用 6 个决策模型搭建魔方决策竞技场（Laya、Clef、GLiNER 2.5、Kev 与 Strands Decider）

What happens when six small AI models — four of them running on a laptop — are asked to solve the same scrambled cube, one move at a time? Most "AI plays a game" demos hide the interesting part. They show a model winning, and you are left to guess how much of that was the model and how much was the harness. This project does the opposite: it puts six models in a glass box, gives them the exact same scrambled Rubik's cube, the exact same question every turn, and shows you — live, in one screen — exactly where each one succeeds and where it falls apart. The result is uncomfortable and honest. A Rubik's cube is a planning problem: you need lookahead and search. The models in this arena are decision models: fast, calibrated judgments about the current state. They are not the same thing, and the arena makes that difference visible. 1. The premise: decision models vs. planning A "decision model" (sometimes called a system-one model) does one thing well: it looks at a situation, weighs a closed set of options, and returns a choice with a confidence. It is fast and it is often surprisingly well-calibrated. A Rubik's cube is not a decision problem. It is a search problem. From a scrambled state there is an optimal solution length, and finding it means exploring a tree of moves. A model that only ever looks at the current state — with no memory of the plan and no lookahead — is being asked to do something it was not built for. So the arena asks a precise question: Given the same cube and the same one-move question, how good are six real decision models at picking moves that actually help? Everything else in the project exists to answer that question fairly. 2. The two hard rules The whole comparison is only meaningful if the conditions are identical, so the code enforces two rules everywhere: Every engine gets the identical state and the identical question. No engine gets extra context, a better prompt, retries, or a second look. Same 54 stickers, same 18 options, same instructions. Nothing is ever faked. If a model or its server is missing, its panel reads unavailable with the exact reason, and the race simply runs without it. There are no placeholder scores and no synthetic results. That second rule sounds obvious until you build a demo. It is very tempting to fill a dead panel with something plausible. The arena refuses. 3. Architecture ┌──────────────────────────────────┐ │ browser http://127.0.0.1:8077 │ │ ui/index.html (CSS-3D cubes) │ └───────────────┬──────────────────┘ GET / GET /api/engines │ POST /api/run GET /api/run/ GET /api/chart/ │ ┌───────────────▼──────────────────┐ │ server.py (stdlib HTTP) │ │ Hub: loaded engines + active run │ │ CHART_LOCK serialises matplotlib │ └───────┬───────────────────┬───────┘ │ │ one thread per engine (run in parallel) │ ▼ ▼ ┌────────────────────┐ ┌─────────────────────────────┐ │ arena.run() │ │ charts.build_scorecard() │ │ guided | greedy │ │ -> results/chart_.png │ │ | staged │ │ progress · latency · │ └─────────┬──────────┘ │ guided · final │ │ └─────────────────────────────┘ each turn: decide(state, question) ▼ ┌────────────────────────────────────────────────────────────────┐ │ engines.py — one contract, six adapters │ │ laya MLX, in-process gliner CPU, in-process │ │ strands HTTP 127.0.0.1:8010 kev HTTP 127.0.0.1:8008 │ clef / clef-flash Cloudflare Workers AI (HTTPS) │ └────────────────────────────────────────────────────────────────┘ │ ▼ ┌────────────────────────────────────────────────────────────────┐ │ cube.py (pycuber state + metrics) │ │ solver.py (kociemba plan / first move) │ └────────────────────────────────────────────────────────────────┘ Enter fullscreen mode Exit fullscreen mode The stack is deliberately boring: a stdlib HTTP server, a single HTML file, and plain Python. There is no framework, no bundler, and no build step. You can read the whole thing in an afternoon. Tech stack Layer Choice Cube simulation pycuber Optimal search kociemba (two-phase) Local runtimes MLX (laya), torch/MPS (strands), transformers/CPU (gliner2) Cloud inference