diff --git a/benchmark/benchmark_ane_inference.py b/benchmark/benchmark_ane_inference.py
index cdf9e9d50..f48623ffe 100644
--- a/benchmark/benchmark_ane_inference.py
+++ b/benchmark/benchmark_ane_inference.py
@@ -203,6 +203,7 @@ def save_plots(
     ax1.set_xlabel("env_num")
     ax1.set_ylabel("Mean time (ms)")
     ax1.set_xscale("log", base=2)
+    ax1.set_yscale("log")
     ax1.set_xticks(env_nums)
     ax1.set_xticklabels([str(v) for v in env_nums])
     ax1.grid(True, alpha=0.3)
@@ -229,6 +230,7 @@ def save_plots(
     ax2.set_xlabel("env_num")
     ax2.set_ylabel("Envs per second")
     ax2.set_xscale("log", base=2)
+    ax2.set_yscale("log")
     ax2.set_xticks(env_nums)
     ax2.set_xticklabels([str(v) for v in env_nums])
     ax2.grid(True, alpha=0.3)
diff --git a/benchmark/benchmark_backends.py b/benchmark/benchmark_backends.py
index 591cee712..b4fa3496c 100644
--- a/benchmark/benchmark_backends.py
+++ b/benchmark/benchmark_backends.py
@@ -188,7 +188,7 @@ def ew():
 
 def main():
     parser = argparse.ArgumentParser(description="Benchmark NumPy/Torch/MLX compute.")
-    parser.add_argument("--sizes", type=str, default=",".join(str(2**k) for k in range(5, 15)))
+    parser.add_argument("--sizes", type=str, default=",".join(str(2**k) for k in range(5, 12)))
     parser.add_argument("--warmup", type=int, default=1)
     parser.add_argument("--repeat", type=int, default=2)
     parser.add_argument("--dtypes", type=str, default="float16,float32")
diff --git a/benchmark/benchmark_mlp_inference.py b/benchmark/benchmark_mlp_inference.py
index 421c52452..4e760675e 100644
--- a/benchmark/benchmark_mlp_inference.py
+++ b/benchmark/benchmark_mlp_inference.py
@@ -1017,6 +1017,7 @@ def save_plots(
     ax1.set_xlabel("env_num")
     ax1.set_ylabel("Mean time (ms)")
     ax1.set_xscale("log", base=2)
+    ax1.set_yscale("log")
     ax1.set_xticks(env_nums)
     ax1.set_xticklabels([str(v) for v in env_nums])
     ax1.grid(True, alpha=0.3)
@@ -1043,6 +1044,7 @@ def save_plots(
     ax2.set_xlabel("env_num")
     ax2.set_ylabel("Envs per second")
     ax2.set_xscale("log", base=2)
+    ax2.set_yscale("log")
     ax2.set_xticks(env_nums)
     ax2.set_xticklabels([str(v) for v in env_nums])
     ax2.grid(True, alpha=0.3)
diff --git a/scripts/completions/unilab.bash b/scripts/completions/unilab.bash
index cc4895634..b324e051d 100644
--- a/scripts/completions/unilab.bash
+++ b/scripts/completions/unilab.bash
@@ -15,7 +15,15 @@ _unilab_uv_complete() {
         return 0
     fi
 
+    if [[ ${#candidates[@]} -eq 0 ]]; then
+        if [[ $COMP_CWORD -le 2 || ${COMP_WORDS[2]} != "train" && ${COMP_WORDS[2]} != "eval" ]]; then
+            compopt -o default -o bashdefault 2>/dev/null || true
+            return 1
+        fi
+        return 0
+    fi
+
     COMPREPLY=("${candidates[@]}")
 }
 
-complete -F _unilab_uv_complete uv
+complete -o default -o bashdefault -F _unilab_uv_complete uv
diff --git a/scripts/completions/unilab.zsh b/scripts/completions/unilab.zsh
index 3552e4986..3e0f01977 100644
--- a/scripts/completions/unilab.zsh
+++ b/scripts/completions/unilab.zsh
@@ -12,7 +12,12 @@ _unilab_uv_complete() {
             || uv run --no-sync python -m unilab.tools.completion --cword "$((CURRENT - 1))" -- "${words[@]}" 2>/dev/null
     )" || return 0
 
