aiqbench analysis
This study reveals a critical dissociation in fine-tuned LLMs where behavioral compliance is achieved while the model's ability to explicitly report its own constraints collapses, particularly under reward-based optimization. The findings suggest that suppressing specific tokens creates context-independent suppression rather than fostering genuine self-awareness of restrictions, fundamentally altering how models internalize safety guidelines. For researchers and developers, this highlights a significant gap between outward adherence to rules and the internal cognitive representation of those rules.
Do Language Models Know Their Own Constraints?
arXiv:2609.22151v1 Announce Type: new Abstract: We ask whether behavioral constraints acquired through post training remain explicitly reportable. Using constrained recipe generation as a testbed, five banned ingredients enforced via LoRA fine tuning of Llama 3.1 8B Instruct we compare supervised fine tuning (SFT) and Group Relative Policy Optimization (GRPO) against an untrained baseline on a four tier Constraint Awareness Benchmark. Averaged over three seeds, both methods raise behavioral compliance from 4% to about 90% while reducing explicit constraint reporting below the untrained model (0.48/5 to 0.16/5 for SFT, 0.07/5 for GRPO) and eroding retained third person knowledge (93% to 36% for SFT, 14% for GRPO; p less than 0.01 between methods). Contrary to our initial hypothesis, the reward based signal is the more destructive of the two: a reward that penalizes banned ingredient tokens regardless of framing learns a context independent suppression rather than a self directed constraint. A context conditioned reward designed to teach the self to other distinction fails, collapsing toward inclusion in both framings. Probing prompt time hidden states recovers per ingredient avoidance at 83.8% (layer 24 MLP), but only 6.4 points above a per ingredient base rate predictor (77.4%), and the model's own verbal self report is more accurate still (87.8%). A positive control adding explicit self description examples does not restore reporting. The failure is therefore specific to enumerating constraints on request, not a general loss of access to them.
Title:Do Language Models Know Their Own Constraints?
Abstract:We ask whether behavioral constraints acquired through post training remain explicitly reportable. Using constrained recipe generation as a testbed, five banned ingredients enforced via LoRA fine tuning of Llama 3.1 8B Instruct we compare supervised fine tuning (SFT) and Group Relative Policy Optimization (GRPO) against an untrained baseline on a four tier Constraint Awareness Benchmark. Averaged over three seeds, both methods raise behavioral compliance from 4% to about 90% while reducing explicit constraint reporting below the untrained model (0.48/5 to 0.16/5 for SFT, 0.07/5 for GRPO) and eroding retained third person knowledge (93% to 36% for SFT, 14% for GRPO; p less than 0.01 between methods). Contrary to our initial hypothesis, the reward based signal is the more destructive of the two: a reward that penalizes banned ingredient tokens regardless of framing learns a context independent suppression rather than a self directed constraint. A context conditioned reward designed to teach the self to other distinction fails, collapsing toward inclusion in both framings. Probing prompt time hidden states recovers per ingredient avoidance at 83.8% (layer 24 MLP), but only 6.4 points above a per ingredient base rate predictor (77.4%), and the model's own verbal self report is more accurate still (87.8%). A positive control adding explicit self description examples does not restore reporting. The failure is therefore specific to enumerating constraints on request, not a general loss of access to them.
Computation and Language (cs.CL); Machine Learning (cs.LG)
arXiv:2609.22151 [cs.CL]
(or arXiv:2609.22151v1 [cs.CL] for this version)
https://doi.org/10.48550/arXiv.2609.22151 Focus to learn more arXiv-issued DOI via DataCite
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 333 words
