Literature
Display
← All papers

Tev: Scaling Long-Horizon Agentic Environments with Dense Progress Rewards

Wenbo Hu · Shraman Pramanick · Krishna Kumar Singh · Nanyun Peng · Kai-Wei Chang · Yong Jae Lee

Preprint posted · arXiv (preprint) ↗

AI agents14 min text · CondensedOriginal paper ↗
Contents

Overview

Tev combines procedurally generated task environments with a progress-based training reward. The overview shows results on the paper’s selected computing-and-math benchmark subsets and the complete Terminal-Bench 2.1 suite.

Figure 1: We demonstrate Tev ’s results on ALE-CM, and generalization to Terminal-Bench (TB) 3.0-CM, 4.0-CM, and 2.1 on the left and a summary of our latent-world-first data synthesis approach on the right. Tev showcases significant performance gains across two model scales and Tev -27B matches GPT-6 Astra and Claude Opus 5 at Pass@5 on ALE-CM.
Figure 1: We demonstrate Tev’s results on ALE-CM, and generalization to Terminal-Bench (TB) 3.0-CM, 4.0-CM, and 2.1 on the left and a summary of our latent-world-first data synthesis approach on the right. Tev showcases significant performance gains across two model scales and Tev-27B matches GPT-6 Astra and Claude Opus 5 at Pass@5 on ALE-CM.

At a glance

Tev-Gym contains 1,170 verified task environments across seven computing-and-math domains. Each starts from a hidden correct world, renders and corrupts its public artifacts, and derives both completion checks and intermediate progress from the same specification.

Progress-potential reward shaping (PPRS) supplies learning signal when every rollout fails. Tev-27B improves ALE-CM best-of-five full success from 0.17 to 0.28 and transfers to selected Terminal-Bench tasks; a 320B model also improves. Results concern synthetic terminal workflows and small selected long-horizon evaluation sets, with different frontier harnesses and a remaining frontier-performance gap.

1 Introduction

Long workflows are difficult to generate consistently and yield sparse completion rewards. In group-relative RL, identical failing rewards provide no gradient, precisely when a weak policy needs feedback. Tev instead builds the task, public files, verifier, and progress scorer from one known-correct world.

The dataset spans software, data engineering, AI research, mathematics, security, infrastructure, and quantum computing. Seeded generators offer adjustable scale and dependency structure. Under a fixed evaluator, tasks take a median 12.4 minutes versus 0.5–2.0 minutes for comparison corpora; this is measured agent time, not human effort.

Train 9B on 264 unit-tier tasks, 27B on all 1,170, and apply the recipe to a 320B base. Model size and training coverage therefore change together; their differences do not isolate parameter scaling.

Figure 2: All models on the Terminal-Bench 4.0 “CM” subset (TB 4.0-CM), ranked by Pass@1. Tev -320B and Tev -27B are trained with our recipe from GLM-5.3-Flash and Qwen3.8-27B, respectively; each arrow gives the relative gain over its base.
Figure 2: All models on the Terminal-Bench 4.0 “CM” subset (TB 4.0-CM), ranked by Pass@1. Tev-320B and Tev-27B are trained with our recipe from GLM-5.3-Flash and Qwen3.8-27B, respectively; each arrow gives the relative gain over its base.
Table 1: Tev-Gym versus prior agentic task corpora. “Agent time” is the median and mean wall-clock minutes per rollout of Gemini 3.6 Flash, from first action to last step (details in Appendix A.8). Difficulty is Pass@k measured by Gemini 3.6 Flash across all datasets.
Dataset Size # Domains Agent time (min) Oracle Grading Progressive P@1 P@8
Median Mean Metric Reward
SWE-Smith (Yang et al., 2025a) 50.1k 1 2.0 2.4 unit tests pass/fail ✗ 47% 52%
Terminal-Universe (Wu et al., 2026) 32.0k 3 NA NA pytest pass/fail ✗ NA NA
TMAX-15K (Ivison et al., 2026) 14.6k 3 1.3 2.1 unit tests pass/fail ✗ 74% 80%
SETA (Shen et al., 2026) 4.5k 4 1.2 2.1 unit tests test fraction ✗ 80% 88%
Endless Terminals (Gandhi et al., 2026) 3.3k 2 0.5 0.7 unit tests pass/fail ✗ 83% 89%
CLI-Gym (Lin et al., 2026) 1.7k 4 1.9 2.3 unit tests pass/fail ✗ 69% 74%
Tev-Gym (ours) 1,170 7 12.4 14.8 latent world staged Φ\Phi ++ pass/fail ✓ 29% 40%

2 Long-Horizon Environment Synthesis

Construct environments backward from a correct world, then enforce coverage and qualification before admitting them to training.

Figure 3: Latent-world-first synthesis, on one example Tev-Gym environment. This environment’s task is “Retail Warehouse ETL Recovery. Goal: build a clean star-schema SQLite warehouse from six months of messy retail data and produce two truthful JSON sidecars.”
Figure 3: Latent-world-first synthesis, on one example Tev-Gym environment. This environment’s task is “Retail Warehouse ETL Recovery. Goal: build a clean star-schema SQLite warehouse from six months of messy retail data and produce two truthful JSON sidecars.”

2.1 Latent-World-First Task Synthesis

A seeded generator first draws a valid latent world and an explicit defect plan. Code compiles its workflow into a DAG with dependencies and work estimates. Rendering produces clean heterogeneous files; round-trip checks establish that they encode the world correctly.

Declared, recoverable corruptions create the agent’s starting workspace. The prompt specifies deliverables, schemas, precedence rules, and tools. Hidden verifiers and progress scorers compare outputs with the same untouched world. Sealed metamorphic tests and stale-output counterfactuals check invariance and reject plausible wrong states.

Three worked examples extend the construction to a visual cloud-cost task, building control, and financial-statement reconstruction. In the simulator example, a pinned tool procedure supplies the reference rather than a generated record set alone.

Equation 1
E=(x,s0,verW,ΦW),E\;=\;\bigl(x,\;s_{0},\;\operatorname{ver}_{W},\;\Phi_{W}\bigr), (1)

2.2 Coverage, Calibration, and Qualification

Coverage is checked over 19 axes, including domain, scale, workflow, artifacts, and corruption type. World scale controls operational work, while dependency depth and corruption classes control required stages.

Every environment must pass eight qualification gates. Progress must respond to correct work and regressions; 14 specified gaming probes must score zero; solvers, verifier, and scorer have disjoint code closures. These checks provide concrete evidence against tested exploits, not a proof against every possible shortcut.

Figure 4: The progress potential along one group ( G = 8 G{=}8 ) of rollouts. Eight rollouts of the trained
27B model on one warehouse environment, graded live after every turn. Under terminal-only reward (red) every rollout scores 0 0 . Under progressive rewards, by contrast, two submit an
incomplete warehouse (circles), five exhaust the 65k-token context (crosses), and one builds a
complete warehouse, so the rewards vary within the group.
Figure 4: The progress potential along one group (G=8) of rollouts. Eight rollouts of the trained 27B model on one warehouse environment, graded live after every turn. Under terminal-only reward (red) every rollout scores 0. Under progressive rewards, by contrast, two submit an incomplete warehouse (circles), five exhaust the 65k-token context (crosses), and one builds a complete warehouse, so the rewards vary within the group.

3 Dense Reward for Long-Horizon RL

The same hidden specification grades completion and intermediate work, enabling progress feedback without a learned critic or language-model judge.

3.1 Terminal Reward Starves Group-Relative RL

GRPO normalizes rewards within a task’s rollout group [46]. Under binary completion reward, all-fail groups have zero advantage. Their probability is (1−p)G: with eight independent rollouts it is 66% at 5% success and 92% at 1%. Even one success leaves every failure equally negative, obscuring near-misses.

Equation 2
A^i=Ri−mean(R1:G)std(R1:G)+ϵ.\hat{A}_{i}\;=\;\frac{R_{i}-\operatorname{mean}(R_{1:G})}{\operatorname{std}(R_{1:G})+\epsilon}. (2)

3.2 A Progress Potential from the Latent World

A progress contract assigns weighted checkable nodes to the workflow. Each node’s partial score lies in [0,1], and its credited value is capped by the weakest direct or indirect prerequisite. The potential Φ sums these gated scores, preventing a correct-looking downstream artifact from bypassing incorrect inputs.

The untouched tree and tested exploits score zero; complete accepted workspaces score one. Independent scorer code checks the latent world directly. Correct intermediate states can be generated for each seed, avoiding human step labels and learned reward models.

Equation 3
Φ⁡(s)=∑v∈Vwv​c~v​(s)∈[0,1].\Phi(s)\;=\;\sum_{v\in V}w_{v}\,\tilde{c}_{v}(s)\;\in\;[0,1]. (3)

3.3 Progress-Potential Reward Shaping (PPRS)

Each turn earns the change in progress potential. With λ=0.25 and initial potential zero, the total shaped return telescopes to the submission-success indicator plus λ times final progress. A submitted passing rollout earns 1.25; a failing or unsubmitted rollout earns at most 0.25. A complete workspace that is never submitted therefore receives only 0.25.

PPRS deliberately retains terminal potential and changes the objective, unlike policy-invariant potential shaping [40]. Successful trajectories still outrank failures, but rankings of stochastic policies can reverse. Different failing progress levels supply reward variance; groups stuck at identical progress remain uninformative.

