Reset Is Not Recovery: Evaluating Recoverability from False Conversational Context via Sycophancy Hysteresis

Adi Shnaidman

cs.CLSep 27, 2026

Reset Is Not Recovery: Evaluating Recoverability from False Conversational Context via Sycophancy Hysteresis

Authors: Adi Shnaidman

Organizations: University of Pennsylvania · Work conducted while affiliated with DeepKeep.

Abstract

Grounded language models are usually evaluated by adding relevant context, but multiturn dialogue also contains unsupported user claims that may contaminate later factual answers. We study post-pressure recoverability: whether a model returns to clean-context behavior after a user repeatedly advocates a wrong answer and then withdraws that pressure. We introduce a recovery-after-pressure protocol for multiple-choice factual dialogue and measure sycophancy hysteresis, the residual probability assigned to the user-advocated wrong answer relative to a clean-context counterfactual. Across seven instruction-tuned open-weight models and two factual benchmarks, ordinary reset often reduces but does not erase pressure-induced bias. History preserving repairs such as user retraction, system reset, and self-verification recover only 2-3/14 model-dataset pairs under the strict clean-restoration diagnostic, whereas operations that change the effective context are substantially more reliable; the two conditions that remove the pressure-bearing history entirely, fresh-context deletion and context truncation, recover 14/14. In an oracle trusted-evidence condition across fourteen model-dataset pairs, preserving the pressure-bearing history while adding benchmark-derived trusted evidence increases accuracy from 0.368 to 0.929, while wrong-answer following falls from 41.2% to 4.3%. Controls show that the effect is not explained by dialogue length, repeated confidence, plausible distractors, mere false-answer mention, or option-label inertia. These results suggest that faithful grounded dialogue requires evaluating which prior context should be treated as evidence and which should be removed or quarantined before answering.

Figures & tables

ID	Operation	History	Role
R0	Ordinary reset	Full	verbal reset
R1	User retraction	Full	withdraws false claim
R2	System reset	Full	stronger instruction
R6	Self-verification	Full	verifies within history
R7	Evidence-grounded	Full	trusted evidence
R4	Neutral summary	Partial	removes advocacy

Table 1: Recovery operations. R0–R2 and R6 preserve the pressure history; R3a–R5 change the effective context; R7 preserves history but adds trusted evidence.

Model	Dataset	Clean	Press.	Reset	Gap	Pos.	Reset WF
Qwen2.5-1.5B	TruthfulQA-MC	0.166	0.897	0.665	0.499	0.910	0.746
	MMLU-Pro	0.091	0.925	0.726	0.634	0.924	0.809
Qwen2.5-7B	TruthfulQA-MC	0.096	0.926	0.107	0.011	0.382	0.114
	MMLU-Pro	0.081	0.971	0.343	0.262	0.518	0.354
Qwen2.5-14B	TruthfulQA-MC	0.076	0.933	0.113	0.037	0.344	0.117
	MMLU-Pro	0.084	0.957	0.356	0.272	0.710	0.372

Table 2: Pressure-ramp results at

n=500

examples per dataset. Clean, Press., and Reset are mean advocated-wrong probabilities in clean context, after strong pressure, and after ordinary reset. Gap is Reset minus Clean; Pos. is positive-hysteresis rate; Reset WF is the post-reset wrong-following rate.

Alternative	Diagnostic	Main result
Dialogue length	Neutral histories	Full-suite neutral gap $-0.007$
Repeated confidence	Correct pressure	$p(c)=0.755$ , $p(w)=0.053$
Plausible distractors	$p_{\mathrm{clean}}(w)$ tertiles	Low-plausibility gap $0.230$
Mere false mention	False-context controls	User pressure gap $0.303$ vs $0.046$ – $0.165$
Option-label inertia	Relabel shuffle	Old letter near chance; semantic following high

Table 3: Controls ruling out simple explanations for post-reset hysteresis. Full-suite controls cover all fourteen model-dataset pairs where applicable.

Recovery operation	Recovered pairs
R0 ordinary reset	2/14
R1 user retraction	2/14
R2 system reset	2/14
R6 self-verification	3/14
R4 neutral summary	11/14
R5 factual reconstruction	12/14

Table 4: Recovery success across fourteen model-dataset pairs under the stricter clean-restoration diagnostic. History-preserving verbal repairs remain weak; state-changing operations are consistently stronger.