Cloudflare Workers AI (clef, clef-flash) HTTP client httpx Web server Python stdlib http.server + ThreadingHTTPServer UI single-file HTML + CSS 3D transforms + vanilla JS Charts matplotlib (Agg) Terminal rich Automation Playwright Env uv + Python 3.12 4. The cube model Everything starts with a deterministic scramble. Given a seed, every engine sees the same cube — and the same cube every time you rerun the demo. # cube.py FACES = "URFDLB" MOVES = [f + s for f in FACES for s in ("", "'", "2")] # 18 quarter/half turns def scramble(seed: int, n: int = 25) -> tuple[Cube, list[str]]: rng = random.Random(seed) seq: list[str] = [] prev_face = None while len(seq) int: n = 0 for f in FACES: ctr = centre_colour(cube, f) face = _face(cube, f) for r in range(3): for c in range(3): if face[r][c].colour == ctr: n += 1 return n Enter fullscreen mode Exit fullscreen mode The model never sees a Cube object. It sees a compact text state, identical for every engine: # cube.py def state_text(cube: Cube) -> str: rows = facelet_rows(cube) body = "\n".join(f"{f}: {rows[f]}" for f in FACES) return ( "3x3 Rubik's cube. Faces U R F D L B, each 9 stickers read row by row " "seen from outside. Letters are sticker colours (W Y G B R O). A face is " "solved when all 9 stickers match its fixed centre sticker.\n" f"{body}\n" f"Correct stickers: {solved_facelets(cube)}/54. " f"Complete faces: {solved_faces(cube)}/6." ) Enter fullscreen mode Exit fullscreen mode 5. The solver bridge (and its encoding trap) Guided mode needs an optimal-ish solver. The project uses kociemba, but wiring it up has a classic gotcha: pycuber colours its stickers with colour names, and Kociemba wants a 54-character string of face letters in URFDLB order. Hardcoding the colour→face map is wrong, because pycuber's default cube is not the usual white-up arrangement. The fix is to read the map off the centres, which never move: # solver.py FACES = "URFDLB" def colour_map(cube: Cube) -> dict[str, str]: """colour name -> face letter, taken from the fixed centres.""" return {cube.get_face(f)[1][1].colour: f for f in FACES} def facelet_string(cube: Cube) -> str: """54-char Kociemba facelet string, URFDLB order, row-major from outside.""" m = colour_map(cube) out: list[str] = [] for f in FACES: face = cube.get_face(f) for r in range(3): for c in range(3): out.append(m[face[r][c].colour]) return "".join(out) @lru_cache(maxsize=200_000) def _solve_cached(state: str) -> str: return kociemba.solve(state) Enter fullscreen mode Exit fullscreen mode Every state is cached by its facelet string. The race revisits states constantly (neutral moves, repeated positions), and Kociemba is by far the slowest part of the loop — the cache is the difference between a snappy demo and a stall. Two more things worth knowing if you build on this: kociemba has no Apple-silicon wheel, so it always compiles a small C extension. If that build fails, the module degrades cleanly: greedy and staged modes still work, and guided mode says exactly what is missing. The solver's solution length is near-optimal, not provably optimal. That is why guided runs take optimal + a few moves rather than exactly optimal. 6. One contract, six engines This is the heart of the project. Six very different runtimes — an in-process MLX model, a torch server, a CPU classifier, two cloud endpoints, and a local OpenAI-ish server — are all reduced to one method: # engines.py def decide(self, state: str, question: dict, qid: str = "move") -> Decision: ... @dataclass class Decision: engine: str choice: str # e.g. "R2" confidence: float | None probabilities: dict[str, float] # over the 18 moves (may be empty) latency_ms: float error: str | None = None Enter fullscreen mode Exit fullscreen mode The question is a closed choice over the 18 legal moves, with a description for each — the shape that strands-decider requires, and that the others accept too: # engines.py _FACE_NAME = {"U": "up", "R": "right", "F": "front", "D": "down", "L": "left", "B": "back"} MOVE_CRITERIA = {} for _f, _n in _FACE_NAME.items(): MOVE_CRITERIA[_f] = f"turn the {_n} face clockwise 90 degrees" MOVE_CRITERIA[_f + "'"] = f"turn the {_n} face anticlockwise 90 degrees" MOVE_CRITERIA[_f + "2"] = f"turn the {_n} face 180 degrees" def move_question(instructions: str) -> dict: return {"type": "choice", "instructions": instructions, "criteria": dict(MOVE_CRITERIA)} Enter fullscreen mode Exit fullscreen mode Most engines share an HTTP transport, so adding one is usually a tiny subclass: # engines.py class HttpEngine(Engine): def probe(self) -> tuple[bool, str | None]: """A dead server must read as unavailable, not as 'ready'.""" if not self.health_url: return True, None try: r = httpx.get(self.health_url, timeout=3.0) except Exception as e: return False, f"no server at {self.health_url} ({type(e).__name__})" return True, None def decide(self, state, question, qid="move") -> Decision: body = {"state": state, "model": self.model, "questions": {qid: question}} t0 = time.perf_counter() try: r = self.client.post(self.url, json=body) r.raise_for_status() resp = r.json() except Exception as e: return Decision(self.name, "", None, latency_ms=(time.perf_counter() - t0) * 1000, error=f"{type(e).__name__}: {e}") dt = (time.perf_counter() - t0) * 1000 choice, conf, probs = _parse_answers(resp, qid) return Decision(self.name, choice, conf, probs, dt) Enter fullscreen mode Exit fullscreen mode The probe() method is why a dead Kev server shows unavailable: no server at 127.0.0.1:8008 instead of silently reporting "ready" and failing on every move. The six models Model Params Where How Laya (aac6fef/laya-mlx) 421M local, Apple silicon in-process MLX Strands Decider 2B 1.9B local, torch/MPS HTTP POST /v1/systemone GLiNER2.5-Decide 340M local, CPU in-process classify_text Clef-Flash 9B Cloudflare Workers AI HTTPS Clef 27B Cloudflare Workers AI HTTPS Kev-0.8B 0.8B local server :8008 HTTP POST /v1/systemone Why two run in the cloud: this was built on an Apple M2 mini with 8 GB of RAM. A 27B model does not fit at any quantization, so Clef and Clef-Flash are called on Workers AI where they cost no local RAM. The other four run locally — including on a 16 GB laptop — which is the point of the demo. 7. The guided loop: the solver plans, the model decides The unaided models cannot solve a cube. That is measured, not assumed: after 24 moves each, three different architectures all landed on exactly the sticker count they started from. So guided mode separates the two jobs cleanly — the solver is the search, the model is the decision layer inside it. # arena.py (abridged) plan = solution(cube) # a concrete, known-good move list correct = plan[0] # the solver's optimal next move d = engine.decide(state_text(cube), move_question(GREEDY_Q)) mv = pick_move(d, policy, rng) # sample the distribution, or take argmax play, tag, replan = correct, "corrected", False if mv == correct: tag = "hit" # matched the solver exactly elif mv in MOVES: probe = cube.copy(); apply_move(probe, mv) d_m = len(solution(probe)) if d_m 0: play, tag, replan = mv, "neutral", True # wasted turn (budgeted) neutral_budget -= 1 else: play, tag = correct, "corrected" # solver overrides else: play, tag = correct, "corrected" # unusable answer apply_move(cube, play) plan = solution(cube) if replan else plan[1:] Enter fullscreen mode Exit fullscreen mode The model gets the same 18-move question as the unaided race, so the two are directly comparable. Each pick is scored: Outcome Meaning hit the model picked exactly the solver's next move accepted a different move that still made strict progress neutral the move did not change the distance — a wasted turn (budgeted) corrected the model's move was worse or unusable, so the solver played its own move The cube is guaranteed to advance because the solver always holds a valid plan, and after the neutral budget is spent every move must strictly reduce the distance. But here is the subtlety that trips people up: a corrected turn still costs one of the moves you allocated. The solver is doing the solving, and the model is being measured — but the clock runs on every turn, including the ones where the model was overruled. With an optimal distance of 22 and a neutral budget of 6, a