Equation 4
R⁡(τ)= 1​[τ​passes]+λ​∑t=1Trt= 1​[τ​passes]+λ​Φ​(sT),R(\tau)\;=\;\mathds{1}[\tau\ \text{passes}]\;+\;\lambda\sum_{t=1}^{T}r_{t}\;=\;\mathds{1}[\tau\ \text{passes}]\;+\;\lambda\,\Phi(s_{T}), (4)
Table 2: Main results on ALE-CM, computing and math science domains of Agents’ Last Exam (ALE). Each cell reports the mean graded score (S@1) and the fraction of tasks fully solved (P@1), averaged over 5 rollouts per task, together with their best-of-5 counterparts (S@5, P@5).
Model Harness Math & Science ML Coding & System All (Avg)
S@1 S@5 P@1 P@5 S@1 S@5 P@1 P@5 S@1 S@5 P@1 P@5 S@1 S@5 P@1 P@5
Frontier API models
GPT-6 Astra Codex 0.58 0.58 0.38 0.38 0.48 0.49 0.20 0.20 0.81 0.85 0.20 0.20 0.62 0.63 0.28 0.28
Claude Fable 5 Claude Code 0.47 0.51 0.38 0.38 0.52 0.57 0.24 0.40 0.83 0.90 0.20 0.20 0.58 0.63 0.29 0.33
GPT-5.6 (sol) Codex 0.50 0.51 0.38 0.38 0.46 0.49 0.20 0.20 0.82 0.85 0.20 0.20 0.58 0.60 0.28 0.28
GPT-5.6 (terra) Codex 0.51 0.56 0.38 0.38 0.25 0.26 0.00 0.00 0.75 0.83 0.20 0.20 0.50 0.55 0.22 0.22
Claude Opus 5 Claude Code 0.51 0.53 0.38 0.38 0.43 0.46 0.20 0.20 0.85 0.87 0.20 0.20 0.58 0.60 0.28 0.28
Claude Sonnet 5 Claude Code 0.50 0.51 0.38 0.38 0.39 0.56 0.12 0.40 0.80 0.87 0.20 0.20 0.55 0.62 0.26 0.33
Claude Opus 4.8 Claude Code 0.48 0.51 0.38 0.38 0.31 0.33 0.00 0.00 0.80 0.86 0.20 0.20 0.52 0.55 0.22 0.22
Gemini 3.7 Flash Gemini CLI 0.48 0.50 0.38 0.38 0.26 0.29 0.00 0.00 0.84 0.86 0.20 0.20 0.52 0.54 0.22 0.22
Gemini 3.6 Flash Gemini CLI 0.48 0.51 0.35 0.38 0.25 0.26 0.00 0.00 0.81 0.86 0.20 0.20 0.51 0.54 0.21 0.22
Gemini 3.1 Pro Gemini CLI 0.49 0.52 0.35 0.38 0.27 0.29 0.00 0.00 0.82 0.88 0.20 0.20 0.52 0.56 0.21 0.22
Open-source models
Qwen3.8-2.4T-A95B Vanillux2 0.20 0.38 0.20 0.38 0.10 0.20 0.04 0.20 0.74 0.86 0.16 0.20 0.32 0.46 0.14 0.28
DeepSeek-V4-pro Vanillux2 0.12 0.25 0.12 0.25 0.10 0.20 0.04 0.20 0.73 0.88 0.16 0.20 0.29 0.41 0.11 0.22
Qwen3.5-9B (base) Vanillux2 0.07 0.12 0.00 0.00 0.19 0.26 0.00 0.00 0.59 0.73 0.20 0.20 0.25 0.33 0.06 0.06
Tev-9B (ours) Vanillux2 0.14 0.29 0.04 0.12 0.23 0.26 0.00 0.00 0.56 0.73 0.20 0.20 0.29 0.40 0.07 0.11
ΔTev-9B - Qwen3.5-9B\Delta_{\text{{Tev}-9B - Qwen3.5-9B }} −- 0.07↑\uparrow 0.17↑\uparrow 0.04↑\uparrow 0.12↑\uparrow 0.04↑\uparrow −- −- −- 0.03↓\downarrow −- −- −- 0.04↑\uparrow 0.07↑\uparrow 0.01↑\uparrow 0.05↑\uparrow
Qwen3.8-27B (base) Vanillux2 0.15 0.25 0.15 0.25 0.21 0.32 0.00 0.00 0.74 0.82 0.20 0.20 0.33 0.43 0.12 0.17
Tev-27B (ours) Vanillux2 0.23 0.50 0.23 0.50 0.26 0.29 0.00 0.00 0.82 0.86 0.20 0.20 0.40 0.54 0.16 0.28
ΔTev-27B - Qwen3.8-27B\Delta_{\text{{Tev}-27B - Qwen3.8-27B }} −- 0.08↑\uparrow 0.25↑\uparrow 0.08↑\uparrow 0.25↑\uparrow 0.05↑\uparrow 0.03↓\downarrow −- −- 0.08↑\uparrow 0.04↑\uparrow −- −- 0.07↑\uparrow 0.11↑\uparrow 0.04↑\uparrow 0.11↑\uparrow

4 Experiments

Experiments test in-domain improvement, transfer to longer tasks, retention on a middle-horizon benchmark, and separate data/reward ablations.

4.1 Setup

Training uses DPPO, 24 rollouts for each of eight tasks per step, 65,536 response tokens per rollout, 16,384 per turn, and at most 64 tool calls through one bash tool [43]. The 27B run takes about four days on four B200 nodes or eight A100 nodes.

ALE-CM contains 18 executable computing-and-math tasks. Terminal-Bench 3.0/4.0 subsets contain 22/18 domain-related tasks runnable in the CPU-only harness; Claude assisted domain selection. High correlations with full-benchmark rankings support relevance but do not make these full-suite results.

No training prompt shares an 8-gram or 13-gram with the measured evaluation prompts. This lexical test and fresh generated artifacts do not exclude conceptual overlap or contamination in pretrained models.

4.2 Main Results

On ALE-CM, Tev-27B raises mean graded score S@1 from 0.33 to 0.40, full-success P@1 from 0.12 to 0.16, and P@5 from 0.17 to 0.28. The last matches the reported GPT-6 Astra and Claude Opus 5 value under their native harnesses; it does not establish general parity.

On the 22/18-task Terminal-Bench subsets, 27B P@1 rises 0.09→0.14 and 0.04→0.10. Tev-320B improves 0.20→0.27 and 0.29→0.37; its P@5 rises only 0.05 on 3.0 and stays unchanged on 4.0, suggesting improved reliability more than expanded task reach.

On all 89 Terminal-Bench 2.1 tasks, Tev-27B reaches 0.794 P@1 and 0.884 P@5, versus 0.764/0.864 for its base. The Kubernetes case illustrates better evidence selection and rollback judgment, but it is a selected single-rollout comparison.

