Cross-domain encoding models reveal shared and domain-specific neural representations across language and mathematics.
5 blocks of code here have a permanent identifier (oscr:b:…), listed with the matches below: citable for good, with a page of its own. Select a match and press Cite to copy it. What a block identifier is.
The 6 matches
- [1] § MATERIALS AND METHODS › Functional connectivity analyses ↔ LM_FC_TestData.py, lines 316 to 369 · score 0.73 · right hemisphere, left hemisphere, MNE Connectivity, ROI, math · block
oscr:b:0YE8T0GZD270A-E· Report a mismatch - [2] § MATERIALS AND METHODS › Functional connectivity analyses ↔ LM_FC_TestPred.py, lines 384 to 438 · score 0.72 · right hemisphere, left hemisphere, MNE Connectivity, ROI · block
oscr:b:PRNJ20VJYAXEW-E· Report a mismatch - [3] § MATERIALS AND METHODS › Feature extraction ↔ LM_MakeFeatures_LLM.py, lines 60 to 121 · score 0.64 · GPT NeoX, LLM jp, RakutenAI2, Gemma2, layers, models · block
oscr:b:AVXKZFSNHD1ES-Q· Report a mismatch - [4] § RESULTS › Cross-domain prediction extends beyond simple visual or memory factors and is reproducible across different LLMs ↔ LM_MakeFeatures_LLM.py, lines 207 to 276 · score 0.62 · GPT NeoX, LLM jp, RakutenAI2, Gemma2, Llama3, encoding · block
oscr:b:BPY07ZFPGMN4P-D· Report a mismatch - [5] § MATERIALS AND METHODS › Preprocessing ↔ LM_VertexRidge_Train.py, lines 20 to 65 · score 0.62 · remove motor, temporal delay, regression, filtering, training, components · block
oscr:b:ZZSPZAXX89TC3-2· Report a mismatch - [6] § MATERIALS AND METHODS › Vertex-wise encoding model fitting ↔ LM_VertexRidge_Train.py, lines 20 to 65 · score 0.60 · cross validation, temporal delays, regression, training, vertex, weight · block
oscr:b:ZZSPZAXX89TC3-2· Report a mismatch
Paper
Loaded from Europe PMC by your browser; OSCR does not store it: doi.org · Europe PMC
The paper is loaded when this pane is shown.
The authors' code
Python · 787 lines · 32 KB · CC-BY-4.0 · 2 matches
- import numpy as np
- import os
- import csv
- import sys
- from scipy.signal import decimate
- import copy
- targ_layer = int(sys.argv[1])
- print("Target layer: ", targ_layer)
- model_name = "Llama3"
- from transformers import AutoTokenizer, AutoModelForCausalLM
- model = AutoModelForCausalLM.from_pretrained("elyza/Llama-3-ELYZA-JP-8B", output_hidden_states=True)
- tokenizer = AutoTokenizer.from_pretrained("elyza/Llama-3-ELYZA-JP-8B")
- #model_name = "GPT-Neox"
- #from transformers import GPTNeoXJapaneseTokenizer, GPTNeoXJapaneseModel
- #model = GPTNeoXJapaneseModel.from_pretrained("abeja/gpt-neox-japanese-2.7b", output_hidden_states=True)
- #tokenizer = GPTNeoXJapaneseTokenizer.from_pretrained("abeja/gpt-neox-japanese-2.7b")
- #model_name = "Gemma2"
- #from transformers import AutoTokenizer, AutoModelForCausalLM
- #token = "XXXYYYZZZ"
- #tokenizer = AutoTokenizer.from_pretrained("google/gemma-2-2b-jpn-it", token=token)
- #model = AutoModelForCausalLM.from_pretrained("google/gemma-2-2b-jpn-it", output_hidden_states=True, token=token)
- #model_name = "Qwen25"
- #from transformers import AutoTokenizer, AutoModelForCausalLM
- #model = AutoModelForCausalLM.from_pretrained("abeja/ABEJA-Qwen2.5-7b-Japanese-v0.1", output_hidden_states=True)
- #tokenizer = AutoTokenizer.from_pretrained("abeja/ABEJA-Qwen2.5-7b-Japanese-v0.1")
- #model_name = "RakutenAI2"
- #from transformers import AutoTokenizer, AutoModelForCausalLM
- #model = AutoModelForCausalLM.from_pretrained("Rakuten/RakutenAI-2.0-8x7B", output_hidden_states=True)
- #tokenizer = AutoTokenizer.from_pretrained("Rakuten/RakutenAI-2.0-8x7B")
- #model_name = "LLM-jp"
- #from transformers import AutoTokenizer, AutoModelForCausalLM
- #model = AutoModelForCausalLM.from_pretrained("llm-jp/llm-jp-3-13b", output_hidden_states=True)
- #tokenizer = AutoTokenizer.from_pretrained("llm-jp/llm-jp-3-13b")
- n_dim_pca_initial = 300
- n_dim_pca = 100
- curr_dir = os.getcwd()
- stim_dir = curr_dir + "/Stimuli"
- feature_dir = curr_dir + "/Features"
- np.random.seed(1234)
- def find_subsequence_last(seq, subseq):
- """Return the starting index of the last occurrence of subseq in seq. Return -1 if not found."""
