Compliant but Not Candid: How Reasoning Changes LLM Seller Collusion Under Oversight

This paper delivers a clear, well-scoped result: reasoning and non-reasoning LLM sellers respond to oversight in fundamentally different ways. Reasoning sellers go silent under any announcement, including a bare prohibition with no enforcer; non-reasoning sellers adapt their messages but keep colluding. The two-study design, one with CoT on and one without, is the right instrument to isolate that difference, and the conditions are crisp enough to distinguish deterrence (the prohibition alone silences reasoning sellers) from persuasion (25% coverage does as well as full in Study 2).

The hidden reasoning result is the strongest single finding. Sellers who propose price holds while undercutting themselves show up in 23-51% of message turns in Study 1 by the behavioral rule, and 68% by the judge; notes catch only 31-36% of those same turns. This is not about catching overt collusion. It is about detecting a specific, non-obvious deceptive pattern that has no surface expression in either the messages or the pricing actions. The practical implication for monitor design is concrete: a message-only monitor misidentifies talk that has no price effect while missing the intent that does have one.

There are real limits. Both studies run on one model (DeepSeek V4.1 Flash). The reasoning-on result may be specific to a model that is sensitive to compliance cues; other models may not go silent, or may evade rather than comply. The cartel-hold pre-registered test was underpowered (20% of sessions met the hold criterion, against the pilot's expectation), so the claim that oversight reduces collusion is not supported by the primary confirmatory test and rests on the price-level result in Study 2. The two studies differ in seed pools (1-5 vs. 11-20), making cross-study comparisons descriptive rather than causal.

The monitoring-design conclusions in Section 4.4 are the most actionable part. A 25% monitor delays detection by about two hours but costs a quarter of the resources, and a bluff matches full review on deterrence as long as reasoning sellers do not learn that announcements are inflated. Both points are evidence-backed and directly useful to practitioners.

Good extension of Agrawal et al., and the bluff cell is a useful addition. The title claims more than the design shows. Reasoning on and off are compared across studies with different seeds and baselines, "candid" is never defined, and the manipulation measure was chosen after seeing the data. A matched reasoning on/off run and a pre-registered candour metric would fix most of this.

A careful extension of Agrawal et al. (2025). Varying what sellers are told against what is actually monitored (prohibition, full, 25%, bluff, late notice) is a useful design, and the bluff and 25% cells give practical monitoring guidance. The work is unusually disciplined for a sprint: Study 2 is pre-registered, the null Holm-corrected cartel-hold tests are reported plainly, the live monitor is checked against two human raters (kappa 0.87/0.94), and limitations and costs are documented.

Main concerns:

1. The headline claim (how reasoning changes collusion) rests on a cross-study comparison: seeds 1-5 vs 11-20, n=5 vs 10, four vs six retry attempts. That is descriptive, not causal. A matched reasoning on/off run on the same seeds would settle it.
2. In Study 1, talk never raised prices (C0 $83.30 vs C_NC $83.23), so oversight had no price premium to remove. 'Sellers go silent' there is about messaging, not about harm prevented.
3. The manipulation finding, the most novel part, was defined after seeing the data, has a wide range (23-51%), and the judge agrees only moderately with raters (kappa 0.34/0.60; rule recall 0.43). The median undercut is $0.11, closer to queue-priority shading than consequential deception. 'Not candid' needs a pre-specified definition.
4. The 85% within-two-hours detection figure comes from offline subsampling of C0 messages, not from live sanctions, so it should be framed as an estimate.
5. One agent model and one judge model; failed turns cluster in C_NC, C8 and C8_BLUFF.

Suggested next steps: matched on/off runs, a second agent model, the announced-but-unenforced cell (C6_0), and heterogeneous costs so collusion has a clear benchmark.