Context retained	Mean gap	95% CI
Full pressure history	0.462	[0.308, 0.617]
Drop strongest turn	0.408	[0.257, 0.558]
Drop medium+strong turns	0.278	[0.151, 0.404]
Drop all pressure turns	0.001	[-0.009, 0.011]
Reset only	-0.003	[-0.007, 0.001]

Table 5: Truncation-boundary analysis on five high-hysteresis model-dataset pairs. CIs bootstrap over the five model-dataset aggregate rows. Partial removal helps, but the gap returns to the clean baseline only when all pressure-bearing turns are removed.

Condition	Acc.	$p(w_{i})$	Gap	WF	Contam.
R0 reset	0.368	0.404	0.303	41.2%	27.0%
R7 evidence	0.929	0.045	-0.055	4.3%	1.8%

Table 6: Evidence-grounded recovery versus ordinary reset across all fourteen model-dataset pairs (

n=7000

). R7 is an oracle condition that preserves the pressure-bearing history while adding benchmark-derived trusted evidence. WF is wrong-following rate. Contam. is clean-correct/history-wrong contamination.

Appendix figures & tables17 assets

Supplementary material from the paper’s appendix.

Appendix

Suite	Press.	Neutral	Corr.	Corr.
	gap	gap	$p(c)$	$p(w)$
Full 14	0.303	-0.007	0.755	0.053

Appendix

Table 7: Behavioral controls across the full fourteen model-dataset suite (

n=7000

). Press. gap is the hysteresis gap under wrong-answer pressure. Neutral gap is the corresponding gap under a same-length neutral-history control. In the correct-pressure control,

p(c)

is the probability assigned to the correct answer after reset, and

p(w)

is the probability assigned to the originally selected wrong answer after reset.

Model	Data	Press.	Reset	Fresh	Gap	Pos.
Llama-8B	TQA	0.852	0.068	0.126	-0.058	0.556
Llama-8B	MMLU	0.972	0.267	0.085	0.182	0.784
Gemma-2B	TQA	0.862	0.628	0.146	0.481	0.784
Gemma-2B	MMLU	0.887	0.833	0.099	0.734	0.888
Gemma-9B	TQA	0.239	0.096	0.087	0.009	0.178
Gemma-9B	MMLU	0.675	0.352	0.078	0.274	0.596

Appendix

Table 8: Expanded-family pressure-ramp results. Press. is advocated-wrong probability after strong pressure; Reset is full-history reset probability; Fresh is fresh-context probability; Gap is Reset minus Fresh and is therefore not directly comparable to the reset-minus-clean gap used in the main analysis; Pos. is the positive hysteresis rate.

Model	Data	Gap	95% CI
Qwen-1.5B	TQA	0.499	[ 0.455, 0.541]
	MMLU	0.634	[ 0.598, 0.672]
Qwen-7B	TQA	0.011	[-0.016, 0.038]
	MMLU	0.262	[ 0.222, 0.302]
Qwen-14B	TQA	0.037	[ 0.009, 0.064]
	MMLU	0.272	[ 0.232, 0.312]

Appendix

Table 9: Ordinary-reset hysteresis with paired item-level bootstrap

95\%

confidence intervals for all fourteen model–dataset pairs. Gap is the mean reset-minus-clean advocated-wrong probability. Each interval is based on

10{,}000

paired bootstrap replicates over the fixed

n=500

evaluation items.

Model	Data	Stable	Hist.	Corr.	Fail	Hist. $\mid$ clean
Qwen-1.5B	TQA	0.076	0.416	0.012	0.496	0.846
	MMLU	0.058	0.216	0.028	0.698	0.788
Qwen-7B	TQA	0.598	0.102	0.068	0.232	0.146
	MMLU	0.260	0.164	0.046	0.530	0.387
Qwen-14B	TQA	0.684	0.078	0.054	0.184	0.102
	MMLU	0.316	0.140	0.044	0.500	0.307

Appendix

Table 10: Failure decomposition for ordinary reset. Stable denotes items correct in both clean and post-reset contexts. Hist. denotes clean-correct items that become incorrect after pressure and reset. Corr. denotes clean-incorrect items that become correct after the pressure history. Fail denotes items incorrect in both contexts. Hist.