- L, l = len(seq), len(subseq)
- for s in range(L - l, -1, -1):
- if seq[s:s+l] == subseq:
- return s
- return -1
- #Language
- print(" Loading Lang training data...")
- targ_out_lang = []
- for targ_run in range(4): #Use 5th run as test run, 1 to 4th runs are used for training
- print(targ_run)
- with open(os.path.join(stim_dir, "LangStim_Run0{:d}.csv".format(targ_run+1)), encoding="shift-jis") as f:
- reader = csv.reader(f)
- stim_all = [row for row in reader]
- EPS = 1e-12
- for targ_stim in range(len(stim_all)):
- print(' ', targ_stim)
- # Temporarily store embeddings within a sentence/trial
- targ_out_sent = []
- text_len = len(stim_all[targ_stim])
- for within_text_pos in range(text_len):
- # Build the incremental context (w1..wk) in the forward direction
- prefix_words = stim_all[targ_stim][0:(within_text_pos + 1)]
- input_text = ' '.join(prefix_words)
- # Word newly presented at this step
- target_word = stim_all[targ_stim][within_text_pos]
- # Tokenizer
- enc = tokenizer(input_text, return_tensors="pt")
- input_ids = enc["input_ids"].to(model.device)
- attention_mask = enc.get("attention_mask", None)
- if attention_mask is not None:
- attention_mask = attention_mask.to(model.device)
- # Output
- outputs = model(
- input_ids=input_ids,
- attention_mask=attention_mask,
- output_hidden_states=True,
- use_cache=False
- )
- hs = outputs.hidden_states
- # ---- Safely handle the first word (within_text_pos == 0) ----
- T = input_ids.shape[1] # sequence length
- if within_text_pos == 0:
- start, end = 1, T #BOS is shown in Llama3, Gemma2, RakutenAI2, LLM-jp
- #start, end = 0, T #No BOS is shown in GPT-Neox, Qwen25
- else:
- ids_full = input_ids[0].tolist()
- ids_word = tokenizer(" " + target_word, add_special_tokens=False)["input_ids"]
- start = find_subsequence_last(ids_full, ids_word)
- if start == -1:
- print("input_text:", input_text)
- print("target_word:", target_word)
- print("ids_full:", ids_full)
- print("ids_word:", ids_word)
- raise ValueError(f"target_word span not found: {target_word} in {input_text}")
- end = start + len(ids_word)
- # mean pooling
- span_vecs = hs[targ_layer][0, start:end, :]
- targ_out = span_vecs.mean(dim=0).detach().cpu().numpy()
- targ_out_lang.append(targ_out)
- #Math
- print(" Loading Math training data...")
