Skip to content

Latest commit

 

History

1 Commit

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Axon — Research Workspace

This workspace is for the second paper in the line that started with PROBE.

Where PROBE proposes a fix (an explicit, revisable belief so a frozen model stops confabulating), Axon proposes the diagnostic: a way to measure where a model's reasoning stops holding together and what the failure looks like right at that edge.

The two papers are meant to be read as a pair. Axon measures the boundary; PROBE gives the model a way to act safely near it.

Working Research Direction

Current working direction (working thesis, not the final title). Updated 2026-07-21 after the deep literature sweep — see notes/DEEP_RESEARCH_2026-07-21.md:

Detecting the confident-wrong regime under rule shift, where standard uncertainty methods fail — as the trigger for PROBE

The pivot in plain terms: existing tools (semantic entropy, self-consistency) catch a model when it is unsure. They break when the model is confidently wrong — it repeats the same wrong answer, so it looks certain. That failure is worst exactly when the rules of the task CHANGE: the model keeps confidently applying a rule that just became false. Axon detects that "confidently wrong after a change" moment early, before the model acts on the stale belief. That is precisely the moment PROBE is built to fix.

(The earlier framing — "branch until the tree stops reconverging" — turned out to be largely published already, on static tasks. See DEEP_RESEARCH and RELATED_WORK. Branching / divergence is kept as the mechanism, but the novel target is the confident-wrong, rule-shift regime.)

Project Name

  • Repo / codename: axon
  • System name: Axon
  • One-line pitch: a step-level detector that fires when a frozen model becomes confidently wrong after the rules change — where semantic entropy and self-consistency go blind.

Contributions (locked — detail in notes/CONTRIBUTIONS.md)

  • C-A — Confident-wrong detector under rule shift (headline). A zero-call trace signal that is informative under broad shifts, while its G0 component can invert under partial shifts; this boundary is part of the result, alongside the comparison with semantic entropy and self-consistency.
  • C-B — Early warning / lead time (support). The signal fires before the post-shift accuracy collapse, with measurable lead time. (Salvages the old early-warning idea, now on the novel regime.)
  • C-C — Cross-scale (support). The blind spot and the Axon fix both hold across 2-3 model sizes, so it is not a weak-model artifact.
  • C-D — Pressure cooker (optional robustness). Sealed multi-model debate as a second, independent way to surface divergence; framed around "debate collapses to a confident majority." Cut first if over budget.
  • Old branch-structure idea (C1) is demoted to a mechanism section, not a headline.

Derived novelty claims (grown from the pivot, novelty-checked 2026-07-21)

  • N1 — Active probe beats passive monitoring (strongest new claim). One targeted counterfactual probe step ("if the old rule still held, what would happen?") detects the confident-wrong crossing earlier than any passive signal. This operationalizes the word "probe" and is the tightest unification with PROBE. Comparative result confirmed missing from the literature.
  • N2 — The cling curve. After a rule shift, accuracy drops immediately but confidence lags; measure the gap as a scalar "cling time." It is PROBE's motivating failure turned into a number, and it predicts how much PROBE helps per task.
  • N3 — (stretch) Detectability boundary. Some rule shifts are intrinsically undetectable early (when the new rule agrees with the old on recent observations); characterize lead time vs old/new rule overlap. An honest limit result.

Paper in one sentence: existing uncertainty methods detect an unsure model but fail on a confidently wrong one; Axon detects the confident-wrong crossing under rule shift — early, actively (N1), and measurably (N2) — giving PROBE the trigger it needs.

Why This Is Worth Doing (and is still open)

Confirmed by the deep sweep (notes/DEEP_RESEARCH_2026-07-21.md). What is already settled (do NOT claim): reasoning collapses past a complexity threshold; semantic entropy / self-consistency signal uncertainty; error can be predicted early from the trace on static tasks (AUROC ~0.80). What is still OPEN (Axon's target):

  1. The confident-wrong blind spot. Self-consistency and semantic entropy collapse when the model repeats the same wrong answer. Cross-model disagreement is the only proposed fix, and it is new and narrow.
  2. Non-stationarity. The collapse / uncertainty literature uses static math/QA. Almost nobody studies early detection under a RULE SHIFT. STALE (2026) shows even the best model scores only 55.2% at noticing its belief is stale — open ground.
  3. Detector-to-agent loop. No prior work pairs a boundary detector with a belief-revision agent (PROBE) on the same non-stationary tasks.

Budget Philosophy (important, this is a student project)

Frontier models and GPU clusters are NOT required, and are arguably the wrong tool.

  • Small open models (7B-8B: Llama 3.1 8B, Qwen, Mistral) break sooner and more visibly, so the whole collapse boundary can be mapped cheaply.
  • Use 2-3 model sizes to get the "boundary scales with capacity, failure mode stays the same" story.
  • A single frontier spot-check (via OpenRouter, on the best result only) is a confirmation at the very end, not the experiment.

The honest framing is a strength: "we map the full boundary on small models where it is reachable, and confirm the pattern persists on a frontier model."

Relationship to PROBE

  • PROBE = the fix (explicit belief + contradiction detection + surgical revision).
  • Axon = the diagnostic (detect the distribution boundary / collapse onset).
  • Shared thesis: models fail on novel inputs because they cannot detect their own distribution boundary. Axon measures where that boundary is; PROBE lets the model act near it without confabulating.

Status

Stage 4 pilot complete on Llama 3.1 8B. The active stale probe reached AUROC 0.950, 95 percent CI [0.875, 1.000], on shift steps 10 through 14. It beat reversed agreement by 0.150 AUROC, 95 percent CI [0.009, 0.340]. Caveat: the gap over cling timing was not clean, 0.200 with 95 percent CI [-0.018, 0.485], so cling timing remains a serious baseline.

About

Research workspace for detecting confident-wrong reasoning under rule shifts in language models.

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages