Pith. sign in

Researcher Evidence Record

Fabien Roger

This bounded record lists 19 Pith paper rows and 0 imported work rows attributed to this corpus identity. The enumerated, non-disputed paper rows include cs.LG, cs.AI, cs.CL work dated 2022 to 2026. The record describes sources and coverage; it makes no judgment about the person.

Compiled coverage vector

Measured lane counts only. Not a trust score or person verdict.

Identity 2 canonical identity row plus typed public identifiers Pith papers 19 attributed paper rows in the Pith corpus, itemized below Imported works 0 imported works beyond the paper ledger Current reviews 11 of 19 papers carrying a current Pith review; presence, not outcome Sourced citations 12 of 19 works with a source-named citation count Coauthors 50 distinct coauthors across shared paper rows
Enumerated paper scope: 4 fields (cs.LG, cs.AI, cs.CL, +1 more) · 2022-2026 sources: authors, author_identifiers · paper_authors · author_works · current_verdicts · cited_works

A sourced case file for attributed work. It is neither a profile score nor a verdict about this researcher.

Attributed works

A bounded ledger from the Pith paper and imported-work queries. Counts and source confidence stay with each work.

  1. 2026 Pith paper

    Overthinking: Amplifying Reasoning Weights to Extract Learned Secrets

    cs.AI provisional current review present

    paper paper evidence challenge this paper

    Sources and evidence
    Authorship source
    arxiv_oai
    Printed name
    Fabien Roger
    Author position
    3
    Identity state
    provisional
    Source confidence
    0.7
    Review coverage
    Measured: a current Pith review exists.
    Citation counts
    No source count is attached to this work row.
  2. 2026 Pith paper

    (Mis)generalization of Helpful-only Fine-tuning

    cs.LG provisional current review present

    paper paper evidence challenge this paper

    Sources and evidence
    Authorship source
    arxiv_oai
    Printed name
    Fabien Roger
    Author position
    3
    Identity state
    provisional
    Source confidence
    0.7
    Review coverage
    Measured: a current Pith review exists.
    Citation counts
    No source count is attached to this work row.
  3. 2026 Pith paper

    SLEIGHT-Bench: A Benchmark of Evasion Attacks Against Agent Monitors

    cs.CR provisional current review present

    paper paper evidence challenge this paper

    Sources and evidence
    Authorship source
    arxiv_oai
    Printed name
    Fabien Roger
    Author position
    4
    Identity state
    provisional
    Source confidence
    0.7
    Review coverage
    Measured: a current Pith review exists.
    Citation counts
    No source count is attached to this work row.
  4. 2026 Pith paper

    Classifier Context Rot: Monitor Performance Degrades with Context Length

    cs.AI provisional current review present

    paper citation record paper evidence challenge this paper

    Sources and evidence
    Authorship source
    backfill
    Printed name
    Fabien Roger
    Author position
    2
    Identity state
    provisional
    Source confidence
    0.7
    Review coverage
    Measured: a current Pith review exists.
    Citation counts
    • 1 pith inbound references from cited_work_pith_inbound_counts
  5. 2026 Pith paper

    How Useful Is Cross-Domain Generalization for Training LLM Monitors?

    cs.AI provisional current review present

    paper paper evidence challenge this paper

    Sources and evidence
    Authorship source
    backfill
    Printed name
    Fabien Roger
    Author position
    2
    Identity state
    provisional
    Source confidence
    0.7
    Review coverage
    Measured: a current Pith review exists.
    Citation counts
    No source count is attached to this work row.
  6. 2026 Pith paper

    Narrow Secret Loyalty Dodges Black-Box Audits

    cs.CR provisional current review present

    paper paper evidence challenge this paper

    Sources and evidence
    Authorship source
    arxiv_oai
    Printed name
    Fabien Roger
    Author position
    2
    Identity state
    provisional
    Source confidence
    0.7
    Review coverage
    Measured: a current Pith review exists.
    Citation counts
    No source count is attached to this work row.
  7. 2025 Pith paper

