RUBICON

Do Language Models Know When They've Crossed the Point of No Return?

Srija Mukhopadhyay*  ·  Adnan Qidwai*  ·  Jaydeep Sen  ·  Manish Shrivastava

IIIT Hyderabad · IBM Software Innovation Lab · *equal contribution · EMNLP 2026, Budapest

The agent can still reach the exit.

Move your mouse over the grid. Tiles you pass turn into walls too.

Alea iacta est.

The die is cast. Julius Caesar, upon crossing the Rubicon, 49 BCE.

The point of no return never looks like one.

Got only 2 mins? Paper at a glance!

0–4%

Models can detect impossibility but usually do not admit it! When we ask "Is it still solvable?", models get it right up to 100% of the time. When we ask for the best next move to win in the same state, all models except GPT-5.4 report it only 0 to 4% of the time.

77%

Models override their own reasoning. In up to 77% of wrong answers, the reasoning trace already says the puzzle is unsolvable, but the model still gives a move.

80–100%

Models blame more recent actions. When asked which action caused the failure, 80 to 100% of wrong answers point to a later step than the real cause.

An example

Emergence of
a dead end

Mosaic Grid · a puzzle from the datasetStill winnable

The puzzles

Wrong moves can make
the task impossible

We curate out-of-distribution puzzle environments where every move carries an irreversible effect. Puzzles work well here because simple solvers give us the exact right answer for every step, and new game mechanics stop models from relying on memorized solutions. We also check every puzzle against greedy shortcut strategies, so that solving it needs real planning.

3puzzle games
300puzzles, 100 per game
2,100test questions
31layout templates
25greedy strategies every puzzle beats
11models tested

Findings

Models can detect impossibility but
usually do not admit it

01

The integration gap

We test the same state through two different prompts: "Is it still solvable?" vs "What is the best next move to win?". Except GPT-5.4, most models do well when asked explicitly but fail badly when asked for the next move. We tested each model on 300 unsolvable puzzles, with 100 from each game.

Prompt

"Is it still solvable?"

Share of unwinnable states each model correctly called unwinnable.

02

Self-override in model responses

Through model traces on questions where they gave the wrong answer, we discovered that models realise that a given state is unsolvable, but choose to give a move anyway, overriding their findings. They spend a lot of tokens in this dilemma: wrong answers from Qwen3.5-397B average 37,581 tokens.

■ knows it can't win■ gives a move anyway
03

Tracing failure to its cause

Given the initial state of the puzzle and a list of actions, we ask models to identify the action that first made the puzzle unsolvable. When they get it wrong, their answer usually lands on the step where the failure becomes obvious, not the action that actually caused it. Most models also get worse as the gap between the cause and the end of the trace grows.

Play

Can you spot the
point of no return?

These are puzzles from the dataset. Just like the models, you will not be told when a move makes the puzzle impossible. If you think it can't be won anymore, press the red button. You can also turn on the solver light to check as you go.

Be the model

Answer the same
questions

Each question is an item from the dataset, with the same starting board and list of moves the models got. Answer it, then compare with the eleven models’ overall accuracy for that task and game.

Results

All results

Accuracy (%) for all eleven models. Click a column to sort. The best score in each column is in bold.

Cite

BibTeX

@inproceedings{mukhopadhyay2026rubicon,
  title     = {{RUBICON}: Do Language Models Know When They've Crossed the Point of No Return?},
  author    = {Mukhopadhyay, Srija and Qidwai, Adnan and Sen, Jaydeep and Shrivastava, Manish},
  booktitle = {Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing},
  year      = {2026},
  address   = {Budapest, Hungary}
}