pith. sign in

arxiv: 2502.09487 · v3 · pith:72APJCVSnew · submitted 2025-02-13 · 💻 cs.CL · cs.AI· cs.LG

Internal narratives parameterise affective states

Pith reviewed 2026-05-23 03:12 UTC · model grok-4.3

classification 💻 cs.CL cs.AIcs.LG
keywords internal narrativesaffective statesdepressionlanguage model embeddingssymptom covarianceconstruct validitytemporal dynamics
0
0 comments X

The pith

High-dimensional text representations of internal narratives mirror the latent geometry of depression by preserving symptom covariances.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tests whether written descriptions of depressive symptoms can be turned into quantitative parameters that capture the structure of affective states. Across two studies with 1257 participants, large-language-model encodings of those descriptions predicted standardised depression scores, but only when the specific pattern of covariances among symptoms was retained. This preservation is presented as necessary for the representations to reflect the disorder's underlying geometry. The second study further shows that shifts in the encoded narratives precede and predict shifts in self-reported affect, while initial narrative severity forecasts the size of later change.

Core claim

Verbal descriptions of symptom-specific thoughts captured granular information predictive of standardised, self-reported depression scores. Critically, preserving the specific covariance between symptoms is essential for construct validity, suggesting high-dimensional text representations mirror the latent geometry of the disorder. Quantified changes in internal narratives led to changes in self-report, while the baseline narrative severity predicted the magnitude of subsequent affective change.

What carries the argument

Large-language-model representations of participants' written symptom descriptions and their subspaces, used to parameterise the structure and dynamics of depressive states.

If this is right

  • Preserving the specific covariance structure among symptoms is required for text-based measures to maintain construct validity with depression.
  • Alterations in the content of internal narratives produce measurable shifts in subsequent self-reported affective states.
  • The severity encoded in a person's baseline narrative forecasts the extent of affective change that will follow exposure to emotional material.
  • Affective states function both to constrain the organisation of internal narratives and to integrate contextual input into self-report.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Interventions that rewrite symptom narratives could be tested for downstream effects on reported mood.
  • If the covariance-preserving property holds, routine writing samples might allow longitudinal tracking of depressive geometry without repeated questionnaires.
  • The same encoding approach could be examined for disorders whose symptom covariances differ from those of depression.

Load-bearing premise

Large language model embeddings of participants' written symptom descriptions faithfully capture the latent structure of their internal narratives without systematic distortion from model training data or architecture.

What would settle it

If text embeddings that scramble or ignore the specific covariances among symptoms predict depression scores equally well as those that preserve the covariances, the necessity of covariance preservation for construct validity would be falsified.

Figures

Figures reproduced from arXiv: 2502.09487 by Jakub Onysk, Quentin J. M. Huys.