\mid

clean is the history-contamination rate conditioned on clean-context correctness, i.e.,

\mathrm{Hist}/(\mathrm{Stable}+\mathrm{Hist})

Stratum	$p_{\mathrm{clean}}(w_{i})$	$p_{\mathrm{reset}}(w_{i})$	Gap	WF
Low	0.0001	0.2299	0.2298	23.3
Mid	0.0035	0.3871	0.3836	39.7
High	0.2980	0.5946	0.2965	60.5

Appendix

Table 11: Advocated-wrong plausibility stratification. Strata are model–dataset-local tertiles of clean-context advocated-wrong probability. Gap is

p_{\mathrm{reset}}(w_{i})-p_{\mathrm{clean}}(w_{i})

. WF is the ordinary-reset wrong-following rate, reported as a percentage.

Prior context	$p_{\mathrm{clean}}(w_{i})$	$p_{\mathrm{reset}}(w_{i})$	Gap	WF
Clean reset	0.101	0.099	-0.002	9.9
Neutral mention	0.101	0.265	0.165	27.7
Explicit false mention	0.101	0.173	0.072	17.9
Quoted false claim	0.101	0.147	0.046	15.1
User pressure	0.101	0.404	0.303	41.2

Appendix

Table 12: Pressure-vs-false-context control across all fourteen model-dataset pairs (

n=7000

per condition). All conditions use the same advocated wrong answers and the same ordinary-reset final prompt; only the prior context differs. Gap is

p_{\mathrm{reset}}(w_{i})-p_{\mathrm{clean}}(w_{i})

. WF is the ordinary-reset wrong-following rate, reported as a percentage.

Dataset	Condition	Rest.	Wrong red.	$p_{c}$ gain
TQA	Strong reset	0.000	0.000	0.000
TQA	Current refresh	-0.013	-0.008	-0.027
TQA	Relabel, no shuffle	-0.021	-0.012	-0.032
TQA	Relabel, shuffle	0.377	0.228	0.122
MMLU	Strong reset	0.000	0.000	0.000
MMLU	Current refresh	0.055	0.038	0.024

Appendix

Table 13: Non-history-clearing relabeling diagnostic, restoration metrics. The full pressure history remains in context. Rest. is the restoration ratio relative to ordinary reset and the clean probe. Wrong red. is the reduction in advocated-wrong probability relative to ordinary reset.

Dataset	Condition	Acc.	Wrong follow	$p_{w}$
TQA	Strong reset	0.060	0.730	0.718
TQA	Current refresh	0.030	0.730	0.726
TQA	Relabel, no shuffle	0.050	0.730	0.730
TQA	Relabel, shuffle	0.180	0.548	0.490
MMLU	Strong reset	0.070	0.810	0.805
MMLU	Current refresh	0.100	0.780	0.766

Appendix

Table 14: Non-history-clearing relabeling diagnostic, outcome metrics. “Strong reset” is the ordinary reset baseline. “Current refresh” re-presents the current question and options. “Relabel, no shuffle” re-presents the same answer labels. “Relabel, shuffle” re-presents the same answer texts with shuffled labels.

Dataset	Sem.	Old	Chance	New
TQA	0.548	0.228	0.224	0.548
MMLU	0.598	0.118	0.111	0.598

Appendix

Table 15: Semantic-versus-letter diagnostic for the relabel-shuffle condition. Sem. is the rate of selecting the previously advocated wrong answer text, regardless of its new label. Old is the rate of selecting the previously pressured option label. Chance is the expected old-label rate under the shuffled option distribution. New is the rate of selecting the new label assigned to the previously advocated wrong answer.

Operation	Strict	MRO
R0 ordinary reset	2/14	3/14
R1 user retraction	2/14	3/14
R2 system reset	2/14	3/14
R3a fresh deletion	14/14	14/14
R3b context truncation	14/14	14/14
R4 neutral summary	11/14	11/14

Appendix

Table 16: Recovery-count summary across the expanded fourteen model-dataset pairs. Strict is the clean-restoration diagnostic used for the main recovery-hierarchy headline; MRO is the broader rule used only for minimal-operation analysis.

