Skip to content
OpenTrain AIFor AI Companies
← Back to explorer

Closing the Activation-Cone Blind Spot: Response-Time Probing and Unified Defense

Subhadip Mitra · Jun 28, 2026 · Citations: 0

Automatic Metrics
Open arXiv Find Implementation RSS feed Shortlist (0)

How to use this page

Low trust

Use this as background context only. Do not make protocol decisions from this page alone.

Best use

Background context only

What to verify

Validate the evaluation procedure and quality controls in the full paper before operational use.

Evidence quality

Low

Derived from extracted protocol signals and abstract evidence.

Abstract

Inference-time safety methods for large language models have proliferated, yet no systematic comparison exists. We evaluate five defense paradigms (no defense, static steering, CAST, AlphaSteer, probe-gated) across seven instruction-tuned models (7-31B) and five attack types (GCG, AutoDAN, DeepInception, prefilling, intent laundering). Our central finding: prompt-time activation defenses are structurally blind to prefilling attacks. AlphaSteer achieves 0% attack success on GCG, AutoDAN, and intent laundering but 50% on prefilling. We prove a corollary: any defense that gates intervention on a single layer's activation alignment with a benign reference (cone, subspace, or null-space) is blind to attacks that craft activations to lie inside that reference, whether checked at prompt time or per token. As its constructive contrapositive we introduce response-time probing: a linear probe on the model's hidden state at the first generated tokens, with AUROC 0.97-1.00 across all seven models. Combined with a halt, it cuts prefilling attack success to 0/40 on every model with 0% benign false positives, outperforming Llama Guard 3. Cross-template generalisation depends on probe depth, so we scope the claim to the canonical prefilling-template family. Composing the response-halt with AlphaSteer's null-space steering gives an orthogonal split (the halt catches prefilling, AlphaSteer catches semantic attacks), reaching defense success 0.983 on Mistral and 0.994 on Llama and dominating both components. We further show MMLU fails to capture steering's true utility cost, which appears as behavioral hedging rather than factual loss, and that diverse negative training sets cut probe false positives from 80-100% to near zero. Code, attacks, per-sample results, and the judge prompt are released.

Abstract-only analysis — low confidence

All signals on this page are inferred from the abstract only and may be inaccurate. Do not use this page as a primary protocol reference.

  • This paper looks adjacent to evaluation work, but not like a strong protocol reference.
  • The available metadata is too thin to trust this as a primary source.
Human Eval Hub LLM-as-Judge Hub Pairwise Preference Hub

Should You Rely On This Paper?

This paper is adjacent to HFEPX scope and is best used for background context, not as a primary protocol reference.

Best use

Background context only

Use if you need

A benchmark-and-metrics comparison anchor.

Main weakness

This paper looks adjacent to evaluation work, but not like a strong protocol reference.

Trust level

Low

Usefulness score

5/100 • Low

Treat as adjacent context, not a core eval-method reference.

Human Feedback Signal

Not explicit in abstract metadata

Evaluation Signal

Detected

Usefulness for eval research

Adjacent candidate

Extraction confidence 45%

If you are doing eval pipeline work, start here:

Human Eval Hub LLM-as-Judge Hub Pairwise Preference Hub Tool-Use Eval Hub

What We Could Verify

These are the protocol signals we could actually recover from the available paper metadata. Use them to decide whether this paper is worth deeper reading.

Human Feedback Types

missing

None explicit

No explicit feedback protocol extracted.

"Inference-time safety methods for large language models have proliferated, yet no systematic comparison exists."

Evaluation Modes

partial

Automatic Metrics

Includes extracted eval setup.

"Inference-time safety methods for large language models have proliferated, yet no systematic comparison exists."

Quality Controls

missing

Not reported

No explicit QC controls found.

"Inference-time safety methods for large language models have proliferated, yet no systematic comparison exists."

Benchmarks / Datasets

partial

MMLU

Useful for quick benchmark comparison.

"We further show MMLU fails to capture steering's true utility cost, which appears as behavioral hedging rather than factual loss, and that diverse negative training sets cut probe false positives from 80-100% to near zero."

Reported Metrics

partial

Auroc

Useful for evaluation criteria comparison.

"As its constructive contrapositive we introduce response-time probing: a linear probe on the model's hidden state at the first generated tokens, with AUROC 0.97-1.00 across all seven models."

Human Feedback Details

  • Uses human feedback: No
  • Feedback types: None
  • Rater population: Not reported
  • Expertise required: Coding

Evaluation Details

  • Evaluation modes: Automatic Metrics
  • Agentic eval: None
  • Quality controls: Not reported
  • Evidence quality: Low
  • Use this page as: Background context only

Protocol And Measurement Signals

Benchmarks / Datasets

MMLU

Reported Metrics

auroc

Research Brief

Metadata summary

Inference-time safety methods for large language models have proliferated, yet no systematic comparison exists.

Based on abstract + metadata only. Check the source paper before making high-confidence protocol decisions.

Key Takeaways

  • Inference-time safety methods for large language models have proliferated, yet no systematic comparison exists.
  • We evaluate five defense paradigms (no defense, static steering, CAST, AlphaSteer, probe-gated) across seven instruction-tuned models (7-31B) and five attack types (GCG, AutoDAN, DeepInception, prefilling, intent laundering).
  • Our central finding: prompt-time activation defenses are structurally blind to prefilling attacks.

Researcher Actions

  • Compare this paper against others mentioning MMLU.
  • Check the full text for explicit evaluation design choices (raters, protocol, and metrics).
  • Use related-paper links to find stronger protocol-specific references.

Caveats

  • Generated from abstract + metadata only; no PDF parsing.
  • Signals below are heuristic and may miss details reported outside the abstract.

Recommended Queries

MMLU

Research Summary

Contribution Summary

  • Inference-time safety methods for large language models have proliferated, yet no systematic comparison exists.
  • We evaluate five defense paradigms (no defense, static steering, CAST, AlphaSteer, probe-gated) across seven instruction-tuned models (7-31B) and five attack types (GCG, AutoDAN, DeepInception, prefilling, intent laundering).
  • As its constructive contrapositive we introduce response-time probing: a linear probe on the model's hidden state at the first generated tokens, with AUROC 0.97-1.00 across all seven models.

Why It Matters For Eval

  • Inference-time safety methods for large language models have proliferated, yet no systematic comparison exists.
  • Code, attacks, per-sample results, and the judge prompt are released.

Researcher Checklist

  • Gap: Human feedback protocol is explicit

    No explicit human feedback protocol detected.

  • Pass: Evaluation mode is explicit

    Detected: Automatic Metrics

  • Gap: Quality control reporting appears

    No calibration/adjudication/IAA control explicitly detected.

  • Pass: Benchmark or dataset anchors are present

    Detected: MMLU

  • Pass: Metric reporting is present

    Detected: auroc

Related Papers

Papers are ranked by protocol overlap, extraction signal alignment, and semantic proximity.