- targ_out_math = []
- for targ_run in range(4): #Use 5th run as test run, 1 to 4th runs are used for training
- print(targ_run)
- with open(os.path.join(stim_dir, "MathStim_Run0{:d}.csv".format(targ_run+1))) as f:
- reader = csv.reader(f)
- stim_all = [row for row in reader]
- EPS = 1e-12
- for targ_stim in range(len(stim_all)):
- print(' ', targ_stim)
- # Temporarily store embeddings within a sentence/trial
- targ_out_sent = []
- text_len = len(stim_all[targ_stim])
- for within_text_pos in range(text_len):
- # Build the incremental context (w1..wk) in the forward direction
- prefix_words = stim_all[targ_stim][0:(within_text_pos + 1)]
- input_text = ' '.join(prefix_words)
- # Word newly presented at this step
- target_word = stim_all[targ_stim][within_text_pos]
- # Tokenizer
- enc = tokenizer(input_text, return_tensors="pt")
- input_ids = enc["input_ids"].to(model.device)
- attention_mask = enc.get("attention_mask", None)
- if attention_mask is not None:
- attention_mask = attention_mask.to(model.device)
- # Output
- outputs = model(
- input_ids=input_ids,
- attention_mask=attention_mask,
- output_hidden_states=True,
- use_cache=False
- )
- hs = outputs.hidden_states
- # ---- Safely handle the first word (within_text_pos == 0) ----
- T = input_ids.shape[1] # sequence length
- if within_text_pos == 0:
- start, end = 1, T #BOS is shown in Llama3, Gemma2, RakutenAI2, LLM-jp
- #start, end = 0, T #No BOS is shown in GPT-Neox, Qwen25
- else:
- ids_full = input_ids[0].tolist()
- ids_word = tokenizer(" " + target_word, add_special_tokens=False)["input_ids"]
- start = find_subsequence_last(ids_full, ids_word)
- if start == -1:
- print("input_text:", input_text)
- print("target_word:", target_word)
- print("ids_full:", ids_full)
- print("ids_word:", ids_word)
- raise ValueError(f"target_word span not found: {target_word} in {input_text}")
- end = start + len(ids_word)
- # mean pooling
- span_vecs = hs[targ_layer][0, start:end, :]
- targ_out = span_vecs.mean(dim=0).detach().cpu().numpy()
- targ_out_math.append(targ_out)
- #Concatenate lang and math features
- t_targ_out_trn = targ_out_lang + targ_out_math
- #Standardize
- mean = np.nanmean(t_targ_out_trn)
- std = np.nanstd(t_targ_out_trn)
- t_targ_st = (t_targ_out_trn - mean) / std
- t_targ_st = np.nan_to_num(t_targ_st)
- # Dimensionality reduction
- print("Performing PCA...")
- from sklearn.decomposition import PCA
- model_pca = PCA(n_components=n_dim_pca_initial)
- model_pca.fit(t_targ_st)
- t_stim_trn = t_targ_st @ model_pca.components_.T
- half_len = int(np.shape(t_stim_trn)[0]/2)
- lang_trn = t_stim_trn[0:half_len, :]
- math_trn = t_stim_trn[half_len:, :]
- print(" Total explained variance: {} %".format(np.round(np.sum(model_pca.explained_variance_ratio_)*100, 3)))
- # Use top 100 dimensions
- use_dimensions = list(range(100))
- t_stim_trn = t_targ_st @ model_pca.components_[use_dimensions].T
- half_len = int(np.shape(t_stim_trn)[0]/2)
- lang_trn = t_stim_trn[0:half_len, :]
- math_trn = t_stim_trn[half_len:, :]
- print(" Total explained variance after correction: {} %".format(np.round(np.sum(model_pca.explained_variance_ratio_[use_dimensions])*100, 3)))
- print("Processing test data based on PCA result ...")
- # Language test run (5th run)
- targ_run = 4
- targ_out_lang = []
- with open(os.path.join(stim_dir, "LangStim_Run0{:d}.csv".format(targ_run+1)), encoding="shift-jis") as f:
- reader = csv.reader(f)
- stim_all = [row for row in reader]
- EPS = 1e-12
- for targ_stim in range(len(stim_all)):
- print(' ', targ_stim)
- # Temporarily store embeddings within a sentence/trial
- targ_out_sent = []
- text_len = len(stim_all[targ_stim])
- for within_text_pos in range(text_len):
- # Build the incremental context (w1..wk) in the forward direction
- prefix_words = stim_all[targ_stim][0:(within_text_pos + 1)]
- input_text = ' '.join(prefix_words)
- # Word newly presented at this step
- target_word = stim_all[targ_stim][within_text_pos]
- # Tokenizer
- enc = tokenizer(input_text, return_tensors="pt")
- input_ids = enc["input_ids"].to(model.device)
- attention_mask = enc.get("attention_mask", None)