    Chain of Thought Monitorability: A New and Fragile Opportunity for AI Safety

    cs.AI provisional current review present

    paper citation record paper evidence challenge this paper

    Sources and evidence
    Authorship source
    arxiv_oai
    Printed name
    Fabien Roger
    Author position
    32
    Identity state
    provisional
    Source confidence
    0.7
    Review coverage
    Measured: a current Pith review exists.
    Citation counts
    • 2 external cited by from pith
    • 49 pith inbound references from cited_work_pith_inbound_counts
  8. 2025 Pith paper

    Why Do Some Language Models Fake Alignment While Others Don't?

    cs.LG provisional current review present

    paper citation record paper evidence challenge this paper

    Sources and evidence
    Authorship source
    arxiv_oai
    Printed name
    Fabien Roger
    Author position
    7
    Identity state
    provisional
    Source confidence
    0.7
    Review coverage
    Measured: a current Pith review exists.
    Citation counts
    • 5 pith inbound references from cited_work_pith_inbound_counts
  9. 2025 Pith paper

    Reasoning Models Don't Always Say What They Think

    cs.CL provisional current review present

    paper citation record paper evidence challenge this paper

    Sources and evidence
    Authorship source
    arxiv_oai
    Printed name
    Fabien Roger
    Author position
    10
    Identity state
    provisional
    Source confidence
    0.7
    Review coverage
    Measured: a current Pith review exists.
    Citation counts
    • 12 external cited by from pith
    • 64 pith inbound references from cited_work_pith_inbound_counts
  10. 2025 Pith paper

    Auditing language models for hidden objectives

    cs.AI provisional measured, no current review

    paper citation record paper evidence challenge this paper

    Sources and evidence
    Authorship source
    arxiv_oai
    Printed name
    Fabien Roger
    Author position
    26
    Identity state
    provisional
    Source confidence
    0.7
    Review coverage
    Measured: no current Pith review exists.
    Citation counts
    • 2 external cited by from arxiv_reference
    • 17 pith inbound references from cited_work_pith_inbound_counts
  11. 2025 Pith paper

    A Frontier AI Risk Management Framework: Bridging the Gap Between Current AI Practices and Established Risk Management

    cs.AI provisional current review present

    paper citation record paper evidence challenge this paper

    Sources and evidence
    Authorship source
    arxiv_oai
    Printed name
    Fabien Roger
    Author position
    3
    Identity state
    provisional
    Source confidence
    0.7
    Review coverage
    Measured: a current Pith review exists.
    Citation counts
    • 2 pith inbound references from cited_work_pith_inbound_counts
  12. 2024 Pith paper

    Alignment faking in large language models

    cs.AI provisional current review present

    paper citation record paper evidence challenge this paper

    Sources and evidence
    Authorship source
    arxiv_oai
    Printed name
    Fabien Roger
    Author position
    4
    Identity state
    provisional
    Source confidence
    0.7
    Review coverage
    Measured: a current Pith review exists.
    Citation counts
    • 20 external cited by from pith
    • 77 pith inbound references from cited_work_pith_inbound_counts
  13. 2024 Pith paper

    Do Unlearning Methods Remove Information from Language Model Weights?

    cs.LG provisional measured, no current review

    paper citation record paper evidence challenge this paper

    Sources and evidence
    Authorship source
    arxiv_oai
    Printed name
    Fabien Roger
    Author position
    2
    Identity state
    provisional
    Source confidence
    0.7
    Review coverage
    Measured: no current Pith review exists.
    Citation counts
    • 10 pith inbound references from cited_work_pith_inbound_counts
  14. 2024 Pith paper

    Stress-Testing Capability Elicitation With Password-Locked Models

    cs.LG provisional measured, no current review

    paper citation record paper evidence challenge this paper

    Sources and evidence
    Authorship source
    arxiv_oai
    Printed name
    Fabien Roger
    Author position
    2
    Identity state
    provisional
    Source confidence
    0.7
    Review coverage
    Measured: no current Pith review exists.
    Citation counts
    • 6 pith inbound references from cited_work_pith_inbound_counts
  15. 2023 Pith paper