Figure 1
Figure 1. Figure 1: A: Study 1 design. In study 1, open-ended and equivalent multiple-choice depression symptom questions were asked. Following open-ended questions, participants completed multiple￾choice PHQ-9, GAD-7 and SDS questionnaires, assessing depression and anxiety. B: LLM prompt design and sampling. We take the open-ended question, OQq, and participant’s open-ended answer, Aq,i, forming participant’s QA-pair. We app… view at source ↗
Figure 2
Figure 2. Figure 2: A-H: Violin, box and scatter plots, with correlations (p < 0.001) of subject’s item-level scores vs the best performing model (Gemma2-9B) scores for the PHQ-9 items given the corre￾sponding open-ended QA-pair. I: Correlation map showcasing performance (p < 0.001) of each LLM at predicting PHQ-9 scores from open-ended QA-pairs, see scatter plots in Appendix B.5.2. Next, we evaluate LLMs’ response generalisa… view at source ↗
Figure 3
Figure 3. Figure 3: A: Subjects’ ground-truth pairwise correlations between PHQ-8 item scores and SDS scores. B: Correlations between true SDS item scores (x-axis) and Gemma2-9B recovered scores for that question given each of the open-ended PHQ-9 QA-pairs (y-axis). Significant correlations are in cool-warm colour. The correlations not reaching significance are in gray scale. C: Subject’s total scores for the SDS depression q… view at source ↗
Figure 4
Figure 4. Figure 4: A-I: Violin, box and scatter plots, with correlations (p < 0.001) of subject’s item-level z-scores vs the best performing sSAE model (layer 42, Gemma2-9B) predicted scores for PHQ￾9 items, given the average hidden state across all open-ended responses for each participant. J: Correlation between participants’ item-level scores and sSAE predicted scores for each layer and each question. K: Confusion matrix … view at source ↗
Figure 5
Figure 5. Figure 5: The effect of hidden states sSAE perturbation on sampled item-level PHQ-9 responses, [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: A: Violin, box and scatter plots of changes in cognitive measures after mood-induction (FU) for each condition (negative ML - mood low; positive MH - high mood) relative to baseline. Interventions are effective as the differences for PHQ-9 Q2 score, momentary mood and word recall sentiment show significant differences between conditions. B: sSAE latent Q2 scores for each par￾ticipant’s created, autobiograp… view at source ↗
Figure 7
Figure 7. Figure 7: A-H: Violin, box and scatter plots, with correlations (p < 0.001) of subject’s item-level scores vs the best performing model (Mistral-7B-OpenOrca) scores for the PHQ-9 items given the corresponding open-ended QA-pair. 0 1 2 3 Subject score 0 1 2 3 LLM scoreA Q1: r=0.440 p-val: 1.03e-36 0 1 2 3 Subject score 0 1 2 3 LLM scoreB Q2: r=0.687 p-val: 1.55e-107 0 1 2 3 Subject score 0 1 2 3 LLM scoreC Q3: r=0.64… view at source ↗
Figure 8
Figure 8. Figure 8: A-H: Violin, box and scatter plots, with correlations (p < 0.001) of subject’s item-level scores vs the best performing model (Gemma2-2B) scores for the PHQ-9 items given the corre￾sponding open-ended QA-pair. 17 [PITH_FULL_IMAGE:figures/full_fig_p017_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: A-H: Violin, box and scatter plots, with correlations (p < 0.001) of subject’s item-level scores vs the best performing model (Llama3.1-8B) scores for the PHQ-9 items given the corre￾sponding open-ended QA-pair. 0 1 2 3 Subject score 0 1 2 3 LLM scoreA Q1: r=0.380 p-val: 6.75e-27 0 1 2 3 Subject score 0 1 2 3 LLM scoreB Q2: r=0.696 p-val: 1.40e-111 0 1 2 3 Subject score 0 1 2 3 LLM scoreC Q3: r=0.614 p-val… view at source ↗
Figure 10
Figure 10. Figure 10: A-H: Violin, box and scatter plots, with correlations (p < 0.001) of subject’s item￾level scores vs the best performing model (Llama3.2-3B) scores for the PHQ-9 items given the corresponding open-ended QA-pair. B.5.3 QUESTIONNAIRE GENERALISATION LOGIT SAMPLING -REMAINING MODELS RESULTS 18 [PITH_FULL_IMAGE:figures/full_fig_p018_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Column I (A, D, G): Subject’s total scores for each questionnaire against the Gemma2- 9B total score estimate (Pearson correlation reported). Column II (B, E, H): Subjects’ pairwise correlations between PHQ-9 item scores and SDS, GAD-7 and PHQ-9 scores. Column III (C, F, I): Correlations between participant’s true item score for each questionnaire (x-axis) and Gemma2-9B recovered score for that question g… view at source ↗
Figure 12
Figure 12. Figure 12: Column I (A, D, G): Subject’s total scores for each questionnaire against the Mistral￾7B-OpenOrca total score estimate (Pearson correlation reported). Column II (B, E, H): Subjects’ pairwise correlations between PHQ-9 item scores and SDS, GAD-7 and PHQ-9 scores. Column III (C, F, I): Correlations between participant’s true item score for each questionnaire (x-axis) and Mistral-7B-OpenOrca recovered score … view at source ↗
Figure 13
Figure 13. Figure 13: Column I (A, D, G): Subject’s total scores for each questionnaire against the Gemma2- 2B total score estimate (Pearson correlation reported). Column II (B, E, H): Subjects’ pairwise correlations between PHQ-9 item scores and SDS, GAD-7 and PHQ-9 scores. Column III (C, F, I): Correlations between participant’s true item score for each questionnaire (x-axis) and Gemma2-2B recovered score for that question g… view at source ↗
Figure 14
Figure 14. Figure 14: Column I (A, D, G): Subject’s total scores for each questionnaire against the Llama3.1- 8B total score estimate (Pearson correlation reported). Column II (B, E, H): Subjects’ pairwise correlations between PHQ-9 item scores and SDS, GAD-7 and PHQ-9 scores. Column III (C, F, I): Correlations between participant’s true item score for each questionnaire (x-axis) and Llama3.1-8B recovered score for that questi… view at source ↗
Figure 15
Figure 15. Figure 15: Column I (A, D, G): Subject’s total scores for each questionnaire against the Llama3.2- 3B total score estimate (Pearson correlation reported). Column II (B, E, H): Subjects’ pairwise correlations between PHQ-9 item scores and SDS, GAD-7 and PHQ-9 scores. Column III (C, F, I): Correlations between participant’s true item score for each questionnaire (x-axis) and Llama3.2-3B recovered score for that questi… view at source ↗
Figure 16
Figure 16. Figure 16: Plot of training and validation loss for the best hyper-parameter setting sSAE based on [PITH_FULL_IMAGE:figures/full_fig_p023_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Screenshot of the mood rating VAS scale used in Study 3. [PITH_FULL_IMAGE:figures/full_fig_p025_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Screenshot of the PHQ-9 rating VAS scale used in Study 3 for example question [PITH_FULL_IMAGE:figures/full_fig_p025_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: sSAE latent scores for each participant’s created, autobiographical diary entry show the [PITH_FULL_IMAGE:figures/full_fig_p028_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: sSAE latent scores for each participant’s positive re-evaluations open-ended descriptions. [PITH_FULL_IMAGE:figures/full_fig_p028_20.png] view at source ↗
Figure 21
Figure 21. Figure 21: sSAE latent scores for each participant’s baseline mood open-ended descriptions. [PITH_FULL_IMAGE:figures/full_fig_p029_21.png] view at source ↗
Figure 22
Figure 22. Figure 22: sSAE latent scores for each participant’s baseline energy levels open-ended descriptions. [PITH_FULL_IMAGE:figures/full_fig_p029_22.png] view at source ↗
read the original abstract