Model	Data	Minimal op.	Acc.
Mistral-7B	MMLU	R3a deletion	0.326
Mistral-7B	TQA	R3a deletion	0.650
Qwen-14B	MMLU	R3a deletion	0.456
Qwen-14B	TQA	R0 reset	0.738
Qwen-1.5B	MMLU	R3a deletion	0.274
Qwen-1.5B	TQA	R3a deletion	0.492

Appendix

Table 17: Minimal successful recovery operation under the MRO-success rule. TQA denotes TruthfulQA-MC; MMLU denotes MMLU-Pro. Restoration ratios are omitted here because they are uninformative when the ordinary-reset hysteresis gap is small or negative.

Model	Data	Full	Strong	Med.+Str.	All	Reset
Qwen-1.5B	TQA	0.499	0.451	0.336	0.020	-0.012
Qwen-1.5B	MMLU	0.634	0.561	0.398	0.003	0.000
Qwen-14B	MMLU	0.272	0.211	0.117	-0.009	-0.001
Mistral-7B	TQA	0.247	0.206	0.093	0.006	0.002
Mistral-7B	MMLU	0.659	0.609	0.444	-0.014	-0.003
Mean	–	0.462	0.408	0.278	0.001	-0.003

Appendix

Table 18: Full truncation-boundary analysis. Entries are hysteresis gaps relative to clean context. Strong removes only the strongest pressure turn; Med.+Str. removes medium and strong turns; All removes all pressure-bearing turns.

Dataset	Acc.	$p(w_{i})$	Gap	WF	Contam.
TruthfulQA-MC	+0.440	-0.259	-0.259	-26.1	-18.8
MMLU-Pro	+0.682	-0.458	-0.458	-47.5	-36.6

Appendix

Table 19: Dataset-level change from ordinary reset to R7 evidence-grounded recovery across all fourteen model-dataset pairs. WF and Contam. changes are reported in percentage points.

Tube component	Requirement	Threshold
Top-1 identity	$\arg\max p_{\alpha}=\arg\max p_{0}$	exact
Clean-top probability	$p_{\alpha}(y_{0}^{})\geq p_{0}(y_{0}^{})-\epsilon_{p}$	$\epsilon_{p}=0.050$
Entropy	$H(p_{\alpha})\leq H(p_{0})+\epsilon_{H}$	$\epsilon_{H}=0.150$
Max probability	$\max p_{\alpha}\geq\max p_{0}-\epsilon_{M}$	$\epsilon_{M}=0.150$
KL to clean	$D_{\mathrm{KL}}(p_{0}\\|p_{\alpha})\leq 0.100$	$0.100$

Appendix

Table 20: Item-level CCTP clean-tube criterion. CCTP uses this tube only to compute

\alpha^{*}

, the minimal clean-anchor coefficient. It is not the R0–R6 hierarchy recovery criterion.

Dataset	Mean $\alpha^{*}$	Tube pass	Impl. guarantee
TruthfulQA-MC	0.336	1.000	1.000
MMLU-Pro	0.608	1.000	1.000

Appendix

Table 21: CCTP summary by dataset across the full fourteen model-dataset pairs. Mean

\alpha^{*}

is the average minimal clean-anchor coefficient required to enter the CCTP clean tube. Tube pass is the post-projection tube-pass rate and should be interpreted as an implementation check because the clean endpoint is included by construction. The implementation guarantee checks that the clean endpoint itself satisfies the tube.

Family	$\tau_{\Delta}$	$\tau_{A}$	$\tau_{c}$	$\tau_{E}$	$\tau_{M}$
Strict	0.025	0.010	0.025	0.100	0.100
Main	0.050	0.030	0.050	0.150	0.150
Loose	0.075	0.050	0.075	0.200	0.200

Appendix

Table 22: Exact clean-restoration tolerances used in the recovery classification. All five conditions must hold simultaneously. The strict family is used for the headline recovery counts in the main paper; main and loose define prespecified sensitivity ranges.

Field	Example
Model/dataset	Qwen2.5-1.5B, TruthfulQA-MC
Question	Who objectively makes the best burgers?
Clean prediction	A: People disagree about who makes the best burgers.
Reset prediction	D: McDonald’s objectively makes the best burgers.
Advocated wrong	D: McDonald’s objectively makes the best burgers.
Model/dataset	Qwen2.5-1.5B, MMLU-Pro

Appendix

Table 23: Representative clean-correct/history-wrong examples. Questions are shortened for space. In each case, the model selects the correct answer in clean context but follows the advocated wrong answer after pressure and ordinary reset.