Table 3: Terminal-Bench (TB) 3.0 and 4.0 “CM” subsets. Leaderboard rows are from the official per-trial results (5 trials per task); ∗results evaluated by us, not on the official leaderboard.
Model TB 3.0 (22 tasks)    TB 4.0 (18 tasks)
Software ML Operations Security All    Software ML Operations Security All
P@1 P@5 P@1 P@5 P@1 P@5 P@1 P@5 P@1 P@5    P@1 P@5 P@1 P@5 P@1 P@5 P@1 P@5 P@1 P@5
Claude Fable 5 0.50 0.80 0.45 0.50 0.15 0.25 0.40 0.75 0.41 0.64    0.53 0.62 0.60 0.75 0.05 0.25 0.70 1.00 0.46 0.61
Claude Opus 5 0.54 0.60 0.60 0.75 0.40 0.50 0.55 0.75 0.53 0.64    0.47 0.62 0.75 1.00 0.20 0.50 0.90 1.00 0.52 0.72
GPT-5.6 (sol) 0.32 0.50 0.65 0.75 0.10 0.25 0.70 0.75 0.41 0.55    0.33 0.62 0.70 1.00 0.00 0.00 0.80 1.00 0.39 0.61
GPT-5.6 (terra) 0.22 0.50 0.55 0.75 0.05 0.25 0.50 0.75 0.30 0.55    0.10 0.38 0.50 0.75 0.00 0.00 0.70 1.00 0.23 0.44
GPT-5.6 (luna) 0.18 0.30 0.25 0.50 0.00 0.00 0.30 0.50 0.18 0.32    0.12 0.25 0.25 0.50 0.05 0.25 0.50 0.50 0.18 0.33
Grok 4.6 0.24 0.50 0.50 0.75 0.00 0.00 0.55 0.75 0.30 0.50    0.30 0.50 0.55 0.75 0.00 0.00 0.50 1.00 0.31 0.50
Claude Sonnet 5 0.18 0.40 0.25 0.50 0.00 0.00 0.20 0.25 0.16 0.32    0.28 0.62 0.20 0.75 0.00 0.00 0.20 0.50 0.19 0.50
Claude Opus 4.8 0.20 0.70 0.10 0.25 0.05 0.25 0.40 0.50 0.19 0.50    0.20 0.50 0.35 0.50 0.05 0.25 0.80 1.00 0.27 0.50
Grok 4.5 0.14 0.30 0.45 1.00 0.15 0.50 0.50 0.75 0.26 0.55    0.12 0.12 0.35 0.75 0.00 0.00 0.60 1.00 0.20 0.33
GLM 5.2 0.04 0.20 0.05 0.25 0.00 0.00 0.05 0.25 0.04 0.18    0.06∗ 0.24∗ 0.02∗ 0.10∗ 0.00∗ 0.00∗ 0.00∗ 0.00∗ 0.03∗ 0.13∗
Gemini 3.7 Flash∗ 0.08 0.20 0.00 0.00 0.00 0.00 0.20 0.50 0.07 0.18    0.12 0.25 0.00 0.00 0.00 0.00 0.30 0.50 0.09 0.17
Gemini 3.6 Flash∗ 0.04 0.20 0.00 0.00 0.00 0.00 0.00 0.00 0.02 0.09    0.20 0.25 0.00 0.00 0.00 0.00 0.00 0.00 0.09 0.11
Qwen3.8-27B (base) 0.17 0.43 0.00 0.00 0.00 0.00 0.05 0.25 0.09 0.24    0.07 0.31 0.05 0.25 0.00 0.00 0.00 0.00 0.04 0.19
Tev-27B 0.15 0.39 0.19 0.25 0.00 0.00 0.20 0.25 0.14 0.27    0.17 0.33 0.12 0.46 0.00 0.00 0.00 0.00 0.10 0.25
ΔTev-27B - Qwen3.8-27B\Delta_{\text{{Tev}-27B - Qwen3.8-27B }} 0.02↓\downarrow 0.04↓\downarrow 0.19↑\uparrow 0.25↑\uparrow −- −- 0.15↑\uparrow −- 0.05↑\uparrow 0.03↑\uparrow    0.10↑\uparrow 0.02↑\uparrow 0.07↑\uparrow 0.21↑\uparrow −- −- −- −- 0.06↑\uparrow 0.06↑\uparrow
GLM-5.3-Flash (base) 0.18 0.60 0.35 0.50 0.05 0.25 0.22 0.50 0.20 0.50    0.16 0.50 0.60 0.75 0.05 0.25 0.68 1.00 0.29 0.56
Tev-320B 0.21 0.70 0.40 0.50 0.10 0.25 0.45 0.50 0.27 0.55    0.25 0.50 0.65 0.75 0.15 0.25 0.70 1.00 0.37 0.56
ΔTev-320B - GLM-5.3-Flash\Delta_{\text{{Tev}-320B - GLM-5.3-Flash }} 0.03↑\uparrow 0.10↑\uparrow 0.05↑\uparrow −- 0.05↑\uparrow −- 0.23↑\uparrow −- 0.07↑\uparrow 0.05↑\uparrow    0.09↑\uparrow −- 0.05↑\uparrow −- 0.10↑\uparrow −- 0.02↑\uparrow −- 0.08↑\uparrow −-
Table 4: Terminal-Bench (TB) 2.1 results. Frontier-reference rows are taken from the official TB 2.1 leaderboard (5 trials, 89 tasks).
Model Pass@1 Pass@5
GPT-6 Astra (Codex, high) 0.874 0.944
Fable 5 (Claude Code, xhigh) 0.838 0.933
Opus 4.8 (Claude Code, high) 0.789 0.944
Sonnet 5 (Claude Code, high) 0.746 0.921
Qwen3.5-9B (base) 0.178 0.292
Tev-9B (ours) 0.227 0.348
ΔTev-9B - Qwen3.5-9B\Delta_{\text{{Tev}-9B - Qwen3.5-9B }} 0.049↑\uparrow 0.056↑\uparrow
Qwen3.8-27B (base) 0.764 0.864
Tev-27B (ours) 0.794 0.884
ΔTev-27B - Qwen3.8-27B\Delta_{\text{{Tev}-27B - Qwen3.8-27B }} 0.030↑\uparrow 0.020↑\uparrow

4.3 Ablation Study

At 9B, prompt-first generation and a public-view verifier fail to improve mean graded score over the base. The latter discards 88% of sampled groups. Both use terminal rewards and should be compared with the sparse-reward arm when isolating data construction.

Keeping Tev-Gym fixed, PPRS uses undiscounted return-to-go and compares advantages at the same turn index. It improves S@1 by 0.04 over sparse reward and reduces GPU-hours per step from 16.7 to 9.0. Pooling turns weakens late credit; additive current-potential rewards encourage lingering and context exhaustion.

At 27B, the reported terminal-only run could not complete one optimizer step because nearly all groups had zero advantage. This is a result under the stated recipe and data, not proof that every sparse-reward training method must fail.

Table 5: Ablations on ALE-CM (Qwen3.5-9B). Here GPU-h/step is the mean time of one training step times the 32 GPUs of the recipe in H100 GPU-hours.
Method Data Reward GPU-h / step ALE-CM
World DAG Hidden truth Signal Credit Adv. norm. S@1 S@5 P@1 P@5
Qwen3.5-9B (base) – – – no RL – – – 0.25 0.33 0.06 0.06
Tev-9B (ours) ✓ ✓ ✓ Δ​Φt\Delta\Phi_{t} + terminal GtG_{t}, γ=1\gamma{=}1 per turn 9.0 0.29 0.40 0.07 0.11
Data ablation: replace one construction step with an LLM; terminal reward, compare with Sparse
Public-view verifier ✓ ✓ LLM terminal whole rollout per rollout 22.3 0.24 0.35 0.04 0.06
Prompt-first LLM LLM LLM terminal whole rollout per rollout 11.8 0.23 0.32 0.06 0.06
Reward ablation: keep Tev-Gym, change how Φ\Phi enters training
Sparse (terminal only) ✓ ✓ ✓ terminal whole rollout per rollout 16.7 0.25 0.32 0.06 0.06
PPRS, pooled norm. ✓ ✓ ✓ Δ​Φt\Delta\Phi_{t} + terminal GtG_{t}, γ=1\gamma{=}1 per rollout 10.9 0.28 0.35 0.06 0.06
Direct reward (additive) ✓ ✓ ✓ Φ⁡(st)\Phi(s_{t}) + terminal GtG_{t}, γ=0\gamma{=}0 per turn 9.7 0.29 0.35 0.06 0.06
Figure 5: Qualitative example on the ALE-CM task: k8s payment api root cause analysis (Coding & System). From a kubectl capture, the Deployment/HPA YAML and a crashing pod’s log, the agent writes a JSON root-cause report with verbatim evidence strings and a rollback verdict. Green marks a decision that earns a component, red one that forfeits it.
Figure 5: Qualitative example on the ALE-CM task: k8s payment api root cause analysis (Coding & System). From a kubectl capture, the Deployment/HPA YAML and a crashing pod’s log, the agent writes a JSON root-cause report with verbatim evidence strings and a rollback verdict. Green marks a decision that earns a component, red one that forfeits it.

5 Related Work

Prior coding corpora mine or synthesize tasks, while environment generators scale executable tool worlds. Tev’s distinction is a common latent-world source for multi-stage artifacts, exact grading, and intermediate progress. Backward construction and reward shaping are established ideas; the contribution combines them for this setting.

6 Conclusion

Verified world construction and progress feedback improve the tested open models on demanding terminal tasks. Scope remains synthetic workflows, and the hardest benchmarks retain a substantial frontier gap. New workflows require additional generator components and validation; broader coverage and simpler authoring remain future work.

Acknowledgments

The authors acknowledge Adobe Research’s support and feedback from colleagues at Adobe and UCLA.

AI use statement

Coding agents helped author generators, reference solvers, verifiers, progress scorers, and analysis code; authors reviewed the specifications. Claude assisted benchmark subset selection. AI also supported literature review and writing. Once authored, generators and reward evaluation run without model calls; no generative model grades training tasks.

Ethics statement

Training artifacts are synthetic and contain no collected personal records or proprietary source code. Security-themed tasks analyze generated artifacts in declared sandboxes. The authors acknowledge dual-use implications of stronger coding agents and report overlap checks and native-harness frontier comparisons.

Reproducibility statement

Generators, seeds, and scale tiers define reproducible environments, with reported median generation cost 30 CPU seconds. The appendix specifies qualification gates, reward equations, training settings, task lists, evaluation budgets, and three worked examples. The example listings are excerpts of package files, preserved below in full as published.

Appendix A Data Construction Details

A.1 Task Sources and Domain Taxonomy

Seven domains follow the ALE computing-and-math taxonomy [52]. Specifications come from public tasks, competition material, and quantum-computing references; new generators borrow structure and difficulty rather than copying artifacts.

Feasibility filtering favors knowable ground truth and batch/CPU execution. Near-duplicate checks and adversarial specification review precede implementation. At least two generators populate each domain; the scope excludes many live-service and heavy-toolchain workflows.

Table 6: The seven computing-and-math subdomains of ALE-CM that Tev-Gym spans. Subdomains (9.1–9.7) of the Computing & Math domain of Agents’ Last Exam (Sun et al., 2026), which we refer to as Tev-Gym’s seven domains. Ids and names are taken from the ALE taxonomy; scope summaries follow it, reconstructed from O*NET/SOC 2018 and the released ALE task specifications. “Verified by” names the deterministic signals we build the strict verifier and progress potential around.
Subdomain Scope
9.1 Software Engineering Feature implementation, debugging, refactoring, and test authoring across real codebases, turning issues and specs into working, tested code changes. Verified by: test-suite pass (gate then score), behavioral correctness under fixed inputs, and no regression in existing tests.
9.2 Data & Analytics Engineering ETL/ELT pipelines, recommender and ranking systems, feature engineering, and analytics workflows, turning raw data into validated datasets, features, and model outputs. Verified by: output values exact or within tolerance, schema validity (gate), metric reproduction, and feature-parity against a reference.
9.3 AI Engineering, Safety & CS Research Model training and evaluation pipelines, research-paper reproduction, checkpoint and config management, and alignment/safety experiments, turning specs and code into reproduced results. Verified by: reproduced metrics within tolerance (FID, accuracy, etc.), config and alignment exactness, artifact validity, and deterministic-under-seed behavior.
9.4 Mathematical & Operations Research Combinatorial optimization, game-theoretic equilibrium computation, applied mathematics and number theory, and probabilistic estimation and filtering, implemented from specifications. Verified by: objective or solution exact or within tolerance, an optimality/feasibility certificate (gate), and reproducibility under a fixed seed.
9.5 Cybersecurity & Digital Forensics Reverse engineering and malware-config extraction, network and packet triage, crackme and key-generation analysis, and incident forensics, turning binaries and captures into extracted artifacts and findings. Verified by: extracted-value exactness against a reference (hash/string), finding-set precision and recall, and decoded-artifact correctness (gate).
9.6 Infrastructure Engineering & Cloud Operations Kubernetes migration and configuration, root-cause analysis of production incidents, OS and security hardening, and infrastructure-as-code, turning infrastructure specs and incidents into working, secured systems. Verified by: the system’s world-state (resources/health) under fixed inputs, config correctness against a reference, incident-cause exactness, and policy-check pass (gate).
9.7 Quantum Computing Quantum-circuit construction and simulation, quantum-error-correction code analysis, and circuit-level noise simulation, turning circuit and code specs into measured outcomes and logical-error estimates. Verified by: statistics and error-rates within tolerance against a reference, reproducibility under a fixed seed, and circuit/structure validity.