-    [[ -n "$output" ]] || return 0
+    if [[ -z "$output" ]]; then
+        if (( CURRENT <= 3 )) || [[ ${words[3]} != "train" && ${words[3]} != "eval" ]]; then
+            _files
+        fi
+        return 0
+    fi
 
     local -a candidates
     candidates=("${(@f)output}")
diff --git a/src/unilab/tools/completion.py b/src/unilab/tools/completion.py
index 5ea0e313f..479bdbae1 100644
--- a/src/unilab/tools/completion.py
+++ b/src/unilab/tools/completion.py
@@ -17,6 +17,9 @@
     "train": "unilab.cli:train_main",
     "eval": "unilab.cli:eval_main",
 }
+RUN_PATH_ROOTS = ("benchmark", "scripts")
+RUN_PATH_SUFFIXES = (".py", ".sh")
+RUN_PATH_IGNORED_PARTS = ("__pycache__", "outputs")
 SCRIPT_ASSIGNMENT_PATTERN = re.compile(r'^([A-Za-z0-9_.-]+)\s*=\s*"([^"]+)"\s*(?:#.*)?$')
 COMPLETION_BLOCK_START = "# >>> unilab completion >>>"
 COMPLETION_BLOCK_END = "# <<< unilab completion <<<"
@@ -37,6 +40,7 @@ class CompletionMetadata:
     flags: dict[str, tuple[str, ...]]
     choices: dict[str, dict[str, tuple[str, ...]]]
     tasks: tuple[TaskCompletionEntry, ...]
+    run_paths: tuple[str, ...] = ()
 
 
 def _find_project_root(start: Path) -> Path | None:
@@ -75,6 +79,13 @@ def _training_commands(scripts: Mapping[str, str]) -> tuple[str, ...]:
     )
 
 
+def _project_commands(scripts: Mapping[str, str]) -> tuple[str, ...]:
+    training_commands = _training_commands(scripts)
+    commands = [*training_commands]
+    commands.extend(command for command in sorted(scripts) if command not in training_commands)
+    return tuple(commands)
+
+
 def _parser_for_command(command: str) -> argparse.ArgumentParser:
     return cli._train_eval_parser(mode=command)
 
@@ -162,15 +173,49 @@ def _task_entries(root: Path) -> tuple[TaskCompletionEntry, ...]:
     )
 
 
+def _is_run_path(path: Path) -> bool:
+    return path.is_file() and path.suffix in RUN_PATH_SUFFIXES and path.name != "__init__.py"
+
+
+def _has_run_path_descendant(path: Path) -> bool:
+    return any(
+        child.name != "__init__.py" and child.suffix in RUN_PATH_SUFFIXES
+        for child in path.rglob("*")
+    )
+
+
+def _run_path_entries(root: Path) -> tuple[str, ...]:
+    entries: set[str] = set()
+    for path_root in RUN_PATH_ROOTS:
+        base = root / path_root
+        if not base.is_dir():
+            continue
+        entries.add(f"{path_root}/")
+        for path in base.rglob("*"):
+            if path.name.startswith("."):
+                continue
+            relative_parts = path.relative_to(root).parts
+            if any(part in RUN_PATH_IGNORED_PARTS for part in relative_parts):
+                continue
+            relative_path = path.relative_to(root).as_posix()
+            if path.is_dir():
+                if _has_run_path_descendant(path):
+                    entries.add(f"{relative_path}/")
+            elif _is_run_path(path):
+                entries.add(relative_path)
+    return tuple(sorted(entries))
+
+
 def build_metadata(root: Path | None = None) -> CompletionMetadata:
     selected_root = root or _find_project_root(Path.cwd()) or cli.repo_root()
     scripts = _read_project_scripts(selected_root / "pyproject.toml")
-    commands = _training_commands(scripts)
+    training_commands = _training_commands(scripts)
     return CompletionMetadata(
-        commands=commands,
-        flags={command: _parser_flags(command) for command in commands},
-        choices={command: _parser_choices(command) for command in commands},
+        commands=_project_commands(scripts),
+        flags={command: _parser_flags(command) for command in training_commands},
+        choices={command: _parser_choices(command) for command in training_commands},
         tasks=_task_entries(selected_root),
+        run_paths=_run_path_entries(selected_root),
     )
 
