.0/24.1/24.1	8.9/9.0/10.0/12.2/12.3/12.4	10.7/20.2/23.7/26.7/28.1/28.4
molecular_self_assembly	Optimization	22.4/33.4/34.0/34.1/34.4/34.7	20.2/20.3/20.5/20.7/20.7/20.7	20.8/21.1/21.1/21.5/21.5/21.6	10.0/12.5/12.9/13.0/13.1/13.2	19.4/21.7/21.8/21.8/21.9/21.9
order_addition_permutation_optimization	Optimization	22.6/31.6/34.0/34.4/35.7/36.4	16.7/20.5/21.5/22.4/23.0/23.3	1.6/10.6/13.1/14.0/14.2/14.3	2.0/2.1/23.6/25.8/25.8/33.2	4.6/16.5/17.8/22.9/25.4/30.8
smt_solver	Optimization	10.3/17.4/19.0/23.1/23.3/23.9	7.2/7.8/8.4/8.6/8.6/8.6	6.7/7.9/8.9/9.1/9.1/9.2	2.7/2.7/2.7/2.7/2.7/3.6	1.4/2.8/3.3/3.3/3.3/3.3
treant_forest	Optimization	14.5/15.9/16.1/16.2/16.4/18.0	12.1/14.2/14.9/15.2/15.5/15.6	12.2/12.2/12.7/13.0/13.2/13.3	8.0/11.6/11.7/14.1/14.5/16.9	6.8/8.1/9.7/10.1/12.7/13.5
tree_block_partitioning	Optimization	21.5/30.1/32.4/36.8/37.7/37.7	28.8/31.1/33.0/33.0/35.0/36.4	23.1/26.8/28.8/32.9/34.3/34.3	12.1/15.4/17.1/19.3/20.3/23.4	11.2/11.8/11.9/11.9/14.6/16.1
triangulation_coloring_optimization	Optimization	70.8/71.4/71.9/73.2/73.3/73.4	73.7/74.3/74.5/75.0/75.1/75.2	74.1/74.2/74.3/74.3/74.3/74.3	68.8/71.2/71.6/72.0/72.7/73.0	56.1/58.0/59.0/59.1/59.1/59.3
vehicle_routing_time_windows	Optimization	72.5/72.6/72.9/73.6/73.7/74.0	88.7/89.0/89.4/89.7/89.7/90.8	85.3/88.6/88.7/89.5/89.5/89.6	76.6/76.6/76.6/76.6/76.6/77.9	54.7/76.8/81.9/82.2/82.9/83.1
vibrating_path_graph_coloring	Optimization	19.7/21.1/21.4/22.5/24.5/25.3	10.1/10.5/10.6/10.7/10.7/11.4	18.1/19.4/19.8/23.4/23.6/24.1	9.6/18.3/20.3/22.9/22.9/22.9	12.4/14.4/19.3/19.4/21.8/22.1
warehouse_forklift_routing	Optimization	7.7/9.5/10.4/10.5/11.1/11.2	9.8/11.0/11.8/11.9/12.1/12.6	0.0/0.0/0.0/0.0/0.0/0.0	—/0.0/0.0/0.6/0.7/0.5	0.0/0.0/0.0/0.0/0.0/0.0
wireless_electricity_layout	Optimization	6.5/13.7/14.4/14.5/14.5/14.5	6.2/6.9/7.1/7.1/7.1/7.2	10.9/11.1/11.1/11.1/11.1/11.1	7.2/9.4/6.6/8.1/9.4/9.5	0.0/0.0/0.0/0.0/0.0/0.0
college_english_exam_bank	Knowledge	24.8/28.3/34.8/35.5/35.8/39.8	24.5/35.5/35.5/35.5/37.8/37.8	30.7/30.7/31.3/34.0/34.0/34.5	22.2/26.0/29.3/30.0/32.3/32.5	19.2/21.7/22.5/22.7/29.2/34.7
cta_risk_budget_optimization	Knowledge	42.7/44.8/45.3/45.3/45.3/46.1	43.8/45.8/46.7/46.7/46.7/46.7	46.0/49.0/49.0/49.0/49.8/49.8	38.1/44.8/49.0/49.6/49.6/49.6	44.0/45.6/46.9/46.9/48.1/48.1
k12_math_recommendation	Knowledge	23.6/38.5/41.4/42.0/43.7/44.3	38.5/42.4/42.9/43.5/43.9/44.0	25.9/29.0/30.0/30.8/31.1/31.4	24.8/25.7/31.9/32.5/32.7/32.7	25.6/26.3/26.8/25.7/26.0/26.3
portfolio_risk_calibration	Knowledge	20.1/21.6/23.0/23.6/23.6/24.5	17.3/21.3/22.7/23.5/24.4/25.0	6.0/9.6/10.7/10.7/10.7/10.7	0.0/8.4/8.5/8.9/9.2/9.4	10.4/16.3/16.6/16.7/23.7/23.7