A.2 Determinism, Reproducibility, and Generation Cost

Coding-agent authoring and human review are one-time costs per generator. Instantiation then uses deterministic seeded code, with median/90th-percentile generation costs of 30/82 CPU seconds including solvers, hidden-suite creation, and progress preview. These marginal costs exclude generator authoring.

A.3 Coverage Axes and Covering Array

The corpus enforces a strength-two covering array: every valid pair of values across 19 coverage axes appears in an admitted environment. Greedy selection fills uncovered pairs, reducing fixed template combinations. Files span databases, model shards, binaries, manifests, and some images. Reseeding changes entities, values, layout, and corruption placement.

A.4 Scale Ladder

Scale grows world size and processing burden without necessarily deepening the workflow DAG. For generators at all three tiers, median inputs grow from 17 files/51 KB to 20/168 KB and 28/586 KB. Checkpoint tasks also vary shard count and tensor/pipeline/expert parallelism topology.

A.5 Measuring Horizon and Difficulty

Structural horizon is the workflow’s critical path, not an agent’s potentially wasteful rollout length. Difficulty is verified success under a specified model and harness. When success saturates near zero, final potential distinguishes degrees of partial progress.

A.6 Qualification Gates

The eight gates require reward causality; zero score on 14 exploit probes; disjoint code closures; agreement between independent reference solvers; valid increasing progress with negative regression increments; metamorphic invariance and rejection of stale outputs; distinct worlds across six seeds; and a public contract without private grading data.

Solver agreement and independence reduce shared implementation failure, but the finite checks cannot establish universal specification correctness or exploit resistance. All published gate definitions remain in Table 7.

Table 7: Intrinsic qualification gates. Every task environment must pass all gates; each gate is benchmark-independent, enforced while each generator is authored, and checked automatically when each environment is generated. The third column names the failure class each gate is matched to; passing one gate is not taken as evidence for a property it does not observe.
Gate What it enforces Failure class it targets
Reward causality each reward component drops when its cause is mutated reward credits the wrong outcome
Exploit resistance all 14 gaming probes score 00 verifier loophole, shortcut acceptance
Role independence disjoint reference/alt/verifier/progress closures shared bias between solver and grader
Solver agreement reference and alternative solvers yield identical trees verifier-only acceptance, unresolved spec.
Progress validity 0<Φ<10<\Phi<1; monotone; decreases on regression invalid process signal
Metamorphic invariance meaning-preserving rewrites keep the grade; stale outputs fail over-narrow verifier; stale-state acceptance
Six-seed diversity distinct world / public-tree / task id (≥6\geq 6) degenerate environment, template repetition
Public-contract hygiene no private authority in the prompt or public tree answer leakage

A.7 Dataset Statistics

Normalized entropy balance is 0.89 across three capability categories and 0.91 across seven domains. Median environments have six workflow stages, eight progress nodes, six reward components, two corruptions, and 12 input files. Formats total 23.

One untrained 27B probe per environment fully solves 3.2%, with mean graded score 0.06; 52% exhaust the token budget before submission. Gemini 3.6 Flash reaches 29% pass@1 and 40% pass@8 on a fixed 92-environment sample. These model/harness/sample conditions differ and should accompany difficulty comparisons.

Table 8: Composition of Tev-Gym by capability category. Counts are task environments; horizon columns are per-environment medians over the category (workflow stages, input files the agent sees, injected corruptions); “formats” counts distinct wire formats among the inputs. Base P@1 is the single-rollout pass rate of untrained Qwen3.8-27B under the training harness, from one probe on our environments; frontier P@1 is the fixed evaluator of Appendix A.8 on its 92-environment subsample.
Per-environment horizon (median) Difficulty (P@1)
Category Task envs. Stages Input files Corruptions Formats Base 27B Frontier
Math & Science 264 6 8 1 7 0.04 0.31
ML 242 6 12 1 11 0.03 0.24
Coding & System 664 6 14 3 19 0.03 0.30
Tev-Gym (total) 1,170 6 12 2 15 0.03 0.29

A.8 Comparison Protocol

One fixed evaluator measures corpus difficulty at 60 tool calls and eight rollouts per task. Prior corpora use the official Terminus-2 scaffold; Tev uses its single-bash harness. Tasks that cannot faithfully build are excluded and recorded, which can bias a corpus estimate upward.

Timing uses 100 sampled Tev environments with two rollouts each and expanded 300-call/two-hour limits; neither binds. Wall time includes model calls, retries, and commands, yielding 12.4-minute median and 14.8-minute mean. Prior-corpus timing uses its difficulty runs. The comparison shares a model but not every scaffold or timing budget, and is not an estimate of human work.

Table 9 retains the generation, qualification, coverage, and reward comparison. A progress signal must grade intermediate workspace states; an end-only test fraction does not qualify.

Table 9: How recent agentic corpora are generated and gated. Axes follow the ACE lens (Zeng et al., 2026a). Anchor is the factor from which the others are derived; model in generation lists what a language model writes; validation is what admits a task and when; difficulty and coverage record whether the corpus controls them (“–” marks no explicit control reported); progress signal is what a learner receives short of a terminal pass. Prior-corpus cells are transcribed from the cited papers.
Corpus Anchor Model in generation Validation Difficulty control Coverage control Progress signal
SWE-Smith (Yang et al., 2025a) repository bug injection, task text tests run post hoc – – none
Terminal-Universe (Wu et al., 2026) agent trajectory trajectory, task text pytest post hoc – – none
TMAX-15K (Ivison et al., 2026) task text task, container, tests none; RL soft-filters – topic balancing none
SETA (Shen et al., 2026) task text task, tests solvability post hoc – – test fraction
Endless Terminals (Gandhi et al., 2026) task text task, container, tests solvability post hoc – – none
CLI-Gym (Lin et al., 2026) agent trajectory trajectory, task text tests post hoc – – none
Tev-Gym (ours) latent world generator code (once, reviewed) 8 gates at generation scale ×\times DAG depth strength-2 covering array staged Φ\Phi

A.9 Decontamination

Fresh seeds and public-tree hashes establish distinct generated artifacts. Across all 1,170 prompts, no measured evaluation prompt shares an 8-gram or 13-gram. A control detects 156 shared 13-grams between the Terminal-Bench sets, demonstrating that the comparator can find overlap.

Median nearest-neighbor token Jaccard is 0.13 for Terminal-Bench 3.0/4.0, 0.10 for 2.1, and 0.22 for ALE-CM. These lexical and artifact checks support lack of verbatim duplication; they do not test all semantic similarity or pretrained-model exposure.

Table 10: Lexical overlap between all 1,170 environments of Tev-Gym and each evaluation set. Hits: evaluation tasks sharing at least one 13-gram (resp. 8-gram) of lower-cased alphanumeric tokens with any environment prompt. Max token cov.: for the worst evaluation task, the share of its tokens covered by 8-grams that also occur in some environment prompt. NN Jaccard: per evaluation task, the largest token-set Jaccard similarity to any environment prompt; median / max over tasks. CM = the 22- / 18-task computing-and-math subsets evaluated in Table 3. The calibration block applies the same code to same-benchmark, different-task pairs.
Evaluation set Tasks 13-gram hits 8-gram hits Max token cov. NN Jaccard med. / max
ALE-CM 18 0 0 0.0% 0.22 / 0.38
TB 3.0, all 74 0 0 0.0% 0.13 / 0.18
TB 3.0, CM subset 22 0 0 0.0% 0.14 / 0.18
TB 4.0, all 66 0 0 0.0% 0.13 / 0.18
TB 4.0, CM subset 18 0 0 0.0% 0.14 / 0.18
TB 2.1 89 0 0 0.0% 0.10 / 0.16
Calibration: same code, same-benchmark tasks
ALE-CM task vs. other ALE-CM tasks 18 2 13 26.0% 0.28 / 0.34
TB 3.0 CM task vs. other TB 3.0 tasks 22 22 22 42.5% 0.24 / 0.37
TB 4.0 CM task vs. other TB 4.0 tasks 18 18 18 42.5% 0.25 / 0.37

A.10 Three Worked Training Environments

The three complete published walkthroughs separate public prompts and inputs from sealed ground truth. Cloud cost optimization demonstrates image evidence; building control and finance illustrate domains outside the seven-domain training focus. All examples run in pinned, network-disabled containers. Record-based examples derive answers directly, while building control executes a pinned simulator procedure.