 
@@ -190,6 +235,25 @@ def _matching(candidates: Sequence[str], prefix: str) -> list[str]:
     return [candidate for candidate in candidates if candidate.startswith(prefix)]
 
 
+def _dedupe(candidates: Sequence[str]) -> list[str]:
+    return list(dict.fromkeys(candidates))
+
+
+def _path_choices(candidates: Sequence[str], prefix: str) -> list[str]:
+    choices: set[str] = set()
+    for candidate in candidates:
+        if not candidate.startswith(prefix):
+            continue
+        remainder = candidate[len(prefix) :]
+        if remainder == "":
+            continue
+        if "/" in remainder:
+            choices.add(f"{prefix}{remainder.split('/', 1)[0]}/")
+        else:
+            choices.add(candidate)
+    return sorted(choices)
+
+
 def _option_state(words: Sequence[str], cword: int) -> tuple[dict[str, str], set[str]]:
     values: dict[str, str] = {}
     used: set[str] = set()
@@ -253,7 +317,12 @@ def complete_words(
 
     current = _current_word(words, cword)
     if cword <= 2:
-        return _matching(selected_metadata.commands, current)
+        return _dedupe(
+            [
+                *_matching(selected_metadata.commands, current),
+                *_path_choices(selected_metadata.run_paths, current),
+            ]
+        )
 
     command = words[2]
     if command not in selected_metadata.commands:
diff --git a/tests/test_completion.py b/tests/test_completion.py
new file mode 100644
index 000000000..8ecc1bb88
--- /dev/null
+++ b/tests/test_completion.py
@@ -0,0 +1,53 @@
+from __future__ import annotations
+
+from pathlib import Path
+
+from unilab.tools.completion import build_metadata, complete_words
+
+
+def _write_completion_fixture(root: Path) -> None:
+    (root / "pyproject.toml").write_text(
+        "\n".join(
+            [
+                "[project.scripts]",
+                'train = "unilab.cli:train_main"',
+                'eval = "unilab.cli:eval_main"',
+                'demo = "unilab.cli:demo_main"',
+            ]
+        ),
+        encoding="utf-8",
+    )
+    (root / "benchmark" / "core").mkdir(parents=True)
+    (root / "benchmark" / "benchmark_sim.py").write_text("", encoding="utf-8")
+    (root / "benchmark" / "core" / "runner.py").write_text("", encoding="utf-8")
+    (root / "scripts").mkdir()
+    (root / "scripts" / "play_viser.py").write_text("", encoding="utf-8")
+
+
+def test_uv_run_command_position_includes_project_scripts_and_run_paths(tmp_path: Path) -> None:
+    _write_completion_fixture(tmp_path)
+    metadata = build_metadata(tmp_path)
+
+    assert "demo" in complete_words(["uv", "run", "d"], 2, metadata)
+    assert complete_words(["uv", "run", "b"], 2, metadata) == ["benchmark/"]
+
+
+def test_uv_run_path_completion_is_hierarchical(tmp_path: Path) -> None:
+    _write_completion_fixture(tmp_path)
+    metadata = build_metadata(tmp_path)
+
+    benchmark_choices = complete_words(["uv", "run", "benchmark/"], 2, metadata)
+    assert "benchmark/benchmark_sim.py" in benchmark_choices
+    assert "benchmark/core/" in benchmark_choices
+    assert "benchmark/core/runner.py" not in benchmark_choices
+
+    assert complete_words(["uv", "run", "benchmark/core/"], 2, metadata) == [
+        "benchmark/core/runner.py"
+    ]
+
+
+def test_uv_run_unknown_command_arguments_defer_to_shell_completion(tmp_path: Path) -> None:
+    _write_completion_fixture(tmp_path)
+    metadata = build_metadata(tmp_path)
+
+    assert complete_words(["uv", "run", "benchmark/benchmark_sim.py", ""], 3, metadata) == []