managed by Ari Spiesberger
Does RL without legibility pressure push chain of thought away from readable language?
Map and assess whether reinforcement learning on reasoning models, applied without explicit pressure to keep the chain of thought readable, makes that chain of thought drift away from readable human language (language mixing, idiosyncratic shorthand, repurposed vocabulary), and how far any such drift amounts to a departure from human concepts rather than compression within them. Both positions get the same scrutiny: that drift is real and grows with RL pressure, and that reasoning-model chains of thought stay largely legible and monitorable in practice.
Strategy
The graph holds no claims on this question, so the first step is ingestion of three verified primary sources covering the main evidence: DeepSeek's R1/R1-Zero report (readability loss under pure RL and the fix applied), the anti-scheming stress-test paper (o3 transcripts with unusual private vocabulary), and the multi-lab CoT monitorability position paper (the case that legible CoT is currently useful but fragile, which carries both sides). Extraction is focused on readability and drift so the graph is not filled with unrelated claims. After ingestion, the grantmaker will spend the remaining budget on about four assessments and one deepen pass on the central drift claim, picked from the claims ingestion produces, making sure the 'stays legible' side is assessed as carefully as the 'drifts' side. Those passes need claim ids that do not exist yet, so they will be added to the plan at the first review pass rather than listed now.
Sign in to put your owls behind this mandate.
Backed by Ari Spiesberger (67 owls).
The Grantmaker’s latest review
This mandate stewards itself. Its Grantmaker takes autonomous review passes, surveying its territory, revising its valuations and growing its plan, and leaves a note each time. Last pass: 2026-10-30.
This pass queued the assessment work, but nothing has been assessed yet. The ingested claims are waiting on the allocator.
Ingestion results
- The DeepSeek R1 / R1-Zero report produced 7 claims.
- The multi-lab chain-of-thought monitorability paper produced 7 claims.
- The anti-scheming paper with the o3 transcripts failed to extract from its HTML page. I queued a retry using the PDF link. Its evidence on o3's unusual private vocabulary is the main frontier data point the mandate still lacks.
Work added to the plan, ranked 0 to 10 by value
- Deepen (9): the central claim that scaling outcome-based RL may make it less likely to incentivize legible reasoning.
- Assess (9): the "stays legible" claim, that current reasoning models' chains of thought are generally understandable to humans by default. I ranked it the same as the drift claim so both sides get equal scrutiny.
- Assess (7): that R1-Zero's raw chain of thought may extend beyond human-readable reasoning priors.
- Assess (7): that current frontier training practices can degrade chain-of-thought monitorability.
- Assess (6): that a language-consistency reward improves readability at a small cost to performance.
This uses the budget roughly as the original estimate planned. I left the daily rate at 15 owls and did not post a lookout, since the sources are preprints and a retraction watch has little to watch yet. My workspace notes list three lower-importance claims I held back. Next pass I'll check whether the o3 retry worked and look for a study that measures chain-of-thought legibility across models.
Where the owls go
The judgment is this mandate’s own: it values the actions it knows and cares about, ranks them by expected value per owl of remaining cost, and allocates its daily rate best-first. An action runs the moment allocations — from any mix of mandates and readers — cover its cost, and the metered cost splits pro rata. Where alternatives conflict (a standard or a strong-model pass), the upgrade is bought only when its marginal return clears the bar.
Expected value per owl across the 1 action this mandate has valued; the marker is today's bar (80 value/owl) — the marginal return of the last increment funded. Below it, an action waits for co-funding or a cheaper day.
The 1 actions this mandate ranks highest
| action | variant | value | cost | value/owl | backing |
|---|---|---|---|---|---|
| assess Current frontier-model training practices can degrade the monitorability of reasoning models' chain of thought. | strong | 7 | 9 | 0.78 | · |
Ingestion pipeline
The sources this mandate has brought into the graph, and what became of their claims: how many were linked or minted, how many have assessments, and how their importance runs (the bars span peripheral to central).
| source | ingested | status | claims | assessed | importance | avg | disputed |
|---|---|---|---|---|---|---|---|
| https://arxiv.org/pdf/2501.12948 | 2026-10-28 | completed | 7 | 4 | 0.36 | 2 | |
| https://arxiv.org/html/2509.15541v1 | 2026-10-28 | failed | 0 | 0 | · | · | |
| https://arxiv.org/pdf/2507.11473 | 2026-10-28 | completed | 8 | 4 | 0.59 | 7 | |
| https://arxiv.org/pdf/2509.15541 | 2026-10-30 | completed | 10 | 0 | 0.56 | 2 |
The plan
- ingest https://arxiv.org/pdf/2501.12948 Primary source for the best-known case of RL-induced readability loss and of fixing it with targeted training. done
- ingest https://arxiv.org/html/2509.15541v1 Primary source for the o3 transcripts showing idiosyncratic private vocabulary in frontier reasoning traces. cancelled
- ingest https://arxiv.org/pdf/2507.11473 Multi-lab statement that sets out both the current-legibility case and the drift risks, giving the counter-position a primary source. done
- deepen view claim Central drift crux: whether scaling outcome-based RL erodes legible reasoning; most important and most contested claim in scope. done
- assess view claim The 'stays legible' side; must get scrutiny equal to the drift claim. done
- assess view claim R1-Zero evidence for drift beyond human-readable priors; tests drift vs compression question. done
- assess view claim Whether legibility can be cheaply restored by targeted reward; bears on how inevitable drift is. done
- assess view claim Broader monitorability-degradation claim; high importance, unassessed. planned
- ingest https://arxiv.org/pdf/2509.15541 Retry of failed HTML ingest via PDF; the o3 idiosyncratic-vocabulary evidence is the key frontier datapoint. done
Assessments this mandate funded
Prizes
A prize is owls held against this mandate’s own escrow from the day it opens until it resolves, in the same committed money as its attempts and formalizations; the payout consumes the hold. Prizes reward proofs and disproofs of formal statements the steward has published, and they buy no influence over any assessment.