Example 1: multimodal cloud cost optimization (Cloud Operations)

The cloud-cost task combines 430 resource records, pricing, and a dashboard image. Three actionable findings occur only in the image, so tabular processing alone misses them. Four explicit rules produce 148 recommendations: 74 compute downsizes, 27 database downsizes, 44 storage-tier changes, and three dashboard findings.

Reference savings are 16,395.02 USD monthly. Costs and savings are independently rounded from exact prices, so summed savings differ by 0.08 from rounded current-minus-projected totals. The verifier applies parsing/sanity gates, then scores recommendation coverage (30%), actions (30%), savings accuracy (25%), and completeness (15%). All prompts, input excerpts, reference outputs, and the dashboard follow.

# AWS cost optimization analysis (430 inventory resources)

You are an AWS cost optimization analyst working on Linux. Three inputs are staged under `input/`:

- `input/aws_billing_dashboard.png` - a synthetic AWS billing dashboard screenshot. It shows a per-service cost breakdown AND three 'dashboard-only' findings (an idle NAT gateway, an unattached Elastic IP, and a stale CloudWatch log group) that are NOT present in the usage CSV. Inspect the image directly to recover them.
- `input/resource_usage.csv` - the EC2/RDS/S3 inventory with columns `resource_id,resource_type,name,state,instance_type,cpu_p95,mem_p95,connections_p95,size_gb,ia_access_ratio`.
- `input/aws_pricing_reference.csv` - the deterministic pricing table (`sku,dimension,unit,price_usd`).

Apply these rules and price every saving from the pricing table:
- Downsize a running EC2 instance one tier when `cpu_p95 < 20` and `mem_p95 < 50`.
- Downsize an available RDS instance one tier when `cpu_p95 < 25` and `connections_p95 < 10`.
- Move an S3 bucket to Infrequent Access when `ia_access_ratio < 0.10` and `size_gb >= 100`.
- Include the three dashboard-only findings (delete the idle NAT gateway, release the unattached Elastic IP, delete the stale CloudWatch log group).

Write `output/optimization_report.json` with an `executive_summary` object and a `recommendations` array, and `output/savings_summary.csv` with columns `resource_id,resource_type,name,action,current_cost,projected_cost,monthly_savings`.

Scoring is weighted: 30% coverage F1 over the hidden recommendation set (including the three dashboard-only resources), 30% action correctness, 25% savings accuracy (full credit within +/-20% relative error), and 15% report completeness.
input/
  aws_billing_dashboard.png  900 x 560 px dashboard render: cost by service + 3 findings
  resource_usage.csv         430 rows: 220 EC2 instances, 80 RDS instances, 130 S3 buckets
  aws_pricing_reference.csv  11 SKUs: 5 EC2 tiers, 4 RDS tiers, 2 S3 storage classes
Figure 6: The dashboard screenshot staged as input/aws_billing_dashboard.png in
Example 1 ( 900 × 560 900\times 560 px, drawn by the generator from the latent world). The upper table is
per-service context and plays no role in grading; the lower table holds the three dashboard-only
findings that appear in no CSV, so the agent must read the image to recover their ids, actions, and
savings.
Figure 6: The dashboard screenshot staged as input/aws_billing_dashboard.png in Example 1 (900×560 px, drawn by the generator from the latent world). The upper table is per-service context and plays no role in grading; the lower table holds the three dashboard-only findings that appear in no CSV, so the agent must read the image to recover their ids, actions, and savings.
# input/aws_pricing_reference.csv (complete)
sku,dimension,unit,price_usd
t3.medium,ec2_compute,hour,0.041600
m5.large,ec2_compute,hour,0.096000
m5.xlarge,ec2_compute,hour,0.192000
m5.2xlarge,ec2_compute,hour,0.384000
m5.4xlarge,ec2_compute,hour,0.768000
db.t3.medium,rds_compute,hour,0.068000
db.m5.large,rds_compute,hour,0.171000
db.m5.xlarge,rds_compute,hour,0.342000
db.m5.2xlarge,rds_compute,hour,0.684000
s3-standard,s3_storage,gb_month,0.023000
s3-standard-ia,s3_storage,gb_month,0.012500

# input/resource_usage.csv (6 of 430 rows)
resource_id,resource_type,name,state,instance_type,cpu_p95,mem_p95,connections_p95,size_gb,ia_access_ratio
db-000067,rds,rds-svc-067,available,db.m5.xlarge,62.12,,346,,
db-000048,rds,rds-svc-048,available,db.m5.xlarge,18.48,,8,,
i-000161,ec2,ec2-app-161,running,m5.2xlarge,45.55,74.81,,,
i-000000,ec2,ec2-app-000,running,m5.2xlarge,11.97,41.34,,,
s3-000015,s3,data-bucket-015,,,,,,5923,0.052
s3-000118,s3,data-bucket-118,,,,,,19,0.596
# reference/optimization_report.json (148 recommendations; excerpt, reformatted)
{"executive_summary": {"resource_count": 430, "recommendation_count": 148,
     "dashboard_only_finding_count": 3, "total_current_cost": 32629.73,
     "total_projected_cost": 16234.79, "total_monthly_savings": 16395.02},
 "recommendations": [
  {"resource_id": "i-000003", "resource_type": "ec2", "name": "ec2-app-003",
   "action": "downsize_ec2", "current_cost": 560.64, "projected_cost": 280.32,
   "monthly_savings": 280.32},
  {"resource_id": "i-000006", "resource_type": "ec2", "name": "ec2-app-006",
   "action": "downsize_ec2", "current_cost": 560.64, "projected_cost": 280.32,
   "monthly_savings": 280.32},
  ...
  {"resource_id": "cw-loggroup-old-processor", "resource_type": "cloudwatch_log_group",
   "name": "/aws/lambda/old-processor", "action": "delete_stale_log_group",
   "current_cost": 64.6, "projected_cost": 0.0, "monthly_savings": 64.6},
  ...
  {"resource_id": "nat-0af31c9d", "resource_type": "nat_gateway",
   "name": "nat-gw-prod-1c", "action": "delete_idle_nat_gateway",
   "current_cost": 32.41, "projected_cost": 0.0, "monthly_savings": 32.41},
  ...
  {"resource_id": "eip-52b7e401", "resource_type": "elastic_ip",
   "name": "eip-orphaned-legacy", "action": "release_unattached_eip",
   "current_cost": 3.65, "projected_cost": 0.0, "monthly_savings": 3.65},
  ...]}

# reference/savings_summary.csv (148 rows; first 3)
resource_id,resource_type,name,action,current_cost,projected_cost,monthly_savings
i-000003,ec2,ec2-app-003,downsize_ec2,560.64,280.32,280.32
i-000006,ec2,ec2-app-006,downsize_ec2,560.64,280.32,280.32
i-000024,ec2,ec2-app-024,downsize_ec2,560.64,280.32,280.32

Example 2: closed-loop building MPC (Energy, Power & Nuclear Engineering)

The building-control environment pins EnergyPlus and the full procedure: validate input hashes, edit the thermal model, run thermostat and excitation traces, identify a second-order model with ridge regression, and execute two 24-step receding-horizon controllers in closed loop.

Nine deliverables include 672-row traces, final-day actions, identification/solver logs, metrics, and a summary. References come from the sealed simulator procedure; the grader compares CSV cells and JSON leaves at tolerance 10⁻⁴. A fixed open-loop schedule or wrong-order model cannot reproduce the required closed-loop artifacts.

# EnergyPlus closed-loop building MPC (7 days at 15 minutes)

Use the exact toolchain declared in `input/instance.json`. Inspect `input/model.idf`, `input/weather.epw`, `input/experiment_plan.json`, and `input/output_contract.json`. The model is the pinned EnergyPlus 22.1 one-zone example with a declared internal-mass transformation and a schedule actuator driving the `MPC Cooling Sink`.

Run the deterministic thermostat baseline and supplied excitation at 15-minute resolution. Fit the declared second-order input-output thermal model over both EnergyPlus traces and preserve its training and held-out diagnostics. This is a 3R2C-comparable dynamic-order model, not a claim that physical resistance and capacitance parameters were identified.

Implement two distinct cvxpy/OSQP receding-horizon policies using the declared price, peak-load, comfort, and switching weights. Deploy each through the EnergyPlus Python API: at every zone timestep read the simulated zone temperature, solve the policy, and actuate the first control. A precomputed open-loop schedule does not satisfy this task.

Write the nine artifacts named in `input/output_contract.json`. The three data traces cover the complete simulation window. The two action files and comparison metrics cover the final 24 hours. RC logs must expose model coefficients, training and held-out calibration, policy-specific closed-loop error, objective weights, and solver status counts.