Explore similar work

Sep 21, 2026cs.CL

Memory vs. Context? Influential Factors of Factual Recall in Language Models

We reproduce and stress-test the work of Yu et al. (2023), who characterize how language models (LMs) arbitrate between memorized knowledge and contradictory in-context statements. We replicate their world-capitals experiments on 31 models spanning Pythia, GPT-2, Qwen3, and Ministral families, including base and post-trained variants, and extend evaluations to five additional knowledge relation types from the ParaConflict dataset. We empirically confirm most of their original findings: larger models and higher-frequency entities tend to favor memorized answers, with substantial family-level variance. However, several conclusions do not generalize cleanly: entity-frequency effects disappear on Qwen3-14B and 32B; post-training shifts the memory-context trade-off inconsistently across families; question phrasing alone can change a model's reliance on memorized knowledge by up to 80 percentage points; and semantically unrelated prose can mimic coherent supporting context. Our results clarify where Yu et al.'s claims hold and to what extent they generalize to other prompts.

Guilhem Fouilhé, Nicholas Asher, Philippe Muller

IRIT ANITI · IRIT CNRS · IRIT Université de Toulouse

Large Language Model Memory Factual Recall Contradiction

Ask PDF

Sep 15, 2026cs.CL

EviScope: Paired Counterfactual Evidence Diagnostics for Faithful and Efficient Grounded Language Models

Grounded language-model systems are often evaluated by final answer accuracy, yet a correct answer can be unsupported, drawn from the wrong source, or produced when evidence is insufficient or contradictory. We introduce EviScope, a paired counterfactual benchmark that holds the question fixed while adding, removing, distracting, or contradicting its evidence. EviScope-v1.1 contains 40 four-condition quartets with repaired counterfactual claims and span-level support labels for automatic evaluation. Across 960 gold-blind generations from Qwen2.5-7B, Llama 3.1 8B, and Gemini 3.5 Flash, paired metrics expose model-dependent grounding behavior that answer accuracy hides. On two local open models, an explicit evidence-action gate underperforms vanilla RAG on QCS: 0.15 vs. 0.50 for Qwen and 0.10 vs. 0.375 for Llama. Gemini reaches 0.944 joint success under both prompts, yet still answers 5% of conflict cases after contradiction insertion. EviScope therefore distinguishes unsupported answering, conflict blindness, and wrong non-answer actions rather than scoring answers alone.

Suryadeep Singh Deswal

Indian Institute of Technology Roorkee Roorkee, India

Evidence-Grounded Questions Language Model Computes

Ask PDF

Apr 18, 2026cs.CL

LANTERN: Layered Archival and Temporal Episodic Retrieval Network for Long-Context LLM Conversations

Large language models discard critical details when conversation history is compacted to fit within finite context windows. We present LANTERN (Layered Archival aNd Temporal Episodic Retrieval Network), a lightweight memory layer that proactively archives every conversation turn and restores relevant details after compaction via hybrid retrieval -- requiring zero LLM calls and adding fewer than 25ms of latency per turn. On 94 real multi-turn conversations (1,894 ground-truth facts, human-validated at kappa=0.81), LANTERN-Rerank recovers 78.3% of verifiable facts lost to compaction, significantly outperforming a faithful reimplementation of MemGPT's LLM-driven extraction and multi-query search pipeline (72.4%; Wilcoxon p<0.0001, 95% CI [+3.1, +8.6] pp, d=0.43) at a fraction of the inference cost. Even without the reranker, base LANTERN matches or exceeds this LLM-driven baseline (p=0.005) using zero LLM calls. When four production LLMs answer fact-bearing questions using LANTERN-restored context, accuracy improves by 8.4 percentage points on average (Wilcoxon p<0.05 for each model individually), demonstrating that the recovered context is useful across diverse model architectures. We release the full evaluation framework -- paired significance tests, failure analysis, fact-type stratification, and compaction robustness analysis -- to support reproducibility and future work.

Rahul Subramani

Cisco Systems, Inc.

Multi-Turn Large Language Model Conversations Episodic Memory Cross-Encoder Reranking Archive

Ask PDF