- if attention_mask is not None:
- attention_mask = attention_mask.to(model.device)
- # Output (requires hidden states to be returned)
- outputs = model(
- input_ids=input_ids,
- attention_mask=attention_mask,
- output_hidden_states=True,
- use_cache=False
- )
- hs = outputs.hidden_states
- # ---- Safely handle the first word (within_text_pos == 0) ----
- T = input_ids.shape[1] # sequence length
- if within_text_pos == 0:
- start, end = 1, T #BOS is shown in Llama3, Gemma2, RakutenAI2, LLM-jp
- #start, end = 0, T #No BOS is shown in GPT-Neox, Qwen25
- else:
- ids_full = input_ids[0].tolist()
- ids_word = tokenizer(" " + target_word, add_special_tokens=False)["input_ids"]
- start = find_subsequence_last(ids_full, ids_word)
- if start == -1:
- print("input_text:", input_text)
- print("target_word:", target_word)
- print("ids_full:", ids_full)
- print("ids_word:", ids_word)
- raise ValueError(f"target_word span not found: {target_word} in {input_text}")
- end = start + len(ids_word)
- # mean pooling
- span_vecs = hs[targ_layer][0, start:end, :]
- targ_out = span_vecs.mean(dim=0).detach().cpu().numpy()
- targ_out_lang.append(targ_out)
- # Math test run (5th run)
- targ_out_math = []
- with open(os.path.join(stim_dir, "MathStim_Run0{:d}.csv".format(targ_run+1))) as f:
- reader = csv.reader(f)
- stim_all = [row for row in reader]
- EPS = 1e-12
- for targ_stim in range(len(stim_all)):
- print(' ', targ_stim)
- # Temporarily store embeddings within a sentence/trial
- targ_out_sent = []
- text_len = len(stim_all[targ_stim])
- for within_text_pos in range(text_len):
- # Build the incremental context (w1..wk) in the forward direction
- prefix_words = stim_all[targ_stim][0:(within_text_pos + 1)]
- input_text = ' '.join(prefix_words)
- # Word newly presented at this step
- target_word = stim_all[targ_stim][within_text_pos]
- # Tokenizer
- enc = tokenizer(input_text, return_tensors="pt")
- input_ids = enc["input_ids"].to(model.device)
- attention_mask = enc.get("attention_mask", None)
- if attention_mask is not None:
- attention_mask = attention_mask.to(model.device)
- # Output (requires hidden states to be returned)
- outputs = model(
- input_ids=input_ids,
- attention_mask=attention_mask,
- output_hidden_states=True,
- use_cache=False
- )
- hs = outputs.hidden_states
- # ---- Safely handle the first word (within_text_pos == 0) ----
- T = input_ids.shape[1] # sequence length
- if within_text_pos == 0:
- start, end = 1, T #BOS is shown in Llama3, Gemma2, RakutenAI2, LLM-jp
- #start, end = 0, T #No BOS is shown in GPT-Neox, Qwen25
- else:
- ids_full = input_ids[0].tolist()
- ids_word = tokenizer(" " + target_word, add_special_tokens=False)["input_ids"]
- start = find_subsequence_last(ids_full, ids_word)
- if start == -1:
- print("input_text:", input_text)
- print("target_word:", target_word)
- print("ids_full:", ids_full)
- print("ids_word:", ids_word)
- raise ValueError(f"target_word span not found: {target_word} in {input_text}")
- end = start + len(ids_word)
- # mean pooling
- span_vecs = hs[targ_layer][0, start:end, :]
- targ_out = span_vecs.mean(dim=0).detach().cpu().numpy()
- targ_out_math.append(targ_out)
- #Apply PCA based on fitting with training data (no leakage from test to training data)
- t_targ_st = (targ_out_lang - mean) / std
- t_targ_st = np.nan_to_num(t_targ_st)
- lang_test = t_targ_st @ model_pca.components_[use_dimensions].T
- t_targ_st = (targ_out_math - mean) / std
- t_targ_st = np.nan_to_num(t_targ_st)
- math_test = t_targ_st @ model_pca.components_[use_dimensions].T
- with open(os.path.join(stim_dir, "LangJitter1.csv")) as f:
- reader = csv.reader(f)
- jitters_lang = [row for row in reader]
- with open(os.path.join(stim_dir, "MathJitter1.csv")) as f:
- reader = csv.reader(f)
- jitters_math = [row for row in reader]
- with open(os.path.join(stim_dir, "LangJitter2.csv")) as f:
- reader = csv.reader(f)
- jitters_lang_probe = [row for row in reader]
- with open(os.path.join(stim_dir, "MathJitter2.csv")) as f:
- reader = csv.reader(f)
- jitters_math_probe = [row for row in reader]
- print("Feature downsampling, Lang training runs ...")