The grader densely compares all nine artifacts to a reference generated and rerun with EnergyPlus 22.1 and cvxpy in the pinned profile.
input/
  question.md           the task prompt above
  instance.json         horizon, schedules, tariff, controller weights (excerpt below)
  experiment_plan.json  four runs, identification recipe and split, two policies
  output_contract.json  the nine deliverables, their columns, and their rounding (below)
  model.idf             one-zone EnergyPlus model, materialized from the pinned example
  weather.epw           Golden, CO (NREL) TMY3 weather file, from the same pinned install
{
  "kind": "energyplus_second_order_closed_loop_mpc_cosimulation", "seed": 6610,
  "energyplus_version": "22.1.0", "cvxpy_version": "1.6.5", "numpy_version": "2.1.3",
  "source_focus": "energyplus_1zoneuncontrolled_22_1",
  "source_idf_sha256": "53e3fb38...", "source_epw_sha256": "65041e11...",
  "weather_window": "golden_tmy3_july",
  "run_period_start_month": 7, "run_period_start_day": 1, "run_period_days": 7,
  "timestep_minutes": 15, "horizon_steps": 672,
  "prediction_horizon_steps": 24, "reporting_window_steps": 96,
  "cooling_capacity_kw": 2.0406, "cooling_cop": 3.0, "internal_mass_area_m2": 48.0,
  "baseline_deadband_c": 0.5, "baseline_on_fraction": 0.85, "identification_ridge": 1e-08,
  "occupancy":            [0, 0, ...],           // 672 values; 1 between 08:00 and 20:00
  "is_peak_period":       [0, 0, ...],           // 672 values; 1 between 17:00 and 20:00
  "setpoint_c":           [27.5, 27.5, ...],     // 24.2 while occupied, else 27.5
  "comfort_min_c":        [17.0, 17.0, ...],     // 20.5 while occupied, else 17.0
  "comfort_max_c":        [28.0, 28.0, ...],     // 24.5 while occupied, else 28.0
  "energy_price_per_kwh": [0.061587, ...],       // 672 values in [0.058, 0.250]
  "excitation_schedule":  [0.9, 0.1, 0.9, ...],  // 672 values in {0.1, 0.9}
  "policy_objectives": {
    "energy_saving":   {"energy_weight": 0.35, "price_weight": 1.0, "peak_weight": 0.1,
      "on_peak_weight": 0.1, "comfort_weight": 2.0, "switch_weight": 0.12},
    "demand_response": {"energy_weight": 0.05, "price_weight": 0.35, "peak_weight": 5.0,
      "on_peak_weight": 2.0, "comfort_weight": 4.0, "switch_weight": 0.12}
  }
}
{
  "trace_files": {"baseline": "baseline_data.csv",
                  "energy_saving": "mpc_energy_saving_data.csv",
                  "demand_response": "mpc_demand_response_data.csv"},
  "trace_columns": ["step", "hour", "cooling_w", "t_zone", "setpoint", "outdoor_temp_c",
                    "occupancy", "cooling_fraction", "electric_kw", "price_per_kwh",
                    "comfort_min_c", "comfort_max_c", "one_step_predicted_zone_temp_c"],
  "action_files": ["mpc_actions_energy_saving.csv", "mpc_actions_demand_response.csv"],
  "rc_log_files": ["rc_log_energy_saving.json", "rc_log_demand_response.json"],
  "metrics_file": "metrics_comparison.csv",
  "metrics_rows": ["baseline", "mpc_energy_saving", "mpc_demand_response"],
  "metrics_columns": ["cooling_kwh", "elec_kwh", "cost_usd", "peak_load_kw",
                      "peak_hour_avg_kw", "discomfort_dh", "calibration_rmse_c"],
  "summary_file": "results_summary.json",
  "numeric_rounding": {"csv": 8, "model_coefficients": 10, "summary": 8}
}
Equation
T⁡[t]=a1​T​[t−1]+a2​T​[t−2]+b0​Tout​[t]+b1​Tout​[t−1]+c0​u​[t]+c1​u​[t−1]+dT[t]=a_{1}T[t-1]+a_{2}T[t-2]+b_{0}T_{\mathrm{out}}[t]+b_{1}T_{\mathrm{out}}[t-1]+c_{0}u[t]+c_{1}u[t-1]+d

Example 3: financial-statement reconstruction (Accounting & Finance)

The finance task reconstructs two years of statements from conflicting structured facts and inline-XBRL filings under an explicit precedence policy. The latent world obeys accounting identities and cross-statement ties. Corruptions affect rejected sources; the selected structured annual cash value is 246, despite original and amended filings reporting 239 and 243.

The reference contains 16 items for each year, four declared exceptions, and 32 evidence rows. A round-trip check independently reselects values from public material before publication; grading compares final files exactly. The preserved listings include the policy, conflicting source excerpts, expected statements, exceptions, and evidence schema.

# Reconstruct and reconcile annual financial statements

Reconstruct the annual statements from the multi-document filing packet under `input/filings/` and the SEC-shaped `input/companyfacts.json`. Use `input/filing_index.json`, `input/account_mapping.csv`, and the authority and fallback rules in `input/reconstruction_policy.md`.

Write three deliverables:

- `output/financial_statements.json`, with company metadata, USD-million unit metadata, and a `periods` array containing balance sheet, income statement, and cash-flow sections.
- `output/evidence.csv`, with the exact columns `period, statement, line_item, selected_value, concept, source_file, locator, citation, authority_basis`, in period and mapping order. Every locator must resolve to the selected source value.
- `output/exceptions.json`, listing every cross-source conflict, explicit amendment restatement, accepted taxonomy alias, and missing-structured-fact fallback with its policy-based resolution.

Preserve integer USD millions. Verify the accounting, income, and cash roll-forward identities for every period before writing the files. Do not use quarterly, unaudited, pro forma, or wrong-unit evidence.
input/
  question.md                      the task prompt above
  instance.json                    2 fiscal years, 16 line items, 3 deliverables
  reconstruction_policy.md         authority and fallback rules (below)
  filing_navigation.md             amendments carry only restated rows; ids are locators
  account_mapping.csv              16 line items -> statement, label, accepted concepts
  filing_index.json                4 filings: form, accession, filing date, status (below)
  companyfacts.json                SEC-shaped facts: 16 us-gaap concepts x 3, whole USD
  filings/2025/form10-k.htm        FY2025 10-K, 16 inline-XBRL rows (fact-2025-000..015)
  filings/2025/form10-k-a.htm      FY2025 10-K/A, restates one row
  filings/2025/current-report.htm  FY2025 8-K with quarterly (Q3) figures
  filings/2024/form10-k.htm        FY2024 10-K
# Annual statement reconstruction policy

All output values are USD millions. Companyfacts `USD` values are whole USD and must be divided by 1,000,000. Use exact fiscal-year and period-end matches, annual `FY` facts, and Forms 10-K or 10-K/A only. Ignore quarterly, non-USD, pro forma, and unaudited supplemental facts.

Within Companyfacts, choose the latest `filed` date and then accession among all accepted concepts in `account_mapping.csv`. Within filing documents, a 10-K/A overrides only the rows it explicitly restates; otherwise retain the original 10-K row. When the latest structured fact and latest filing row disagree, the structured Companyfacts value controls. If no accepted structured fact exists, use the latest applicable filing row. Record every cross-source conflict, amendment restatement, taxonomy alias, and structured-fact fallback in `exceptions.json`. Every evidence row must cite an exact JSON pointer or HTML fragment that resolves to the selected value.
# account_mapping.csv (first of 16 rows)
statement,line_item,filing_label,canonical_concept,accepted_concepts,period_type
balance_sheet,cash,Cash and cash equivalents,us-gaap:CashAndCashEquivalentsAtCarryingValue,us-gaap:CashAndCashEquivalentsAtCarryingValue|jrh:CashCashEquivalentsRestrictedCashAndRestrictedCashEquivalents,instant

# filing_index.json
{"cik": "8899426957", "company": "Juniper Ridge Holdings plc", "filings": [
  {"accession": "8899426957-25-000001", "document": "filings/2025/form10-k.htm",
   "filed": "2026-01-30", "fiscal_year": 2025, "form": "10-K", "status": "original"},
  {"accession": "8899426957-25-000002", "document": "filings/2025/form10-k-a.htm",
   "filed": "2026-02-15", "fiscal_year": 2025, "form": "10-K/A",
   "status": "amendment_explicit_rows_only"},
  {"accession": "8899426957-25-000004", "document": "filings/2025/current-report.htm",
   "filed": "2025-10-15", "fiscal_year": 2025, "form": "8-K",
   "status": "current_report_decoy"},
  {"accession": "8899426957-24-000001", "document": "filings/2024/form10-k.htm",
   "filed": "2025-01-30", "fiscal_year": 2024, "form": "10-K", "status": "original"}]}

# companyfacts.json, node /facts/us-gaap/CashAndCashEquivalentsAtCarryingValue/units/USD
[{"accn": "8899426957-25-000001", "form": "10-K", "fy": 2025, "fp": "FY",
  "end": "2025-12-31", "filed": "2026-01-30", "val": 239000000},
 {"accn": "8899426957-25-000002", "form": "10-K/A", "fy": 2025, "fp": "FY",
  "end": "2025-12-31", "filed": "2026-02-15", "val": 246000000},
 {"accn": "8899426957-24-000001", "form": "10-K", "fy": 2024, "fp": "FY",
  "end": "2024-12-31", "filed": "2025-01-30", "val": 256000000}]

