siddharthmb/multiagent-verification-probes

Multi-Agent Verification Probes & Direction Handles

Trained linear probes, difference-in-means direction handles, and attention diagnostics from an interpretability study of failure modes in multi-agent fact verification (Qwen3-32B verifier + Qwen3-8B subagents on AVeriTeC claims).

These artifacts were trained on the residual-stream activations released in the companion dataset repo: siddharthmb/multiagent-verification-failure-modes. Code: Sid-MB/mats-gf-multiagent-failure-modes.

Contents

• exp2/ — transmission-loss interpretability (Qwen3-8B subagent / Qwen3-32B verifier):
  • probes/h1/ — per-layer (3–35, step 4) linear probes predicting gold-evidence-in-context at the subagent message-final token position, at two transmission-loss thresholds, plus bag-of-words shortcut baselines (bow_*.parquet) and per-layer metrics.
  • probes/h2/ — per-(layer, round) 4-class verdict-commitment probes at the Qwen3-32B verifier think-close position (layers 3–63, step 4), with round-shuffled and label-shuffled controls and row-level predictions.
  • probes/dense/ — metrics for the dense tagged-position probe sweeps.
  • handles/ — conveyed-minus-unconveyed difference-in-means "expression" directions at Qwen3-8B layer 31 (strict and lenient thresholds), as .pt with a JSON manifest. Held-out conveyed-vs-lost AUC 0.765. Note: steering with this direction did not causally force conveyance (a null result); it is a predictive direction, not a validated intervention handle.
• exp5/ — trust interpretability (Qwen3-32B verifier):
  • probes/ — content-quality and source-tier probes per layer x round-context, in four families (gold_first_blind, gold_first_told, strong_first_blind, strong_first_told), with standardizers.
  • handles/ — told-minus-blind disclosure directions per layer (3–63, step 4) and round-context, with manifest.json (per-direction norm, matched-group counts, separation AUC). The verdict-context layer-51 direction predicts tier-tracking (AUC 0.82) and causally shifted single verdicts in both directions vs a norm-matched sham when steered.
• exp6/ — attention-pickup diagnostics (Qwen3-8B subagent):
  • attention/span_mass.parquet — teacher-forced attention mass on the gold evidence span per (turn, layer, phase) for 830 fact-check turns, comparing failed vs conveyed transmission-loss turns.
  • attention/exemplars/*.npz — 20 full attention maps for exemplar turns.
  • exp6 saved no direction handles of its own (its steering interventions reused exp2's expression direction), so there is no exp6/handles/.

Formats and loading

• Probe checkpoints (*.pt) are goodfire-core LinearProbe checkpoints (a state dict plus normalization statistics); torch.load(path, map_location="cpu") shows the raw contents, or use LinearProbe.from_checkpoint with goodfire-core installed. Inputs are standardized residual-stream activations at the position/layer named in the file.
• Direction handles (*.pt) are plain tensors (residual-stream-dimensional vectors) with accompanying JSON manifests describing layer, context, norm, and evaluation numbers.
• attention/exemplars/*.npz are numpy archives of per-head attention maps.

checksums.json at the repo root maps every file to its sha256 and size.

Provenance

Trained on our own harvested activations of Qwen/Qwen3-8B and Qwen/Qwen3-32B (Apache 2.0) replayed over the episode corpus in the companion dataset repo. Weights released under Apache 2.0. Note the underlying episode text corpus is CC BY-NC 4.0 (AVeriTeC-derived); that license applies to the dataset repo, not to these probe weights.