carleson_formalization	Formal	4.3/7.7/11.0/12.7/15.0/16.8	6.0/9.5/13.2/16.5/25.3/26.5	1.8/3.5/4.6/5.4/6.3/7.1	1.0/1.7/2.0/2.2/2.2/2.2	0.8/1.3/2.0/2.0/2.3/2.5
combinatorial_games_formalization	Formal	14.5/23.2/27.6/32.1/34.5/35.5	12.0/18.8/24.6/27.2/33.4/38.2	5.9/8.3/11.5/13.5/16.3/17.8	6.7/9.8/14.3/14.9/16.2/16.2	4.3/6.7/7.3/7.4/7.7/7.8
flt_regular_formalization	Formal	31.0/41.8/50.6/50.6/50.6/50.6	43.7/48.3/50.6/66.7/75.1/75.1	1.5/19.5/28.4/41.8/46.0/48.3	14.4/13.4/16.5/18.8/18.8/38.7	5.7/11.9/14.6/14.9/17.2/17.6
lean_analysis_proofs	Formal	17.9/25.1/28.6/30.2/32.6/33.0	16.8/23.2/28.4/33.9/39.0/42.5	3.6/8.1/10.8/12.9/15.5/16.4	5.2/5.9/5.9/5.9/5.9/5.9	5.8/7.3/8.2/8.8/9.3/9.5
new_foundations_consistency	Formal	28.9/36.2/50.0/62.7/64.2/65.1	13.7/38.2/55.1/56.4/65.1/66.5	3.3/12.2/14.9/20.5/30.7/39.8	2.2/3.3/5.1/21.9/24.6/27.0	2.2/3.4/6.5/7.2/10.5/11.4
ordinal_notation_well_foundedness	Formal	10.6/18.4/24.7/24.7/24.7/24.7	13.7/24.7/24.7/24.7/24.7/24.7	1.2/5.5/13.7/15.3/18.4/21.6	2.0/3.5/5.1/5.9/5.9/5.9	3.5/3.5/4.7/4.7/4.7/4.7
pfr_formalization	Formal	32.4/36.9/38.8/40.2/45.6/46.3	30.7/38.3/41.9/47.6/52.7/60.0	10.2/13.7/27.1/34.0/35.9/38.9	8.3/14.9/22.5/26.5/31.3/33.5	9.9/14.7/16.5/17.8/18.5/19.1
sphere_eversion_formalization	Formal	41.7/47.4/49.1/50.4/54.1/55.4	45.0/51.1/55.0/55.9/56.9/58.5	13.3/20.2/32.7/43.7/50.4/51.4	15.5/24.2/26.7/28.7/30.2/30.2	2.9/14.1/22.3/24.9/28.6/29.3
anchorhead_text_adventure	Games	13.3/19.3/19.7/20.3/22.3/22.3	15.0/26.3/31.7/34.3/35.3/36.3	5.0/11.7/13.0/13.3/14.7/17.7	10.7/17.3/19.7/20.3/20.3/20.3	2.0/6.0/7.3/8.0/12.3/14.7
dcss_dungeon_ai	Games	4.2/4.9/5.9/6.3/6.3/8.3	8.9/9.7/10.0/10.0/13.3/13.4	2.6/5.6/5.6/5.6/6.1/6.1	2.8/3.0/3.3/3.3/5.1/7.6	2.8/3.6/4.4/4.5/5.1/5.7
nethack_dungeon_agent	Games	29.7/35.3/36.7/37.3/41.9/41.9	16.6/17.6/18.1/20.6/21.3/22.5	10.9/14.1/15.2/15.8/17.0/20.4	2.3/2.3/15.3/21.6/21.6/21.6	1.0/1.4/2.9/3.2/3.2/3.3
openrct2_theme_park_ai	Games	24.4/24.4/26.0/26.0/27.5/27.5	28.5/28.6/32.7/37.3/37.4/37.6	23.0/23.1/23.1/23.1/23.1/23.1	35.1/36.2/36.2/36.2/36.2/36.2	24.4/24.4/24.4/26.0/26.0/26.0
openttd_transport_ai	Games	50.0/50.4/50.6/51.7/51.8/52.0	10.1/11.6/13.2/21.9/25.6/28.1	10.8/11.4/11.6/11.9/11.9/11.9	0.0/0.0/0.0/0.0/0.0/0.0	4.8/9.2/9.3/12.3/15.2/15.2
trinity_text_adventure	Games	25.0/28.0/29.3/30.0/30.0/30.0	22.3/26.7/28.7/36.3/36.3/40.0	16.3/19.7/22.7/23.3/23.7/27.0	16.0/18.7/24.3/26.0/26.7/26.7	16.3/16.3/17.7/20.0/20.0/20.3
tryst_text_adventure	Games	18.1/33.8/36.7/40.0/40.0/44.3	32.1/42.4/44.3/48.6/55.2/55.7	19.5/20.0/20.0/31.0/38.6/44.3	18.6/28.6/36.2/40.5/42.9/43.3	8.6/11.4/11.4/11.4/11.4/13.8
wesnoth_tactical_ai	Games	84.0/85.3/87.7/87.7/87.7/88.0	64.7/73.0/76.3/78.0/78.3/79.3	79.7/79.7/80.3/80.3/81.3/81.3	75.7/78.3/78.3/78.3/78.3/78.3	17.0/36.3/36.3/36.3/36.3/36.3

Task Taxonomy

EdgeBench contains 134 realistic, diverse tasks spanning six capability categories, of which 51 are publicly released. Each task is designed as a day-scale challenge with a performance ceiling high enough that no current agent can saturate it. Recorded human expert effort averages 57.2 hours per task (up to 320 hours).

Evaluation Harness: SForge

EdgeBench is powered by SForge, a two-container evaluation harness built for long-horizon agent evaluation. Each task materializes as isolated work and judge Docker images — the agent only sees the work environment, while hidden tests run in ephemeral judge containers.

Key mechanisms:

• Two-container isolation — work and judge environments are fully separated, preventing evaluation hacking at its root
• Iterative evaluation with feedback — agents don't submit once at the end for a one-shot score; instead they submit throughout the run, receive granular feedback (pass rates, failing tests, scores), and improve in a closed loop until timeout — the best result across all submissions is the final score
• Long-horizon execution — stop hooks prevent premature agent exit, auto-resume recovers from transient failures, and the Kubernetes backend enables parallel runs at scale

Quick Start

# Install (requires Docker Engine running on a Linux host)
pip install sforge

# 1. Download task definitions
sforge fetch-tasks edgebench

# 2. Pull pre-built Docker images
sforge pull --task ad_placement_optimization --registry seededge

# 3. Start judge server (separate terminal)
sforge serve

# 4. Run an agent
SFORGE_AGENT_API_KEY="sk-xxx" \
  sforge run --task ad_placement_optimization --agent claude-code \
    --model "claude-opus-4-8[1m]" --timeout 43200 --run-id edgebench-001

Step-by-step examples:

• Single task on local Docker — run one task end-to-end with Docker
• All tasks on Kubernetes — run the full suite on a K8s cluster with the official leaderboard setting
• Single task on E2B — run one task in E2B cloud Sandboxes

[!IMPORTANT]

• Official setting — leaderboard numbers use the official experiment YAMLs unchanged, including the time budget, stop hook, auto-eval, submission cooldowns, and hardware resource limits.
• Cost — one 12-hour task with a frontier model can cost hundreds to over a thousand USD; a full ~50-task run is a five-figure spend.
• Scale — the Docker backend suits only a few tasks at a time; for full-suite runs use the Kubernetes backend.

Evaluating your own model / agent:

• Your own model — the built-in Claude Code, Codex, and OpenCode scaffolds work with any compatible API endpoint: point SFORGE_AGENT_API_BASE_URL at your endpoint, set your key via SFORGE_AGENT_API_KEY, and pass your model name via --model. See Supported Agents.
• Your own agent scaffold — just add a new agent under sforge/harness/agent/ (a small Agent subclass declaring how to install and launch it) and register it in the factory, then run with --agent <your-agent>. See Custom Agents.

Full documentation: bytedance-seed.github.io/EdgeBench

Citation

If you find EdgeBench useful in your research, please cite our tech report:

@misc{edgebench2026,
  title  = {EdgeBench: Unveiling Scaling Laws of Learning from Real-World Environments},
  author = {Deyao Zhu and Xin Zhou and Shengling Qin and Xuekai Zhu and Hangliang Ding and Shu Zhong and others},
  year   = {2026},
  url    = {https://arxiv.org/abs/2607.05155},
}

License

• EdgeBench Tasks (task datasets) are released under CC BY 4.0.
• SForge (evaluation harness code) is released under the Apache License 2.0.

Contact

To evaluate on the full 134-task suite, please contact [email protected].

_{Built by ByteDance Seed}