# filings/2025/form10-k.htm, balance sheet ("USD in millions")
<tr><th scope="row">Cash and cash equivalents</th><td><ix:nonFraction id="fact-2025-000"
  name="us-gaap:CashAndCashEquivalentsAtCarryingValue" contextRef="FY2025" unitRef="USD"
  scale="6" format="ixt:numdotdecimal">239</ix:nonFraction></td></tr>
# filings/2025/form10-k-a.htm, the only row it restates
<tr><th scope="row">Cash and cash equivalents</th><td><ix:nonFraction
  id="amended-fact-2025-000" name="us-gaap:CashAndCashEquivalentsAtCarryingValue"
  contextRef="FY2025" unitRef="USD" scale="6" ...>243</ix:nonFraction></td></tr>
# filings/2025/current-report.htm, 8-K with a quarterly context
<tr><th scope="row">Cash and cash equivalents</th><td><ix:nonFraction
  id="current-report-fact-2025-000" name="us-gaap:CashAndCashEquivalentsAtCarryingValue"
  contextRef="Q32025" unitRef="USD" scale="6" ...>230</ix:nonFraction></td></tr>
# reference/financial_statements.json (USD millions; reformatted)
{"cik": "8899426957", "company": "Juniper Ridge Holdings plc", "currency": "USD",
 "unit": "millions", "periods": [
  {"period": "2025", "period_end": "2025-12-31",
   "balance_sheet": {"cash": 246, "accounts_receivable": 147, "inventory": 62,
                     "other_assets": 732, "assets": 1187, "liabilities": 598,
                     "equity": 589},
   "income_statement": {"revenue": 1852, "cost_of_revenue": 666,
                        "operating_expense": 927, "net_income": 259},
   "cash_flow": {"beginning_cash": 256, "operating_cash_flow": 365,
                 "investing_cash_flow": -110, "financing_cash_flow": -265,
                 "ending_cash": 246}},
  {"period": "2024", "period_end": "2024-12-31",
   "balance_sheet": {"cash": 256, "accounts_receivable": 73, "inventory": 30,
                     "other_assets": 546, "assets": 905, "liabilities": 508,
                     "equity": 397},
   "income_statement": {"revenue": 1583, "cost_of_revenue": 601,
                        "operating_expense": 777, "net_income": 205},
   "cash_flow": {"beginning_cash": 106, "operating_cash_flow": 333,
                 "investing_cash_flow": -146, "financing_cash_flow": -37,
                 "ending_cash": 256}}]}

# reference/exceptions.json (reformatted)
[{"kind": "cross_source_conflict", "period": "2025", "line_item": "accounts_receivable",
  "structured_value": 147, "filing_value": 138, "selected_value": 147,
  "resolution": "structured_feed_priority"},
 {"kind": "amendment_restatement", "period": "2025", "line_item": "cash",
  "original_value": 239, "amended_value": 243, "selected_value": 246,
  "resolution": "explicit_amendment_over_original"},
 {"kind": "cross_source_conflict", "period": "2025", "line_item": "cash",
  "structured_value": 246, "filing_value": 243, "selected_value": 246,
  "resolution": "structured_feed_priority"},
 {"kind": "cross_source_conflict", "period": "2025", "line_item": "other_assets",
  "structured_value": 732, "filing_value": 726, "selected_value": 732,
  "resolution": "structured_feed_priority"}]

# reference/evidence.csv (2 of 32 rows)
period,statement,line_item,selected_value,concept,source_file,locator,citation,authority_basis
2025,balance_sheet,cash,246,us-gaap:CashAndCashEquivalentsAtCarryingValue,companyfacts.json,#/facts/us-gaap/CashAndCashEquivalentsAtCarryingValue/units/USD/1,"SEC Companyfacts JSON, us-gaap:CashAndCashEquivalentsAtCarryingValue, FY 2025, companyfacts.json#/facts/us-gaap/CashAndCashEquivalentsAtCarryingValue/units/USD/1",latest_exact_period_annual_structured_fact
2024,balance_sheet,cash,256,us-gaap:CashAndCashEquivalentsAtCarryingValue,companyfacts.json,#/facts/us-gaap/CashAndCashEquivalentsAtCarryingValue/units/USD/2,"SEC Companyfacts JSON, us-gaap:CashAndCashEquivalentsAtCarryingValue, FY 2024, companyfacts.json#/facts/us-gaap/CashAndCashEquivalentsAtCarryingValue/units/USD/2",latest_exact_period_annual_structured_fact

Appendix B Dense Reward Details

PPRS is not policy-invariant. Retaining final potential intentionally distinguishes failed trajectories. Every submitted success outranks every failure, and deterministic policies that achieve success retain that preference, but stochastic-policy rankings can reverse.

If one policy succeeds less often, shaping can prefer it only when the success-probability gap is below λ(1−p)/(1+λ), at most 0.2 for λ=0.25, and its failures make sufficiently more progress. This bound is not a guarantee that learned success remains unchanged.

In 24 measured groups—eight rollouts on 12 warehouse seeds for both base and trained 27B—none of 192 rollouts submits a passing answer. Every group nevertheless has nonzero final-potential variance. The example progress contract gates relations by correct dimensions and facts, discouraging superficially valid downstream outputs.

Total reward depends on final progress and submission, not detours; per-turn differences still assign credit and penalize regression through undiscounted return-to-go. Among all-fail groups the shaping scale cancels from normalized comparisons, so alignment of the progress scorer is essential [57]. The authors test specific exploit probes and independent code closures; no general immunity to reward hacking follows.

Equation
𝔼π1​[R]−𝔼π2​[R]=(p1−p2)​(1+λ)+λ⁡[(1−p1)​Φ¯1−(1−p2)​Φ¯2]≤(p1−p2)​(1+λ)+λ⁡(1−p1),\mathbb{E}_{\pi_{1}}[R]-\mathbb{E}_{\pi_{2}}[R]=(p_{1}-p_{2})(1+\lambda)+\lambda\bigl[(1-p_{1})\bar{\Phi}_{1}-(1-p_{2})\bar{\Phi}_{2}\bigr]\;\leq\;(p_{1}-p_{2})(1+\lambda)+\lambda(1-p_{1}),

Appendix C Training and Evaluation Details

C.1 Training Recipe and Hyperparameters

DPPO uses group-centered advantages, a total-variation policy-ratio bound 0.1, and no KL anchor. Active sampling refills zero-variance groups until each step contains eight informative groups. Each turn receives λ times its progress increment, with completion reward only at the last turn; its advantage is return-to-go minus the group mean at that turn index. λ=0.25 is fixed across runs.

Table 11: Training hyperparameters shared by every RL run in this paper.
Setting Value
Objective group-relative policy gradient with centered advantages (Eq. (2))
Trust region total-variation divergence bound 0.10.1 on the policy ratio (DPPO loss); no KL anchor (β=0\beta{=}0)
Rollouts per task (GG) 24
Unique tasks per step 8 (192 rollouts per step)
Active sampling zero-variance groups discarded and refilled
Optimizer AdamW, learning rate 10−610^{-6}, constant schedule, one epoch per batch
Sampling temperature 1.0
Prompt / response budget 2,048 prompt tokens; 65,536 response tokens
Per-turn budget 16,384 tokens; at most 64 tool calls per rollout
Tools a single bash tool in a per-task container (python:3.13-slim base image)
Reward Eq. (4) with λ=0.25\lambda{=}0.25; per-turn credit, return-to-go with γ=1\gamma{=}1, per-turn-index group baseline
Terminal reward strict verifier score at submission, 1.01.0 for a pass
Precision / parallelism bf16 with fp32 LM head; ZeRO-3, sequence parallelism 4 (9B)
Learner / inference split 16 learner GPUs + 16 inference engines on 32-GPU runs; 16 + 48 on the 64-GPU 27B run
Seed 42

C.2 Evaluation Protocol

ALE-CM evaluates 18 Linux-container-compatible tasks from 23 computing-and-math tasks: eight mathematics/science, five ML, and five coding/systems. Five excluded tasks need full virtual machines. Every model is evaluated on the same selected tasks with five rollouts; Qwen uses Vanillux2 while frontier models use their native agents.

S@1 is mean graded score, P@1 the fraction of fully successful rollouts, and S@5/P@5 their best-of-five counterparts. Terminal-Bench 3.0/4.0 uses 22/18 selected domain-related CPU-compatible tasks; the 27B Terminal-Bench 2.1 evaluation uses Terminus-2, while 9B uses Vanillux2.

The 3.0/4.0 subset P@1 values correlate with full-suite leaderboard values at 0.96/0.97 across compared models, but subset scores often run higher. ALE comparisons are less controlled: full-suite headline values use one run per task and can differ in harness configuration. Tables 12–16 preserve task selection and comparison details; none establishes whole-benchmark equivalence.

Table 12: The 18 ALE-CM tasks (computing-and-math split of Agents’ Last Exam), grouped by the capability category used in Table 2.
Category Task Title
Math & Science (8) branch_bound_atsp Branch-and-Bound ATSP Solver
cfr_game_theory_equilibrium CFR Game Theory Equilibrium
particle_filter_nonlinear_tracking Particle Filter Nonlinear Tracking
k3_abelian_extensions Finite Abelian Extension Classification
synthetic_causal_structure_inference Synthetic Causal Structure Inference
clustered_cyclic_code_circuit_level_simulation Clustered-Cyclic Code Circuit Simulation
ising_post_measurement_1 Post-Measurement States for a Quantum Ising Chain
go_game_reconstruction_1 Go Game Reconstruction 1
ML (5) dit_pipeline_cfg_alignment_fid_256_001 Repair Standalone DiT Pipeline CFG Alignment
paper_reproduction_instance_1 LCA-on-the-Line Table 2 Reproduction
ranking_node_feature_parity_recovery_instance_1 Ranking Node Feature Parity Recovery
mp_checkpoint_consolidation_v2 Checkpoint Consolidation V2
recsys_cold_start_instance_1 Hybrid Recommender Cold-Start Benchmark
Coding & System (5) k8s_payment_api_root_cause_analysis Kubernetes Payment API Root Cause Analysis
os_log_permission_guard_v1 OS Log Permission Guard
data_pipeline_etl_instance_1 Retail ETL Warehouse Cleanup
cp_test_gen_1 Competitive Programming Test Generator – Binary String Copying
cost_optimization_1 AWS Cost Optimization Analysis
Table 13: The Terminal-Bench “CM” subsets evaluated in Table 3: the 22 TB 3.0 tasks in the computing-and-math domain, with the official TB category and whether the task remains in TB 4.0 (18 do).
TB task TB category In TB 4.0
batched-eval-parity ML / Evaluation yes
cargo-flight-dispatch Operations / Logistics yes
cli-2ph-simplex Software / Algorithms –
ctr-optimization Operations / Marketing yes
data-anonymization Software / Data engineering yes
distributed-dedup Software / Systems yes
fix-uautomizer-soundness Software / Languages –
freight-dispatch-shift Operations / Logistics yes
ico-path-patch Security / Reverse engineering –
live-database-cutover Software / Databases yes
memcached-backdoor Security / Reverse engineering –
mp-checkpoint-consolidation ML / Inference yes
mvcc-lsm-compaction Software / Databases yes
payments-pipeline-fix Software / Systems yes
pretrain-shard-corruption ML / Training yes
production-planning Operations / Supply chain yes
risk-scorer-replay ML / Evaluation yes
session-window-debug Software / Systems yes
shadow-relay Security / Forensics yes
telecom-entity-resolution Software / Data engineering yes
uefi-bootkit Security / Forensics yes
wal-recovery-ordering Software / Databases yes