Characterising how we verbalise our feelings is central to psychological assessment and intervention, yet the mapping between narrative and affective state remains poorly understood. Across two large studies (n=1257), we parameterised the structure and dynamics of depressive states by quantifying participants' internal narratives through large-language-model representations and their subspaces. In Study 1, we found verbal descriptions of symptom-specific thoughts captured granular information predictive of standardised, self-reported depression scores. Critically, we show preserving the specific covariance between symptoms is essential for construct validity, suggesting high-dimensional text representations mirror the latent geometry of the disorder. Study 2 probed the temporal dynamics of this relationship as participants engaged with emotional narratives. We found quantified changes in internal narratives led to changes in self-report, while the baseline narrative severity predicted the magnitude of subsequent affective change. By framing affect as a computational state, our results highlight its core, therapeutically pertinent functions: constraining the structure of internal narratives and integrating context to shape self-report.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper claims that across two studies (total n=1257), large-language-model representations of participants' written internal narratives about depressive symptoms can parameterize affective states. Study 1 reports that symptom-specific narrative descriptions capture granular information predictive of standardized self-reported depression scores, with preservation of the specific covariance between symptoms being essential for construct validity and implying that high-dimensional text representations mirror the latent geometry of the disorder. Study 2 examines temporal dynamics during engagement with emotional narratives, finding that quantified changes in internal narratives lead to changes in self-report while baseline narrative severity predicts the magnitude of subsequent affective change. The work frames affect as a computational state with implications for assessment and intervention.

