Fudan University  ·  Shanghai Innovation Institute  ·  ACL 2026 (Findings)  ·  arXiv:2601.11940

ACL 2026 · Findings

When does thinking longer stop helping?

Thinking Traps — after an early wrong commitment, a model elaborates a self-consistent but incorrect prefix. TAAR cuts before the trap and restarts the thought.

TAAR·cut the trap, restart the thought

Don’t argue with the trap. Leave it.

Kang Chen∗  ·  Fan Yu∗  ·  Junjie Nian  ·  Shihan Zhao  ·  Zhuoka Feng  ·  Zijun Yao  ·  Heng Wang  ·  Minshen Yu  ·  Yixin Cao†

Watch the film The idea Read the paper

The one idea

More tokens don't help if they elaborate the wrong start.

Long chain-of-thought scales test-time compute — but longer traces also demand the ability to revise early assumptions. In many failed runs an early wrong commitment dominates everything after it: even when the model later reflects, verifies, or tries an alternative, it refines the error's consequences instead of fixing the root cause. We call this prefix-dominant deadlock a Thinking Trap — and on a curated DAPO-MATH subset, 89% of failures are traps.

Conceptual illustration of Thinking Traps: a reasoner facing a trap can, guided by a diagnostic policy, choose an escape strategy — step over (no intervention), jump (mild intervention), or use a ladder (strong intervention) to get past the trapped region toward the correct answer.
Fig. 1Facing a trap, a diagnostic policy chooses how to escape: step over (no intervention), jump (mild), or a ladder (strong). The trap itself is not worth arguing with.
01

An early commitment dominates

Once a wrong assumption is anchored, the trajectory becomes prefix-dominant: later computation refines the error's consequences rather than revising its root cause.

02

Reflection that fails

Downstream self-repair — reflection, verification, alternative attempts — largely doesn't work: keeping the trapped prefix and correcting after it escapes far less often than cutting the trap out.

→

A test-time-scaling bottleneck

Traps aren't just another error category: they burn large token budgets without buying correctness. Removing them is where extra compute actually pays off.

How TAAR works

Diagnose the trap. Cut before it. Restart — as hard as needed.

TAAR is a test-time controller — it never fine-tunes the base reasoning model. A lightweight diagnostic policy reads a partial trajectory and predicts two signals: where the trap is and how deep it is; the controller then reallocates compute from the trapped continuation toward a fresh re-derivation.

The TAAR framework in three stages. Left: a diagnostic data construction pipeline segments Long-CoT trajectories and labels trap indices and escape probabilities via GLM-4.7 annotation with manual verification. Middle: supervised training of the diagnostic policy to predict the trap index and escape probability from a partial trajectory. Right: an adaptive restart controller that truncates before the predicted trap and selects no, mild, or strong intervention based on the predicted escape probability.
Fig. 2TAAR. Left: build diagnostic labels (trap index + escape probability) with an LLM judge and human checks. Middle: train a small policy to predict them from partial reasoning. Right: a controller truncates before the trap and picks intervention strength from p̂.
  1. diagnose

    Predict the trap index t̂ and escape probability p̂

    A Qwen3-4B policy, SFT'd on GLM-4.7-judged trajectories (93% human agreement), reads a segmented prefix and predicts the earliest wrong-commitment segment (t̂) and how escapable the trap is (p̂) — from as little as 20% of the trace.

    pescape = 1N ∑n 𝟙[Correct(ŷ(n))]
  2. cut

    Truncate before the trap

    Instead of correcting after the error, TAAR keeps only the prefix before the predicted trap segment, Y<t̂, and regenerates from there — removing the wrong commitment from the effective context.

  3. restart

    Adaptively restart by how deep the trap is

    p̂ gates the intervention: ≥ 0.6 step over (keep going), 0.1–0.6 mild resample, ≤ 0.1 strong — higher-temperature resampling with an optional structured reboot suffix that requests a fresh derivation and a constraint checklist.

TAAR touches no base-model weights and uses no ground-truth answers at test time. It is a controller, not a new model — the diagnostic policy decides only where to cut and how strongly to restart under a fixed compute budget.

Why cut, not correct?

Once the trap is set, downstream fixes rarely land.

Two measurements justify TAAR's design. First, where you cut matters enormously: removing the trap beats correcting after it. Second, the trap can be spotted early, so control can happen online — before the long chain even finishes.

