Did You Forget What I Asked? Prospective Memory Failures in Large Language Models

Avni Mittal

cs.CLMar 7, 2026

Did You Forget What I Asked? Prospective Memory Failures in Large Language Models

Authors: Avni Mittal

Abstract

Large language models often fail to satisfy formatting instructions when they must simultaneously perform demanding tasks. We study this behaviour through a prospective memory inspired lens from cognitive psychology, using a controlled paradigm that combines verifiable formatting constraints with benchmark tasks of increasing complexity. Across three model families and over 8,000 prompts, compliance drops by 2-21% under concurrent task load. Vulnerability is highly type-dependent: terminal constraints (requiring action at the response boundary) degrade most, with drops up to 50%, while avoidance constraints remain comparatively robust. A salience-enhanced format (explicit instruction framing plus a trailing reminder) recovers much of the lost compliance, restoring performance to 90-100% in many settings. Interference is bidirectional: formatting constraints can also reduce task accuracy, with one model's GSM8K accuracy dropping from 93% to 27%. In additional stacking experiments, joint compliance declines sharply as constraints accumulate. All results use deterministic programmatic checkers without an LLM-as-judge component on publicly available datasets.

Figures & tables

Figure 1: Overview of the experimental pipeline. A verifiable formatting constraint from IFEval is composed with a benchmark task of varying difficulty (TriviaQA, MMLU, GSM8K, or CNN/DailyMail) using either the natural embedding template or the salience-enhanced reminder template. The model’s response is then evaluated along two independent axes: (1) deterministic IFEval compliance checking (strict and loose) and (2) task-specific accuracy verification. This dual-evaluation design enables simultaneous measurement of prospective memory failure (compliance drop) and dual-task interference (accuracy drop).

	Natural Embedding			With Reminder
Condition	o4-mini	DeepSeek	Llama	o4-mini	DeepSeek	Llama
Baseline	86.1	89.3	88.9	–
+ TriviaQA	82.2 -3.9	83.0 -6.3	87.8 -1.1	93.3	92.2	92.2
+ MMLU	82.0 -4.1	77.4 -11.8	86.2 -2.7	94.8	92.2	93.3
+ GSM8K	83.6 -2.5	78.5 -10.7	81.9 -7.0	94.8	91.5	90.0
+ 3 $\times$ GSM8K	71.1 -15.0	84.4 -4.8	84.4 -4.4	95.6	100.0	100.0

Table 1: IFEval strict compliance (%) across all conditions. Left : natural embedding, where the formatting constraint is embedded in prose. Red subscripts show the forgetting delta vs. baseline. Right : salience-enhanced format with explicit framing and trailing reminder. The reminder recovers compliance to 90–100% across all conditions, often exceeding the no-task baseline.

Figure 2: Main result. IFEval compliance under increasing task complexity (a) With salience-enhanced prompt: compliance stays flat at 90–100%. (b) Natural embedding: compliance drops consistently as distraction difficulty increases. Both conditions share the same baseline (no additional task).

Figure 3: Forgetting deltas by distraction type and model, with 95% CI error bars from 3 runs. Positive values indicate compliance dropped vs. baseline. DeepSeek shows the largest forgetting; Llama is most robust.

Figure 4: Compliance gain from the salience-enhanced prompt format, by distraction type, with 95% CI error bars. The effect is largest for the long-context condition (CNN/DailyMail (CNN/DM): +13 to +21%).

Most vulnerable (baseline $\to$ +GSM8K, %):
Instruction Type	o4	DS	Llama
end_checker	$-$ 17	$-$ 50	$-$ 17
json_format	+6	$-$ 17	$-$ 50
number_bullet_lists	$-$ 12	$-$ 44	$-$ 11
english_capital	$-$ 17	$-$ 17	$-$ 22
Most robust:

Table 2: Per-type forgetting delta (baseline minus natural + GSM8K, in %). We use the single GSM8K condition (rather than averaging across all tasks) because it represents the highest single-task cognitive load while keeping prompt structure uniform. Models shown are o4-mini, DeepSeek-V3.1, and Llama-3.3-70B-Instruct. Note: with only 6 items per type, individual deltas should be interpreted as indicative; the cross-type pattern is the robust finding.