Significance. If the central results hold after appropriate controls, the work has moderate significance for computational approaches to affective science. It offers a concrete demonstration that LLM-derived subspaces can preserve symptom covariance structures, providing a potential bridge between narrative data and latent disorder geometry. The temporal component adds value by linking narrative dynamics to self-report change, which could inform therapeutic applications. Strengths include the large sample and the explicit focus on covariance preservation as a validity criterion.

major comments (2)
  1. [Abstract / Study 1] Abstract and Study 1: the claim that narrative representations are predictive of self-reported depression scores and that covariance preservation is essential for construct validity cannot be evaluated without the statistical methods, controls, sample details, error estimates, or ablation results (e.g., on embedding model choice or comparison to non-LLM baselines). This information is load-bearing for the central claim that text representations mirror disorder geometry.
  2. [Study 1] Study 1: it remains unclear whether the LLM embeddings of symptom descriptions are independent of the self-report validation data or whether training-data leakage or architectural biases introduce circular dependence; this directly affects the interpretation that the representations faithfully capture latent narrative structure without systematic distortion.
minor comments (1)
  1. The manuscript would benefit from explicit definitions or equations for the 'subspaces' derived from the LLM representations and how covariance is quantified and preserved.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive comments. We address each major point below, providing clarifications on methods and independence while revising the manuscript to improve evaluability of the central claims.

read point-by-point responses
  1. Referee: [Abstract / Study 1] Abstract and Study 1: the claim that narrative representations are predictive of self-reported depression scores and that covariance preservation is essential for construct validity cannot be evaluated without the statistical methods, controls, sample details, error estimates, or ablation results (e.g., on embedding model choice or comparison to non-LLM baselines). This information is load-bearing for the central claim that text representations mirror disorder geometry.

    Authors: We agree these details are necessary for full evaluation. The Methods section specifies the statistical approach (ridge regression and canonical correlation analysis with demographic controls for age/gender/education), sample details (Study 1 n=628), bootstrap-derived error estimates, and ablations across embedding models (e.g., all-MiniLM-L6-v2 vs. MPNet) plus non-LLM baselines (TF-IDF, LDA). To make this load-bearing information more accessible without supplements, we have added a concise summary paragraph and ablation table to the main Study 1 Results, plus explicit covariance-preservation metrics with confidence intervals. revision: yes

  2. Referee: [Study 1] Study 1: it remains unclear whether the LLM embeddings of symptom descriptions are independent of the self-report validation data or whether training-data leakage or architectural biases introduce circular dependence; this directly affects the interpretation that the representations faithfully capture latent narrative structure without systematic distortion.

    Authors: Embeddings were produced by applying frozen, pre-trained sentence-transformer models zero-shot to participants' free-text narratives; no fine-tuning or training occurred on any study data, including the separate PHQ-9 self-reports used solely for validation. This design avoids direct circularity. Architectural biases were probed via cross-model ablations (reported in supplementary analyses). We have inserted explicit language in the revised Methods clarifying the independence, zero-shot application, and absence of leakage from validation labels. revision: yes

Circularity Check

0 steps flagged

No significant circularity

full rationale

The abstract and available excerpts frame the work as an empirical parameterization: LLM embeddings of symptom narratives are shown to capture information predictive of independent self-report scores, with covariance preservation treated as a validity check rather than a definitional step. No equations, fitted parameters, or self-citations are quoted that reduce any claimed prediction or geometry to the input data by construction. The central claims rest on statistical associations between distinct measures (text representations vs. standardized scales) and temporal dynamics in Study 2, without evidence of self-definitional loops, renamed fits, or load-bearing self-citations. The derivation is therefore self-contained as an observational mapping rather than a tautology.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

No free parameters, axioms, or invented entities are identifiable from the abstract alone; the central claims rest on unstated assumptions about LLM embedding fidelity and self-report validity.

pith-pipeline@v0.9.0 · 5695 in / 1062 out tokens · 35688 ms · 2026-05-23T03:12:30.297001+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

52 extracted references · 52 canonical work pages · 1 internal anchor

  1. [1]

    arXiv:2306.02707 [cs]. OpenAI. GPT-4 Technical Report, 2023. arXiv:2303.08774 [cs]. Prolific, 2025. URLhttps://www.prolific.com. Quentin Raffaelli, Caitlin Mills, Nadia-Anais de Stefano, Matthias R. Mehl, Kate Chambers, Surya A. Fitzgerald, Ramsey Wilcox, Kalina Christoff, Eric S. Andrews, Matthew D. Grilli, Mary- Frances O’Connor, and Jessica R. Andrews-...