Left: escape rate by cut strategy across model scales — truncating at the trap segment (Cut@Trap) yields much higher escape rates than cutting at post-trap self-repair windows or at random positions. Right: trap position distribution — traps concentrate in the early portion of trajectories, with 44.4% occurring before relative position 0.2.
Fig. 3Left: Cut@Trap escapes far more often than post-trap or random cuts (e.g. 17.5% vs 6.7% vs 4.3% at 20B). Right: traps sit early — 44.4% before relative position 0.2.
①

Cut the trap, don't patch it

Restarting at the trap segment reaches the answer 17.5% of the time vs 6.7% from post-trap reflection windows and 4.3% from random cuts (20B; the same ordering holds at 4B and 8B). Prefix-dominance makes late fixes weak.

②

Traps are diagnosable early

Trap detection exceeds 0.7 AUC with only 20% of the trajectory and saturates by 40–60% — so TAAR can decide whether to intervene online, without waiting for the full chain.

→

p̂ separates deep from shallow

The predicted escape probability is discriminative — correct trajectories cluster near 0.95, incorrect ones near 0.40 — a usable control signal for how strongly to restart.

Left: trap detection AUC-ROC versus chain-of-thought prefix length by model, exceeding 0.7 at only 20% of the trajectory and saturating around 40 to 60 percent. Right: violin plot of predicted escape probability, where correct trajectories concentrate at high values near 0.95 and incorrect ones spread toward lower values near 0.40.
Fig. 4Left: traps are detectable from a short prefix (AUC > 0.7 at 20%). Right: predicted p̂ separates correct from incorrect trajectories — the signal that gates intervention strength.

The payoff

Same compute, more correct answers — no fine-tuning.

Under a matched budget of K=4 sampled paths, TAAR reallocates compute from trapped continuations toward genuinely different solution paths — lifting accuracy on five hard math and science benchmarks without changing a single base-model weight.

89%of curated failures are Thinking Traps
+4.6ptaverage gain on an 8B model over AVG@4
0.7trap-detection AUC from only 20% of the trace

On DeepSeek-R1-8B, TAAR reaches 69.1 average vs 64.5 (AVG@4) and 66.0 (PRM@4) — +4.6 / +3.1 — with the largest gains on the hardest suites (HMMT25 +7.5, BRUMO25 +5.9, AIME24 +5.0). The 4B model gains +1.7 and the 20B model +1.5, all without fine-tuning base weights.

Beats sampling and reward selection

TAAR tops both standard multi-sample averaging (AVG@4) and an outcome-based process-reward selector (PRM@4) on the mid-scale 8B model, and is competitive at 4B and 20B — reallocating budget beats re-ranking trapped candidates.

Where restart helps — and where it doesn't

Gains concentrate where reasoning is fragile. On the very strong 120B model TAAR is competitive but not consistently ahead (79.2): when prefixes are already near-correct, aggressive restarts can discard useful work — an honest limit of trap-cutting.

A controller, not a bigger model

TAAR spends the same tokens differently. It never edits base weights, never sees gold answers at inference, and localizes over paragraph segments — so boundary ambiguity and judge-label noise are real limits. The gain is structural: stop pouring compute into a trapped prefix.

Built from 4 models (4B–120B) · 5 benchmarks · 6,000 trajectories Qwen3-4B · DeepSeek-R1-8B · GPT-OSS-20B / 120B · AIME24 · AIME25 · BRUMO25 · HMMT25 · GPQA-Diamond · K=4 budget · GLM-4.7 judge (93% human agreement)

Watch

Cut the trap, restart the thought — a visual essay

A silent vector essay, drawn live in your browser — no video file. An early wrong commitment traps the thought; reflection fails; diagnose the trap, cut before it, and adaptively restart — step over, resample, or reboot.

Cite

Reference

Kang Chen1,∗, Fan Yu1,∗, Junjie Nian1, Shihan Zhao1, Zhuoka Feng1, Zijun Yao, Heng Wang1, Minshen Yu1, Yixin Cao1,2,†

1 Fudan University    2 Shanghai Innovation Institute    ∗ Equal contribution    † Corresponding author

Correspondence: yxcao@fudan.edu.cn

Links: Paper (PDF)  ·  arXiv  ·  Code

@misc{chen2026thinkingtraps,
  title   = {Thinking Traps in Long Chain-of-Thought:
             A Measurable Study and Trap-Aware Adaptive Restart},
  author  = {Chen, Kang and Yu, Fan and Nian, Junjie and Zhao, Shihan and
             Feng, Zhuoka and Yao, Zijun and Wang, Heng and Yu, Minshen and Cao, Yixin},
  year    = {2026},
  eprint  = {2601.11940},
  archivePrefix = {arXiv},
  primaryClass = {cs.AI},
  note    = {TAAR — Trap-Aware Adaptive Restart. ACL 2026 Findings},
  url     = {https://arxiv.org/abs/2601.11940},
}

Don’t argue with the trap. Leave it — and restart the thought.