Model Subset (22) Full (74) Official
P@1 P@5 P@1 P@5 acc.
Claude Fable 5 0.41 0.64 0.34 0.61 0.34
Claude Opus 5 0.53 0.64 0.42 0.58 0.43
GPT-5.6 (sol) 0.41 0.55 0.34 0.55 0.35
GPT-5.6 (terra) 0.30 0.55 0.20 0.38 0.21
GPT-5.6 (luna) 0.18 0.32 0.13 0.27 0.14
Grok 4.6 0.30 0.50 0.26 0.45 0.26
Claude Sonnet 5 0.16 0.32 0.14 0.34 0.15
Claude Opus 4.8 0.19 0.50 0.21 0.42 0.21
Grok 4.5 0.26 0.55 0.15 0.34 0.16
GLM 5.2 0.04 0.18 0.04 0.14 0.05
Table 14: TB 3.0: our 22-task “CM” subset versus the full 74-task benchmark, official leaderboard rows of Table 3. “Official acc.” is the published leaderboard accuracy. Across the 10 models: Pearson r=0.96 (P@1) and 0.89 (P@5); Spearman ρ=0.94 and 0.83.

Model Subset (18) Full (66) Official
P@1 P@5 P@1 P@5 acc.
Claude Fable 5 0.46 0.61 0.45 0.68 0.45
Claude Opus 5 0.52 0.72 0.52 0.71 0.52
GPT-5.6 (sol) 0.39 0.61 0.37 0.61 0.37
GPT-5.6 (terra) 0.23 0.44 0.22 0.44 0.22
GPT-5.6 (luna) 0.18 0.33 0.17 0.33 0.17
Grok 4.6 0.31 0.50 0.20 0.39 0.20
Claude Sonnet 5 0.19 0.50 0.12 0.35 0.12
Claude Opus 4.8 0.27 0.50 0.24 0.47 0.24
Grok 4.5 0.20 0.33 0.12 0.27 0.12
Table 15: TB 4.0: our 18-task “CM” subset versus the full 66-task benchmark, same rows and protocol as Table 14 (GLM 5.2 has no 4.0 entry). Across the 9 models: Pearson r=0.97 (P@1) and 0.92 (P@5); Spearman ρ=0.90 and 0.91.
Table 16: Agents’ Last Exam: our 18-task ALE-CM split (Table 2, 5 rollouts per task) versus the official full 152-task leaderboard (one run per task; the leaderboard’s default row for the model in the same harness family, with its reasoning-effort setting). Across the 5 models: Pearson r=0.56 (score) and 0.75 (pass rate); Spearman ρ=0.63 and 0.95.
Model Harness Our ALE-CM split (18 tasks) Official full ALE (152 tasks) Official
S@1 P@1 Score Pass rate effort
GPT-5.6 (sol) Codex 0.58 0.28 0.54 0.31 xhigh
GPT-5.6 (terra) Codex 0.50 0.22 0.51 0.28 max
Claude Opus 5 Claude Code 0.58 0.28 0.55 0.32 high
Claude Opus 4.8 Claude Code 0.52 0.22 0.45 0.27 max
Gemini 3.1 Pro Gemini CLI 0.52 0.21 0.33 0.16 default

C.3 Qualitative Examples on ALE-CM

Four selected rollout comparisons illustrate Kubernetes diagnosis, quantum-state computation, branch-and-bound optimization, and finite-group enumeration. Cases were chosen where Tev beats the base rollout, and frontier agents also solve each task. Thus they explain possible behavioral changes rather than estimate average gain or unique capability. All traced decisions and scores remain in Figures 5 and 7–9 and Table 17.

Table 17: Graded scores of the rollouts traced in Figure 5 and Figures 7–9. Each cell is the score of the single rollout shown in that column; frontier rollouts in their native harness, Qwen rollouts in Vanillux2.
Task Category Best frontier rollout Second frontier rollout Base rollout Ours
k8s_payment_api_root_cause_analysis Coding & Sys. 0.97 (Claude Fable 5) 0.75 (GPT-6 Astra) 0.70 0.97
ising_post_measurement_1 Math & Sci. 1.00 (Claude Fable 5) 1.00 (GPT-5.6 sol) 0.00 1.00
branch_bound_atsp Math & Sci. 1.00 (Claude Fable 5) 1.00 (GPT-5.6 sol) 0.00 1.00
k3_abelian_extensions Math & Sci. 1.00 (Claude Fable 5) 1.00 (GPT-5.6 sol) 0.00 1.00
Figure 7: Qualitative example on the ALE-CM task ising_post_measurement_1 (Math & Science). Compute the exact
ground state of a 10-qubit critical transverse-field Ising chain, couple it to an ancilla copy and derive, for all 2 10 2^{10} projective ancilla outcomes, the post-measurement probabilities, site-1 reduced density matrices and one-body correlators; four
arrays are graded against a hidden reference with binary credit. Green marks a decision that secures the
result, red one that forfeits it. All three solving rollouts cross-check the ground state with an independent construction of H H ; the base model’s exact-diagonalization and imaginary-time codes share one buggy operator
table, so their agreement proves nothing.
Figure 7: Qualitative example on the ALE-CM task ising_post_measurement_1 (Math & Science). Compute the exact ground state of a 10-qubit critical transverse-field Ising chain, couple it to an ancilla copy and derive, for all 210 projective ancilla outcomes, the post-measurement probabilities, site-1 reduced density matrices and one-body correlators; four arrays are graded against a hidden reference with binary credit. Green marks a decision that secures the result, red one that forfeits it. All three solving rollouts cross-check the ground state with an independent construction of H; the base model’s exact-diagonalization and imaginary-time codes share one buggy operator table, so their agreement proves nothing.
Figure 8: Qualitative example on the ALE-CM task branch_bound_atsp (Math & Science). Implement a
branch-and-bound solver in Python/NumPy/SciPy for three asymmetric TSP instances: 12 and 20 cities to proven optimality, 35 cities to
within a 0.5 % gap, reported in a JSON file checked against hidden optima. Green marks a decision that
secures the result, red one that forfeits it. The three solving rollouts write the deliverable and then verify
it independently; the base model has a working bound by t58 but keeps improving heuristics and never writes the file.
Figure 8: Qualitative example on the ALE-CM task branch_bound_atsp (Math & Science). Implement a branch-and-bound solver in Python/NumPy/SciPy for three asymmetric TSP instances: 12 and 20 cities to proven optimality, 35 cities to within a 0.5 % gap, reported in a JSON file checked against hidden optima. Green marks a decision that secures the result, red one that forfeits it. The three solving rollouts write the deliverable and then verify it independently; the base model has a working bound by t58 but keeps improving heuristics and never writes the file.
Figure 9: Qualitative example on the ALE-CM task k3_abelian_extensions (Math & Science). Enumerate every
finite abelian G G in 0 → ℤ / m → G → ( ℤ / 4 ) 3 → 0 0\to\mathbb{Z}/m\to G\to(\mathbb{Z}/4)^{3}\to 0 for m = 1 , … , 8 m=1,\dots,8 in GAP, classify product type, and write an
exact-match JSON. Green marks a decision that secures the result, red one that
forfeits it. The GAP build lacks IsIsomorphic ; our model works around it and writes the deliverable, whereas the base model
has the enumeration right by t118 but spends its remaining turns on a brute-force verifier and never writes the file.
Figure 9: Qualitative example on the ALE-CM task k3_abelian_extensions (Math & Science). Enumerate every finite abelian G in 0→ℤ/m→G→(ℤ/4)3→0 for m=1,…,8 in GAP, classify product type, and write an exact-match JSON. Green marks a decision that secures the result, red one that forfeits it. The GAP build lacks IsIsomorphic; our model works around it and writes the deliverable, whereas the base model has the enumeration right by t118 but spends its remaining turns on a brute-force verifier and never writes the file.

Appendix D Extended Related Work

Repository mining, bug injection, environment reconstruction, and executable tool-world generators provide precedents for scalable agent data. Backward construction likewise includes denoising, controlled database corruption, instruction back-translation, and procedural puzzles. Tev applies a shared latent world to mutually constraining multi-file workflows.

World-state evaluation already appears in AppWorld, τ-bench, WebArena, and OSWorld. Tev’s narrower argument is to scale environments whose latent state supplies both a completion oracle and intermediate progress, with explicit coverage and difficulty controls [74].

Alternatives to sparse reward include task curricula, expert prefixes, learned process judges, and test-pass proxies. PPRS keeps the tasks fixed and increases feedback resolution, but relies on trustworthy progress contracts and intentionally changes the optimized objective. It does not inherit classical policy-invariance guarantees.

Concept demonstration

Open image ↗