  2. [2]

    Could you share any activities or events from the past two weeks that made you feel both- ered because of a lack of interest or pleasure in doing them?

  3. [4]

    Can you provide examples of how your sleep has been in the past two weeks? Have you been bothered by challenges with falling asleep, staying asleep, or even sleeping too much?

  4. [5]

    Have you been bothered by your energy levels over the past two weeks? Can you recall situations when it comes to feeling tired/lively or low/high on energy? 14

  5. [6]

    In the past two weeks, have you been bothered about your appetite? Can you describe your typical attitude towards food - maybe you have noticed something unusual, like changes in how much you’re eating or not eating?

  6. [7]

    In the past two weeks, have you been bothered by feelings about yourself? In what situa- tions did you feel proud or like a failure? Did you feel you met your own and your family’s expectations, or let them down?

  7. [8]

    In the past two weeks, have you been bothered by your ability to concentrate and focus? Please describe how it felt to do things that require you to concentrate for a while, like working, reading, or watching movies?

  8. [9]

    These are answered on a scale - Not at all (0), Several days (1), More than half the days (2), Nearly every day (3) (Kroenke et al., 2001)

    Can you describe situations over the past two weeks when you were bothered by feeling slower than usual in terms of thinking, speaking, or just acting - or situations where you felt fidgety and restless? B.3 MULTIPLE-CHOICE QUESTIONS B.3.1 PHQ-9ITEMS Participants were instructed to consider ”Over the last 2 weeks, how often have you been bothered by any o...

  9. [10]

    Little interest or pleasure in doing things

  10. [11]

    Feeling down, depressed, or hopeless

  11. [12]

    Trouble falling or staying asleep, or sleeping too much

  12. [13]

    Feeling tired or having little energy

  13. [14]

    Poor appetite or overeating

  14. [15]

    Feeling bad about yourself - or that you are a failure or have let yourself or your family down

  15. [16]

    Trouble concentrating on things, such as reading the newspaper or watching television

  16. [17]

    Moving or speaking so slowly that other people could have noticed? Or the opposite - being so fidgety or restless that you have been moving around a lot more than usual

  17. [18]

    B.3.2 SDSITEMS These are answered on a scale - A little of the time (1), Some of the time (2), Good part of the time (3), Most of time (4)

    Thoughts that you would be better off dead or of hurting yourself in some way. B.3.2 SDSITEMS These are answered on a scale - A little of the time (1), Some of the time (2), Good part of the time (3), Most of time (4). Questions 2, 5, 6, 11, 12, 14, 16, 17, 18 and 20 are reverse scored Zung (1965)

  18. [19]

    I feel down-hearted and blue

  19. [20]

    Morning is when I feel the best

  20. [21]

    I have crying spells or feel like it

  21. [22]

    I have trouble sleeping at night

  22. [23]

    I eat as much as I used to

  23. [24]

    I notice that I am losing weight

  24. [25]

    I have trouble with constipation

  25. [26]

    My heart beats faster than usual

  26. [27]

    I get tired for no reason

  27. [28]

    My mind is as clear as it used to be

  28. [29]

    I find it easy to do the things I used to

  29. [30]

    I am restless and can’t keep still 15

  30. [31]

    I feel hopeful about the future

  31. [32]

    I am more irritable than usual

  32. [33]

    I find it easy to make decisions

  33. [34]

    I feel that I am useful and needed

  34. [35]

    My life is pretty full

  35. [36]

    I feel that others would be better off if I were dead

  36. [37]

    These are answered on a scale - Not at all (0), Several days (1), More than half the days (2), Nearly every day (3) (Spitzer et al., 2006)