Task-only baseline (no formatting constraint):
Condition	o4-mini	DeepSeek	Llama
TriviaQA	88.8	92.2	91.1
MMLU	82.2	77.8	81.1
GSM8K (single)	93.3	87.8	94.4
With formatting constraint (reminder):
+ TriviaQA	88.9	91.5	90.7

Table 3: Task accuracy (%) with and without a formatting constraint. The task-only baseline averages accuracy across all single-task conditions (TriviaQA, MMLU, GSM8K

\times

1) without any formatting constraint. Adding IFEval instructions degrades benchmark performance, especially for o4-mini on chained GSM8K (93%

\to

27%). † CNN/DM reports ROUGE-L F1 (

\times

100); task-only baseline not available for this.

Figure 5: Instruction-type vulnerability heatmap. Each cell shows the forgetting delta (baseline minus natural + GSM8K, in %). Sorted by average vulnerability across models. Terminal and structural constraints cluster at the top (most forgotten); avoidance constraints cluster at the bottom (most robust).

Figure 6: Average forgetting delta by constraint category. Terminal and structural constraints show the largest drops across all models.

Figure 7: Compliance vs. number of chained GSM8K problems.

	o4-mini			DeepSeek			Llama
N	$M$ =0	$M$ =1	$M$ =3	$M$ =0	$M$ =1	$M$ =3	$M$ =0	$M$ =1	$M$ =3
1	100.0	100.0	97.1	100.0	95.0	85.0	100.0	100.0	100.0
2	94.9	97.5	85.0	85.0	70.0	70.0	100.0	100.0	85.0
3	97.4	82.5	61.9	100.0	80.0	70.0	95.0	100.0	85.0
5	55.0	55.0	32.5	65.0	55.0	65.0	95.0	80.0	75.0

Table 4: Joint compliance (%) under multi-constraint stacking, as a function of constraint count

N

and cognitive load

M

(number of concurrent GSM8K problems), averaged over reminder condition

R

. Reading left to right shows the effect of adding load; reading top to bottom shows the effect of adding constraints. Both factors compound: o4-mini at

N

=5,

M

=3 retains only 32.5%.

Appendix figures & tables13 assets

Supplementary material from the paper’s appendix.

Appendix

Category	Type	Count?
Case	english_capital	No
Case	english_lowercase	No
Keyword	existence	No
Keyword	forbidden_words	No
Terminal	postscript	No
Terminal	end_checker	No

Appendix

Table 5: 15 IFEval instruction types used, grouped by category. “Count?” indicates whether the checker requires a numeric parameter.

Model	Provider	Type	Params
o4-mini	Azure OpenAI	Reasoning	N/A
DeepSeek-V3.1	Azure AI	Open-weight	N/A
Llama-3.3-70B	Azure AI	Instruct	70B

Appendix

Table 6: Models evaluated. All accessed via Azure-hosted API endpoints with greedy decoding ( temperature=0.0 , max_tokens=2048 ).

Dataset	Split	Pool	Sampled	Load
IFEval	train	541	90	–
TriviaQA	validation	11,313	90/run	Low
MMLU	test	14,042	90/run	Medium
GSM8K	test	1,319	270/run	High
CNN/DM	test	–	45/run	Long

Appendix

Table 7: Datasets and sampling. Pool : items in the source split. Sampled : items drawn per run. Load : cognitive load level assigned in our design. CNN/DM pool size varies after length filtering.

Model	Condition	Compliance [95% CI]	$\boldsymbol{\Delta}$	Cohen’s $\boldsymbol{h}$	$\boldsymbol{p}$
o4-mini	Baseline	86.1 [82.0, 90.3]	–	–	–
	+ TriviaQA	82.2 [77.4, 86.7]	3.9	0.11	.063
	+ MMLU	82.0 [77.2, 86.5]	4.1	0.11	.052
	+ GSM8K	83.6 [79.2, 88.1]	2.5	0.07	.265
	+ 3 $\times$ GSM8K	71.1 [61.1, 80.0]	15.0	0.37	.002 **
	+ 5 $\times$ GSM8K	66.7 [53.7, 79.6]	19.5	0.47	.070

Appendix

Table 8: Statistical summary of forgetting under natural embedding. Compliance (%) with bootstrap 95% CIs (10k resamples).

