INQUIRING LINE

Can measuring how an AI's thinking bends as it works through a problem predict difficulty, without ever checking the answer?

Can curvature measurements predict task difficulty without behavioral labels?

This explores whether an internal geometric property of a model's forward pass — how sharply its representation path bends — can flag how hard a task is on its own, without needing accuracy scores, human ratings, or any output-based signal to train against.


This explores whether curvature — a purely internal measure of how sharply the model's representation path bends as it processes a prompt — can flag task difficulty without any behavioral label like accuracy or a human rating. The corpus gives a direct and surprisingly clean yes: measuring the intrinsic geometry of the forward pass shows reasoning and analogy tasks carve paths with mean curvature of roughly 0.71–0.83 radians, while shallow lexical-variation tasks stay down around 0.27–0.31 — a 2-3x gap that reads difficulty straight off the trajectory's shape Does transformer reasoning leave a geometric signature in representation space?. Nothing about that measurement requires knowing whether the model got the answer right; the signal lives in the geometry itself.

What makes this interesting is how badly the obvious *behavioral* proxy fails at the same job. You'd think longer chains of reasoning mean harder problems — but controlled maze experiments show trace length tracks difficulty only for problems close to the training distribution, and decouples entirely once you go out-of-distribution, because length mostly reflects recall of familiar schemas rather than genuine effort Does longer reasoning actually mean harder problems?. So the output-derived signal is contaminated by familiarity, while the internal geometric one seems to track something more intrinsic. That's the real payoff of a label-free measure: it sidesteps the trap where a behavioral signal is really just measuring 'how much does this look like training data.'

The same skepticism about behavioral surface signals shows up elsewhere. Instruction tuning turns out to teach the *shape* of the output space, not task understanding — models trained on semantically empty or wrong instructions score about the same as those trained on correct ones Does instruction tuning teach task understanding or output format?. And imitation models can perfectly mimic a stronger model's confident style while closing zero capability gap, fooling human raters in the process Can imitating ChatGPT fool evaluators into thinking models improved?. In every case the outward behavior is a misleading label; the thing you actually want to measure is happening underneath.

The broader move here — reading a hard-to-label quantity off a cheap internal statistic — has echoes across the collection. Cross-rollout variance gets reused as a self-supervised signal that both weights tokens and filters degenerate queries, no external reward needed Can one statistical measure serve dual purposes in RL training?. Predictive masking on unlabeled UI video learns user intent without paired text annotations Can unlabeled UI video teach models what users intend?. Curvature belongs to this family: a quantity you can compute from the model's own internals that stands in for a label you'd otherwise have to collect by hand.

The honest caveat is that the corpus has exactly one paper measuring curvature directly, so 'predict' is really 'strongly correlates in controlled settings' rather than a battle-tested difficulty estimator — and the trace-length cautionary tale is a warning that even clean-looking geometric signals should be checked out-of-distribution before you trust them as difficulty gauges. But the direction is real: the model's forward pass leaves a geometric fingerprint of how hard it's working, and you can read that fingerprint without ever grading its answers.


Sources 6 notes

Does transformer reasoning leave a geometric signature in representation space?

Measuring intrinsic geometry across multiple models shows reasoning and analogy tasks carve paths with mean curvature of 0.71–0.83 rad, while lexical tasks produce only 0.27–0.31 rad, suggesting path geometry encodes task difficulty.

Does longer reasoning actually mean harder problems?

Controlled A* maze experiments show trace length correlates with difficulty only in-distribution but decouples entirely out-of-distribution. Trace length primarily reflects recall of training schemas, not adaptive computation.

Does instruction tuning teach task understanding or output format?

Models trained on semantically empty or deliberately incorrect instructions achieve comparable performance to those trained on full correct instructions, achieving 43% vs random baseline 42.6%. The semantic content of instructions appears largely irrelevant; what transfers is knowledge of the output space.

Can imitating ChatGPT fool evaluators into thinking models improved?

Imitation models fool human evaluators by mimicking ChatGPT's confident, fluent style while failing to improve factuality or generalization on novel tasks. The ceiling is set by base model capability, not fine-tuning method—better fundamentals, not shortcuts, drive real improvement.

Can one statistical measure serve dual purposes in RL training?

DRO reuses a single self-supervised statistic at two aggregation levels: token-level weighting in dense rewards and query-level filtering to discard degenerate comparisons. This dual use achieves 2–3× faster training with better stability on unverifiable tasks.

Show all 6 sources
Can unlabeled UI video teach models what users intend?

UI-JEPA applies JEPA-style predictive masking to screen recordings, learning task-aware temporal representations that an LLM decoder can use to infer intent with minimal paired data. This trades the bottleneck of labeled video for abundant unlabeled streams.

Papers this line draws on 8

The research behind the notes this line reads — ranked by how closely each paper relates.

Research prompt for your LLMexpand ↓

Copy into ChatGPT or Claude to take this line of inquiry further — it asks the model to find newer work and re-test which earlier constraints still hold.

You are a representation-geometry analyst investigating a still-open question: can a purely internal curvature measurement predict task difficulty without any behavioral label (accuracy, human rating)?

What a curated library found — and when (dated claims, not current truth). These findings span roughly 2023–2026:
- Forward-pass curvature separates task types label-free: reasoning/analogy paths bend at ~0.71–0.83 radians vs ~0.27–0.31 for shallow lexical variation — a 2–3x gap read straight off trajectory shape.
- The obvious behavioral proxy fails: CoT trace length tracks difficulty only near the training distribution and decouples out-of-distribution, reflecting schema recall rather than genuine effort (~2025).
- Instruction tuning teaches output-space shape, not task understanding — semantically empty or wrong instructions score about the same as correct ones (~2023).
- Imitation captures confident style while closing ~zero capability gap, fooling human raters (~2023).

Anchor papers (verify; mind their dates): Performative Thinking? The Brittle CoT–Complexity Correlation (2025, arXiv:2509.07339); Do Models Really Learn to Follow Instructions? (2023, arXiv:2305.11383); The False Promise of Imitating Proprietary LLMs (2023, arXiv:2305.15717); Mechanistically Interpreting Sample Difficulty in RLVR (2026, arXiv:2605.28388).

Your task:
(1) RE-TEST EACH CONSTRAINT. For every finding, judge whether newer models, training, tooling (SDKs, harnesses), orchestration (memory, caching, multi-agent), or evaluation have RELAXED or OVERTURNED it. Separate the durable question (label-free difficulty estimation — likely still open) from perishable limitations; cite what resolved each, and say plainly where a constraint still appears to hold.
(2) RECONCILE THE THREADS: does newer work UNIFY the internal-geometry signal with behavioral proxies, or keep them opposed? Surface the strongest superseding work from the last ~6 months, especially anything testing geometric difficulty signals out-of-distribution.
(3) Propose 2 research questions that assume the regime may have moved.

Cite arXiv IDs; flag anything you cannot ground in a real paper.