guided run can legitimately need up to 28 moves. Set the cap below that and slower models get cut off unsolved while still showing the done status. Policy: how a distribution becomes a move Most of these models return a probability for each of the 18 moves. Policy decides what to do with it: # arena.py def pick_move(d, policy: str, rng: random.Random) -> str: if policy == "sample" and d.probabilities: opts = [(m, v) for m, v in d.probabilities.items() if m in MOVES and v > 0] if opts: total = sum(v for _, v in opts) r = rng.random() * total acc = 0.0 for m, v in opts: acc += v if r None: threads = [] for panel in run["panels"]: if panel["status"] != "ready": continue t = threading.Thread(target=self._one_engine, args=(run, panel), daemon=True) t.start() threads.append(t) for t in threads: t.join() # ... persist the run + render the chart, THEN flip status to "done" Enter fullscreen mode Exit fullscreen mode Every turn, the engine thread fires an on_move callback that updates the shared panel under a lock. The browser polls /api/run/ every 350 ms and redraws the cubes. There is one nasty bug worth calling out, because it is the kind of thing that only shows up in a demo: matplotlib's pyplot is not thread-safe. The scorecard is built on the run thread when a race finishes, and on demand by the /api/chart/ route. If those two ever overlap, the render corrupts and the browser shows a broken image. The fix is two-fold — a lock, and getting the order right: # server.py CHART_LOCK = threading.Lock() # ... in _drive(), after all engine threads join: snapshot = {**run, "status": "done", "finished_at": finished_at} (RESULTS / f"ui_run_{run['id']}.json").write_text(json.dumps(snapshot, indent=2)) with CHART_LOCK: build_scorecard(run, RESULTS / f"chart_{run['id']}.png") with self.lock: run["status"] = "done" # flipped LAST, only after the PNG exists Enter fullscreen mode Exit fullscreen mode The status flips to done after the PNG is on disk, so the page can never request a chart that does not exist yet. The front-end adds a retry with a cache-buster as a belt-and-braces measure. The API is four routes: Route Purpose GET / the single-page UI GET /api/engines per-engine load status (polled on startup) POST /api/run start a race {seed, mode, policy, moves, engines} GET /api/run/ current run state (polled while running) GET /api/chart/ the scorecard PNG 9. The UI: a real 3D cube with no library The cube is not an image or a canvas — it is six DOM faces positioned in 3D space with CSS transforms, slowly rotating. That is the whole trick: /* ui/index.html */ .scene { height: 164px; perspective: 1000px; perspective-origin: 50% 45%; } .cube3d { position: relative; width: var(--cube); height: var(--cube); transform-style: preserve-3d; animation: spin 46s linear infinite; } .face3d.U { transform: rotateX(90deg) translateZ(calc(var(--cube)/2)); } .face3d.D { transform: rotateX(-90deg) translateZ(calc(var(--cube)/2)); } .face3d.F { transform: translateZ(calc(var(--cube)/2)); } .face3d.B { transform: rotateY(180deg) translateZ(calc(var(--cube)/2)); } .face3d.L { transform: rotateY(-90deg) translateZ(calc(var(--cube)/2)); } .face3d.R { transform: rotateY(90deg) translateZ(calc(var(--cube)/2)); } Enter fullscreen mode Exit fullscreen mode Each face is a 3×3 grid of stickers whose colours come straight from the model's net (the per-face colour strings). The layout is three columns by two rows, with a short-screen media query that shrinks the cube so all six stay in one frame. Two front-end details earned their keep during development: Persistence. When a run finishes, runId is cleared and the engine poll resumes. The first version re-rendered empty placeholders and every cube vanished at the moment of victory. The fix keeps the last run's panels and re-renders those instead, so the final solved state stays on screen. The chart is the true "done" signal. startRace() is async, so the status text briefly still reads "models ready". Tests wait on the chart image actually loading (naturalWidth > 0), not on the status string. 