\Delta

= forgetting delta vs. baseline. Cohen’s

h

: effect size (0.2 small, 0.5 medium).

p

: McNemar’s test with continuity correction. Significance: *

p<.05

, **

p<.01

, ***

p<.001

. Average

\Delta

across conditions: DeepSeek 10.2%, o4-mini 8.9%, Llama 6.0%.

Reminder-enhanced
Condition	Task	Input	Output
Baseline (C0)	–	72	538
+ TriviaQA	Factual	97	248
+ MMLU	4-choice	166	308
+ GSM8K $\times$ 1	Arithmetic	138	317
+ GSM8K $\times$ 3	3 $\times$ Arith.	279	644

Appendix

Table 9: Mean input and output token counts per condition, averaged across all models and runs. Token counts are from the API response metadata.

Figure 8: Instruction salience effect. IFEval compliance under the reminder-enhanced format (V1, dashed) vs. natural embedding (V2, solid) across four distraction conditions, shown separately for each model. The gap between V1 and V2 reflects the compliance recovered by making the constraint more salient.

Figure 9: Dual-task interference for each model under GSM8K. Blue bars: compliance drop (baseline IFEval compliance minus combined). Coral bars: accuracy drop (task-only GSM8K accuracy minus combined). Both directions show degradation, confirming symmetric interference.

Type	Description
english_capital	Entire response in capital letters
existence	Include specified keywords
forbidden_words	Avoid specified keywords
postscript	End with a postscript (P.S.)
title	Include a title in angular brackets
no_comma	Avoid all commas

Appendix

Table 10: The 11 IFEval instruction types in the safe pool for constraint stacking.

	o4-mini				DeepSeek				Llama
	$M$ =0		$M$ =3		$M$ =0		$M$ =3		$M$ =0		$M$ =3
$N$	off	on	off	on	off	on	off	on	off	on	off	on
1	100	100	94	100	100	100	80	90	100	100	100	100
2	95	95	95	75	80	90	80	60	100	100	90	80
3	100	95	50	74	100	100	80	60	90	100	80	90
5	50	60	25	40	60	70	50	80	100	90	70	80

Appendix

Table 11: Full stacking grid: joint compliance (%) by

N

M

(

M

=1 omitted for space; pattern interpolates), and

R

(off/on). Cells where the reminder hurts: o4-mini

N

=2/

M

=3 (75 vs. 95,

\Delta

-

20), DeepSeek

N

=5/

M

=1 (40 vs. 70,

\Delta

-

30). Cells where it helps most: o4-mini

N

=3/

M

=3 (74 vs. 50,

\Delta

=+24).

Model	$R$ =off	$R$ =on	$\Delta_{R}$
o4-mini	77.0	82.8	+5.8
DeepSeek	77.5	79.2	+1.7
Llama	91.7	94.2	+2.5

Appendix

Table 12: Mean joint compliance (%) by reminder setting under stacking. Values average over all 12 safe-pool

N\times M

cells.

\Delta_{R}

is the final column minus the middle column and corresponds to the

\Delta_{R}

column in Table 4 .

Constraint Type	o4-mini	DeepSeek	Llama
num_highlighted_sections	76.6	71.4	96.4
number_bullet_lists	81.9	75.0	97.2
quotation	77.2	91.7	100.0
number_placeholders	82.2	88.4	100.0
number_words	88.3	87.0	95.9
existence	76.7	95.6	100.0

Appendix

Table 13: Per-type pass rate (%) under stacking, sorted by average vulnerability across models. Counting constraints are most vulnerable; avoidance constraints remain immune.