    AI Control: Improving Safety Despite Intentional Subversion

    cs.LG provisional measured, no current review

    paper citation record paper evidence challenge this paper

    Sources and evidence
    Authorship source
    arxiv_oai
    Printed name
    Fabien Roger
    Author position
    4
    Identity state
    provisional
    Source confidence
    0.7
    Review coverage
    Measured: no current Pith review exists.
    Citation counts
    • 3 external cited by from pith
    • 23 pith inbound references from cited_work_pith_inbound_counts
  16. 2023 Pith paper

    Preventing Language Models From Hiding Their Reasoning

    cs.LG provisional measured, no current review

    paper citation record paper evidence challenge this paper

    Sources and evidence
    Authorship source
    arxiv_oai
    Printed name
    Fabien Roger
    Author position
    1
    Identity state
    provisional
    Source confidence
    0.7
    Review coverage
    Measured: no current Pith review exists.
    Citation counts
    • 2 external cited by from arxiv_reference
    • 10 pith inbound references from cited_work_pith_inbound_counts
  17. 2023 Pith paper

    Benchmarks for Detecting Measurement Tampering

    cs.LG provisional measured, no current review

    paper citation record paper evidence challenge this paper

    Sources and evidence
    Authorship source
    arxiv_oai
    Printed name
    Fabien Roger
    Author position
    1
    Identity state
    provisional
    Source confidence
    0.7
    Review coverage
    Measured: no current Pith review exists.
    Citation counts
    • 1 pith inbound references from cited_work_pith_inbound_counts
  18. 2023 Pith paper

    Large Language Models Sometimes Generate Purely Negatively-Reinforced Text

    cs.LG provisional measured, no current review

    paper paper evidence challenge this paper

    Sources and evidence
    Authorship source
    arxiv_oai
    Printed name
    Fabien Roger
    Author position
    1
    Identity state
    provisional
    Source confidence
    0.7
    Review coverage
    Measured: no current Pith review exists.
    Citation counts
    No source count is attached to this work row.
  19. 2022 Pith paper

    Language models are better than humans at next-token prediction

    cs.CL provisional measured, no current review

    paper paper evidence challenge this paper

    Sources and evidence
    Authorship source
    arxiv_oai
    Printed name
    Fabien Roger
    Author position
    2
    Identity state
    provisional
    Source confidence
    0.7
    Review coverage
    Measured: no current Pith review exists.
    Citation counts
    No source count is attached to this work row.

Evidence apparatus

The machinery behind this record. Every lane states whether Pith measured it, did not query it, could not reach it, or withheld it.

LaneStateObservedBoundary and source
identity Measured 2 Canonical identity row plus public typed identifiers.
source=authors, author_identifiers
papers Measured 19 of 19 bounded rows Rows attributed to this author UUID in the Pith corpus.
source=paper_authors
works Measured zero 0 of 0 bounded rows Imported works not duplicated by the paper ledger.
source=author_works
reviews Measured 11 of 19 bounded rows Coverage count only. No review outcome is projected onto the person.
source=current_verdicts
citations Measured 12 of 19 bounded rows Counts remain itemized by work and source.
source=cited_works
coauthors Measured 50 of 19 bounded rows Shared-work edges from admitted paper rows.
source=paper_authors
account Unavailable No public count of 1 bounded rows Account metadata is separate from corpus evidence.
source=users.author_id
Public identity sources
  • name variant
    Fabien Roger     backfill
    confidence 0.6

identity history

Enumerated research scope

Fields and dates come only from enumerated, non-disputed Pith paper rows. They do not claim career completeness.

  • cs.LG8 rows
  • cs.AI7 rows
  • cs.CL2 rows
  • cs.CR2 rows
  • 20221 rows
  • 20234 rows
  • 20243 rows
  • 20255 rows
  • 20266 rows
Shared-work index
Record scope

The work queries are bounded. Missing rows may mean measured zero, an unavailable source, a query that did not run, or private data that Pith withheld. The lane table keeps those cases separate.

Paper findings remain attached to papers. They do not become findings about this researcher.