10. The scorecard When a race finishes, the server renders a four-panel dark scorecard with matplotlib: Progress — correct stickers over moves, one line per model. Latency — median ms per move, log scale. Guided resolution — stacked hit / accepted / neutral / corrected bars, so you can see at a glance how much of the solve was the model. Final — stickers correct at the end, solved models marked. Colours are stable per engine key, so a model keeps its colour across runs. 11. What we actually measured Short version; the full write-up is in REPORT.md. Unaided (greedy / staged), same scramble, seed 20261002: after 24 moves each, three different architectures landed on exactly the sticker count they started from — zero net progress, and no face ever completed. Even 60 moves of subgoal-named play (staged mode) topped out at 17 of 54 stickers. Guided mode, same scramble (optimal distance 22): every model solves the cube, because the solver is doing the solving. The honest reading is in the counters: model solved moves hits accepted neutral solver corrected laya-mlx yes 28 0 1 6 21 strands-decider yes 28 0 2 6 20 GLiNER2.5-Decide yes 27 4 0 5 18 Read it carefully: 18 to 21 of every 27–28 moves were the solver overriding the model. laya and strands never once picked the solver's move. GLiNER's 4 hits are an artefact, not skill. It returns a single label with no distribution, so it cannot be sampled — it plays U almost every turn, and the optimal plan happened to start with U four times. Wall clock is the real race. Same solve, but laya finishes in ~5 s against strands' ~64 s and GLiNER's ~95 s. The headline is not "who won". It is that small decision models do not plan, and the arena shows exactly where that breaks down — while never pretending the model did something it didn't. 12. Hard-won lessons A few things that cost real debugging time and are worth stealing: kociemba has no Apple-silicon wheel. It always compiles a C extension. Keep it in its own pip install line, or a failed build aborts the whole command and takes pycuber down with it. matplotlib pyplot is not thread-safe. Serialise every render with a lock, and write the PNG before you advertise that the run is done. A dead server must be probed, not assumed. probe() is what turns a silent failure into an honest unavailable panel. Never wipe the final state on the next poll. Keep the last run's panels and re-render those, or the cubes disappear the instant the race ends. The move cap is part of the experiment. In guided mode a corrected turn still costs a move, so a cap below optimal + neutral_budget silently cuts models off unsolved. argmax on a decision model collapses. Sampling from the returned distribution is the honest way to run a model in a loop. Build on the centre stickers, not on assumed colours. pycuber's default cube is not the usual white-up arrangement, so derive the colour→face map from the fixed centres. 13. Run it, fork it, extend it git clone https://github.com/harishkotra/cube-arena.git && cd cube-arena uv venv --python 3.12 && source .venv/bin/activate uv pip install laya-mlx strands-decider gliner2 pycuber rich matplotlib requests httpx uv pip install kociemba python server.py # open http://127.0.0.1:8077 Enter fullscreen mode Exit fullscreen mode Everything is optional except pycuber: a model without credentials or a running server simply shows unavailable. ./run.sh starts the strands server and the UI together; ./run_kev.sh starts the Kev server. There is no framework and no build step, so contributing is easy. The most common contribution is a new engine: subclass HttpEngine (or Engine), register it in ENGINE_SPECS in server.py, and the UI, charts, and CLI pick it up automatically. Other good first contributions: per-move regret and calibration metrics, a best-of-N mode, an OpenAI-compatible adapter, a /api/runs history index, and a multi-seed suite. The full list is in the README's "Feature ideas" section. If you build something on top of this, I'd love to see it. The whole point is that comparison under identical conditions is more interesting than a leaderboard. Screenshots Watch Code & more: https://www.dailybuild.xyz/project/277-cube-arena