- # Language, training run
- # Make features for 10 Hz (each 100ms)
- intro_dur = 60 #6s
- outro_dur = 60 #6s
- cue_dur = 5 #0.5s
- word_dur = 7 #0.7s for each (5 words in total)
- probe_dur = 20 #2s
- feedback_dur = 10 #1s of T/F feedback
- stim_lang_trn = []
- for targ_run in range(4):
- probe_id = 0 #Initialize
- #Intro
- stim = np.zeros([intro_dur, n_dim_pca])
- for trial_id in range(0,40):
- #Cue
- stim = np.concatenate([stim, np.zeros([cue_dur, n_dim_pca])], axis=0)
- #Stimuli
- stim = np.concatenate([stim, np.tile(lang_trn[200*targ_run + 5*trial_id, :], [word_dur, 1])], axis=0)
- stim = np.concatenate([stim, np.tile(lang_trn[200*targ_run + 5*trial_id+1, :], [word_dur, 1])], axis=0)
- stim = np.concatenate([stim, np.tile(lang_trn[200*targ_run + 5*trial_id+2, :], [word_dur, 1])], axis=0)
- stim = np.concatenate([stim, np.tile(lang_trn[200*targ_run + 5*trial_id+3, :], [word_dur, 1])], axis=0)
- stim = np.concatenate([stim, np.tile(lang_trn[200*targ_run + 5*trial_id+4, :], [word_dur, 1])], axis=0)
- #Jitter
- stim = np.concatenate([stim, np.zeros([10+10*int(jitters_lang[targ_run][trial_id]), n_dim_pca])], axis=0)
- #Probe
- if int(jitters_lang_probe[targ_run][trial_id]) > 0:
- #Probe
- stim = np.concatenate([stim, np.zeros([probe_dur, n_dim_pca])], axis=0)
- #Feedback
- stim = np.concatenate([stim, np.zeros([feedback_dur, n_dim_pca])], axis=0)
- #Probe jitter
- stim = np.concatenate([stim, np.zeros([10*int(jitters_lang_probe[targ_run][trial_id]), n_dim_pca])], axis=0)
- #Increment of probe index
- probe_id = probe_id + 1
- #Outro
- stim = np.concatenate([stim, np.zeros([outro_dur, n_dim_pca])], axis=0)
- #Exclude Intro
- stim = stim[60:,:]
- stim_lang_trn.append(stim)
- print("Feature downsampling, Math training runs ...")
- # Math training run
- stim_math_trn = []
- for targ_run in range(4):
- probe_id = 0 #Initialize
- #Intro
- stim = np.zeros([intro_dur, n_dim_pca])
- for trial_id in range(0,40):
- #Cue
- stim = np.concatenate([stim, np.zeros([cue_dur, n_dim_pca])], axis=0)
- #Stimuli
- stim = np.concatenate([stim, np.tile(math_trn[200*targ_run + 5*trial_id, :], [word_dur, 1])], axis=0)
- stim = np.concatenate([stim, np.tile(math_trn[200*targ_run + 5*trial_id+1, :], [word_dur, 1])], axis=0)
- stim = np.concatenate([stim, np.tile(math_trn[200*targ_run + 5*trial_id+2, :], [word_dur, 1])], axis=0)
- stim = np.concatenate([stim, np.tile(math_trn[200*targ_run + 5*trial_id+3, :], [word_dur, 1])], axis=0)
- stim = np.concatenate([stim, np.tile(math_trn[200*targ_run + 5*trial_id+4, :], [word_dur, 1])], axis=0)
- #Jitter
- stim = np.concatenate([stim, np.zeros([10+10*int(jitters_math[targ_run][trial_id]), n_dim_pca])], axis=0)
- #Probe
- if int(jitters_math_probe[targ_run][trial_id]) > 0:
- #Probe
- #stim = np.concatenate([stim, np.tile(math_probe_trn[8*targ_run + probe_id, :], [probe_dur, 1])], axis=0)
- stim = np.concatenate([stim, np.zeros([probe_dur, n_dim_pca])], axis=0)
- #Feedback
- stim = np.concatenate([stim, np.zeros([feedback_dur, n_dim_pca])], axis=0)
- #Probe jitter
- stim = np.concatenate([stim, np.zeros([10*int(jitters_math_probe[targ_run][trial_id]), n_dim_pca])], axis=0)
- #Increment of probe index
- probe_id = probe_id + 1
- #Outro
- stim = np.concatenate([stim, np.zeros([outro_dur, n_dim_pca])], axis=0)
- #Exclude Intro
- stim = stim[60:,:]
- stim_math_trn.append(stim)
- print("Feature downsampling, Lang test run ...")
