Deletion Sensitivity Is a Property of the Model: A Three-Condition Measurement over 72 Valid Experimental Cells
Organizations: University of Sussex, Brighton, United Kingdom · Chang’an University, Xi’an, China · Shanghai Shuyuan Information Technology Co., Ltd., Shanghai, China · Zhejiang Gongshang University, Hangzhou, China
Abstract
Reasoning models write a trace at every turn of a multi-turn tool session, and later turns read it. Upstream providers hold different retention constraints, so a request forwarded across issuers arrives with its history deleted, and its task cost remains unmeasured. This paper reports a direct, blind, multi-seed measurement of full history deletion in multi-turn tool calling, through three rendering conditions, a calibrated set of 72 valid experimental cells, a multi-seed protocol, and the official checker. The same deletion separates the models by direction and by magnitude, one class losing tens of points, another gaining, and a third unchanged, with the endpoint ordering reproduced across seeds and task subsets modulating the magnitude and, near zero, the sign. A re-dosed model preserves its success across the full deletion range, so the reading belongs to the model rather than the dose. Deletion sensitivity is a model-level property under an operational definition, measured under a fixed task family, scorer, and protocol, and a retrained model is a new model that a two-cell screen, one retention run and one deletion run, re-reads for the cost of two runs. Deletion sensitivity is thus a measurable property of the model, and the reading must enter routing as prior information rather than persist as an invisible default cost.
Figures & tables
| Provider | Trace visibility | Cross-turn default | Stance |
| OpenAI | encrypted, bound to issuer | renders within lineage | recommends carrying |
| Anthropic | summary plain, body encrypted | strips old turns server-side | requires current loop |
| Gemini | encrypted thought signatures | skips old-turn validation | warns of degradation |
| DeepSeek | plain text | API ignores the field | advises discarding |
| Work | Multi-turn | Deletion | Blind grading | Seeds |
| Forget selectively [ 3 ] | partial | selective | – | – |
| When to think [ 4 ] | – | dependence | – | – |
| Shift and collapse [ 1 , 2 ] | – | training-time | – | – |
| BFCL [ 35 ] | yes | – | yes | – |
| Lost in multi-turn [ 17 ] | yes | – | – | – |
| This work | yes | full, three cells | yes | yes |
| Model | Params | Tasks | Valid sessions |
| Qwen3-8B | 8 | 794 | 2382/2380 |
| Qwen3-14B | 14 | 794 | 2375/2382 |
| Qwen3-32B | 32 | 794 | 2382/2382 |
| GLM-Z1-32B | 32 | 794 | 2302/2232 |
| GLM-Z1-9B-0414 | 9 | 200 | 567/591 |
| Baichuan-M2-32B | 32 | 200 | 596/573 |
| Model | Retention | Cross-turn | Mid-session | 95% interval | ||
| Baichuan-M2-32B | 51.8 (309/596) | 0.5 (3/573) | 14.7 (93/632) | |||
| GLM-Z1-9B-0414 | 19.2 (109/567) | 2.7 (16/591) | 6.3 (37/590) | |||
| GLM-Z1-32B | 29.1 (671/2302) | 27.8 (621/2232) | 23.8 (551/2319) | |||
| Ling-mini-2.0 | 15.3 (87/569) | 15.3 (87/567) | 15.4 (87/565) | |||
| Qwen3-8B | 34.4 (819/2382) | 34.9 (831/2380) | 36.2 (861/2381) | |||
| Qwen3-14B | 34.4 (816/2375) | 37.4 (892/2382) | 35.4 (843/2382) |
| Model | Base | Long context | Missing function | Missing parameter |
| Baichuan-M2-32B | ||||
| GLM-Z1-9B-0414 | ||||
| GLM-Z1-32B | ||||
| Ling-mini-2.0 | ||||
| Qwen3-8B | ||||
| Qwen3-14B |
| Model | Ret. | Cross | Mid | Signature |
| Qwen3-8B | 44.4 | 46.1 | 37.0 | compensating |
| Qwen3-14B | 37.1 | 36.4 | 36.6 | low-delta |
| Qwen3-32B | 34.3 | 35.7 | 34.9 | low-delta |
| GLM-Z1-32B | 30.3 | 30.2 | 29.7 | low-delta |
| Ling-mini-2.0 | 38.5 | 34.7 | 36.8 | low-delta |
| R1-Distill-14B | 16.8 | 22.7 | 66.4 | looping |
| Model, task | Retention | Cross-turn | Mid-session |
| Qwen3-8B, base 128 | plan in 6 calls, done | same plan re-derived, 410 calls, done | plan held, 9 calls, done |
| R1-Distill-14B, miss_param 17 | reads the state, answers | searches, 22.4% repeat | ls(a=True) issued 418 times |
| Baichuan-M2-32B, miss_param 26 | 4 calls, correct state | cd(temp) , one call, silence | cd(temp) , stall |
| Screening reading | Routing action | Evidence |
| lower bound exceeds | keep or pin session | derived |
| upper bound falls below | delete, monitor the log | derived |
| interval holds | delete, monitor the log | derived |
| mid-session switch | expect the switch cost | derived |
| summary before delete | summarize, then delete | not yet measured |
LLM Tool Use
Large language model tool use studies how models select, call, and coordinate external tools to extend their reasoning and take actions.
See all 313 papers