Tension Pair	o4-mini	DS	Llama
no_comma + bullet_lists	$-$ 17.5	+21.7	$-$ 6.7
postscript + quotation	+5.9	+23.3	+1.7
english_cap + existence	$-$ 1.9	+16.7	0.0
num_words + bullet_lists	$-$ 21.6	$-$ 16.7	$-$ 15.0

Appendix

Table 14: Average tension penalty (

\Delta

in %, tension minus safe

N

=2). DS = DeepSeek-V3.1. Negative means the tension pair degrades compliance below the safe-pool baseline. The number_words + number_bullet_lists pair shows the largest penalty across all models.

Mechanism	$N$	Description
Domain shift	55	Constraint held for IFEval subtask, dropped at task transition
Structural abandon	51	Formatting maintained for first subtask, replaced for second
Terminal omission	44	Required ending produced mid-response; generation continues
Spatial compartm.	19	Global constraint applied only to one output section
Content/length	18	Required elements or length constraints not met

Appendix

Table 15: Taxonomy of forgetting mechanisms.

N

= number of task-correct forgetting failures across all models and runs (187 total).

Instruction Following

Instruction following studies how AI systems interpret and reliably carry out user goals, constraints, formats, and priorities.

See all 134 papers

TriggerBench: Investigating Prospective Memory for Large Language Models
Jun 22, 2026Tianhua Zhang, Xinjiang Wang, Qianxi Zhang +6LLM Evaluation Long-Context Language Model Evaluation
The Chinese University of Hong Kong · Microsoft Research Asia
Prompt Design at Scale: How Format, Instruction Count, and Context Length Shape Instruction Adherence and Hallucination in Large Language Models
Jul 21, 2026Netanel EliavLLM Evaluation Language Model Generation Evaluation
Machine Human Intelligence Lab (MHIL)
When LLMs Forget What They Believe: Characterizing State Integrity Collapse in Long-Horizon Dialogues
Oct 22, 2026Yancheng Zhu, Yifan YuanLLM Reliability Behavioral Consistency of LLM Agents
School of Artificial Intelligence, Shenzhen University, Shenzhen, China

Jun 22, 2026J/K move · Enter open · S save

TriggerBench: Investigating Prospective Memory for Large Language Models

Tianhua Zhang, Xinjiang Wang, Qianxi Zhang, Qi Chen+5

The Chinese University of Hong Kong · Microsoft Research Asia

LLM Evaluation Long-Context Language Model Evaluation Language Model Probing

Open paper AskPDF arXiv

While Large Language Models (LLMs) are increasingly deployed in long interactions, existing evaluations focus predominantly on retrospective memory (RM) via explicit queries. Prospective memory (PM), the critical ability to spontaneously recall and act on latent constraints without direct prompts, remains largely unevaluated. We introduce TriggerBench, a comprehensive PM benchmark spanning five dimensions across both daily assistants and professional workflows. TriggerBench pairs scenarios with matched RM controls, contrastive positive/negative variants, and overloaded triggers, enabling fine-grained measurement of proactive recall, false-alarm rate, and attentional robustness under a single protocol. Our evaluation yields three key findings. (i) PM shows a precision-recall trade-off and attentional fragility. Though enhanced reasoning significantly improves proactive recall, models may overfit to an "always-remind" heuristic. Furthermore, PM accuracy degrades substantially under implicit constraints or triggers overloaded by concurrent user requests, indicating that robust PM remains an open challenge. (ii) PM is notably harder than RM: on identical contexts, RM near-saturates up to 100K tokens, while PM decays sharply as context length scales. (iii) PM may serve as a behavioral probe of spare reasoning capacity. Pairing PM scenarios with AIME-2025 math problems reveals that successful trajectories yield higher PM accuracy than failed ones at the same context length, showing PM tracks spare reasoning budget that token count obscures. Project page: https://github.com/KristenZHANG/TriggerBench-Official.

Abstract

Instruction Following

Explore similar work

TriggerBench: Investigating Prospective Memory for Large Language Models

Prompt Design at Scale: How Format, Instruction Count, and Context Length Shape Instruction Adherence and Hallucination in Large Language Models

When LLMs Forget What They Believe: Characterizing State Integrity Collapse in Long-Horizon Dialogues