    I still enjoy the things I used to do B.3.3 GAD-7ITEMS Participants were instructed to consider ”Over the last 2 weeks, how often have you been bothered by any of the following problems?”. These are answered on a scale - Not at all (0), Several days (1), More than half the days (2), Nearly every day (3) (Spitzer et al., 2006)

  37. [38]

    Feeling nervous, anxious or on edge

  38. [39]

    Not being able to stop or control worrying

  39. [40]

    Worrying too much about different things

  40. [41]

    Being so restless that it is hard to sit still

  41. [42]

    Becoming easily annoyed or irritated

  42. [43]

    <start_of_turn>user Please answer the following question in detail

    Feeling afraid as if something awful might happen B.4 EXAMPLE ITEM-LEVELLLMPROMPT We used a system prompt ”You are a human participant in a study.” Below is the prompt we used to sample responses from the logits of each response label character (A, B, C, D). <start_of_turn>user Please answer the following question in detail. Question: For the past two wee...

  43. [44]

    For the past two weeks, have you been bothered by your mood and how you felt generally? Were there any situations when you felt down, depressed, or hopeless?

  44. [45]

    See Fig 17

    Have you been bothered by your energy levels over the past two weeks? Can you recall situations when it comes to feeling tired/lively or low/high on energy? D.2.2 MOMENTARY MOOD Before and after mood induction, participants were asked to rate their momentary mood by answer- ing a question ”How happy are you at this moment?” on a continuos visual analogue ...

  45. [46]

    I feel joyful, grateful, and full of positive energy

    Lately, I’ve been really focused on making good memories and enjoying happy moments with my friends and family. I feel joyful, grateful, and full of positive energy. I think it’s because things have been going my way lately — it’s got me feeling excited and motivated

  46. [47]

    I’m surrounded by people I love, and that really makes me happy

    I was feeling elated today — there was so much positivity around me. I’m surrounded by people I love, and that really makes me happy. Work’s been great too; I have awesome colleagues, and we just wrapped up a big project together. I’m excited for what’s next and ready to keep achieving more

  47. [48]

    I’m going out more freely now, with no anxiety, and actually enjoying hanging out with friends

    I’ve been in a happy, light mood and feeling really in control. I’m going out more freely now, with no anxiety, and actually enjoying hanging out with friends. I’ve just been feeling good doing all the regular life stuff. Overall, I’m optimistic, content, and hopeful about what’s ahead

  48. [49]

    I also took some time off to focus on myself and just relax

    I’ve had a really positive couple of weeks, with good things happening both at work and at home. I also took some time off to focus on myself and just relax. It’s such a nice feeling — I just feel happy and content. D.3.4 NEGATIVE MOOD INDUCTION TRANSCRIPTS

  49. [50]

    It’s been tough — I’ve felt lonely and de- pressed, like something bad is hanging over me

    I’ve been feeling really overwhelmed lately. It’s been tough — I’ve felt lonely and de- pressed, like something bad is hanging over me. I can’t even remember the last time I didn’t feel this way

  50. [51]

    It’s hard to enjoy anything

    I feel like I’m wasting my life. It’s hard to enjoy anything. Everything just seems pointless, and I don’t know how to change it. I’m so tired of getting rejection after rejection with job applications. I keep comparing myself to others, and it makes me feel pathetic. The world just feels like a cold, evil place

  51. [52]

    At work, all I can think about is getting home, and once I’m home, I just want to sleep the day away

    I just feel too low to do anything. At work, all I can think about is getting home, and once I’m home, I just want to sleep the day away. Most nights I just vape some weed to try and chill, but the feelings don’t really go away — they’re just always there

  52. [53]

    It was like this wave of sadness hit me, and it hasn’t gone away since

    My birthday was supposed to be a happy time with family, but I just felt hollow and empty. It was like this wave of sadness hit me, and it hasn’t gone away since. I feel so hopeless, and even the things I used to love don’t bring me any happiness anymore. 26 D.4 INSTRUCTIONS AND QUESTIONS We report the exact instructions given to participants when respond...