- #Language, test run
- #Make features for 10 Hz (each 100ms)
- intro_dur = 60 #6s
- outro_dur = 60 #6s
- cue_dur = 5 #0.5s
- word_dur = 7 #0.7s for each (5 words in total)
- probe_dur = 20 #2s
- feedback_dur = 10 #1s of T/F feedback
- stim_lang_test = []
- targ_run = 4
- probe_id = 0 #Initialize
- #Intro
- stim = np.zeros([intro_dur, n_dim_pca])
- for trial_id in range(0,40):
- #Cue
- stim = np.concatenate([stim, np.zeros([cue_dur, n_dim_pca])], axis=0)
- #Stimuli
- stim = np.concatenate([stim, np.tile(lang_test[5*trial_id, :], [word_dur, 1])], axis=0)
- stim = np.concatenate([stim, np.tile(lang_test[5*trial_id+1, :], [word_dur, 1])], axis=0)
- stim = np.concatenate([stim, np.tile(lang_test[5*trial_id+2, :], [word_dur, 1])], axis=0)
- stim = np.concatenate([stim, np.tile(lang_test[5*trial_id+3, :], [word_dur, 1])], axis=0)
- stim = np.concatenate([stim, np.tile(lang_test[5*trial_id+4, :], [word_dur, 1])], axis=0)
- #Jitter
- stim = np.concatenate([stim, np.zeros([10+10*int(jitters_lang[targ_run][trial_id]), n_dim_pca])], axis=0) #test run id = 4
- #Probe
- if int(jitters_lang_probe[targ_run][trial_id]) > 0:
- #Probe
- #stim = np.concatenate([stim, np.tile(lang_probe_test[probe_id, :], [probe_dur, 1])], axis=0)
- stim = np.concatenate([stim, np.zeros([probe_dur, n_dim_pca])], axis=0)
- #Feedback
- stim = np.concatenate([stim, np.zeros([feedback_dur, n_dim_pca])], axis=0)
- #Probe jitter
- stim = np.concatenate([stim, np.zeros([10*int(jitters_lang_probe[targ_run][trial_id]), n_dim_pca])], axis=0)
- #Increment of probe index
- probe_id = probe_id + 1
- #Outro
- stim = np.concatenate([stim, np.zeros([outro_dur, n_dim_pca])], axis=0)
- #Exclude Intro
- stim = stim[60:,:]
- stim_lang_test = stim
- print("Feature downsampling, Math test run ...")
- #Math, test run
- stim_math_test = []
- targ_run = 4
- probe_id = 0 #Initialize
- #Intro
- stim = np.zeros([intro_dur, n_dim_pca])
- for trial_id in range(0,40):
- #Cue
- stim = np.concatenate([stim, np.zeros([cue_dur, n_dim_pca])], axis=0)
- #Stimuli
- stim = np.concatenate([stim, np.tile(math_test[5*trial_id, :], [word_dur, 1])], axis=0)
- stim = np.concatenate([stim, np.tile(math_test[5*trial_id+1, :], [word_dur, 1])], axis=0)
- stim = np.concatenate([stim, np.tile(math_test[5*trial_id+2, :], [word_dur, 1])], axis=0)
- stim = np.concatenate([stim, np.tile(math_test[5*trial_id+3, :], [word_dur, 1])], axis=0)
- stim = np.concatenate([stim, np.tile(math_test[5*trial_id+4, :], [word_dur, 1])], axis=0)
- #Jitter
- stim = np.concatenate([stim, np.zeros([10+10*int(jitters_math[targ_run][trial_id]), n_dim_pca])], axis=0) #test run id = 4
- #Probe
- if int(jitters_math_probe[targ_run][trial_id]) > 0:
- #Probe
- stim = np.concatenate([stim, np.zeros([probe_dur, n_dim_pca])], axis=0)
- #Feedback
- stim = np.concatenate([stim, np.zeros([feedback_dur, n_dim_pca])], axis=0)
- #Probe jitter
- stim = np.concatenate([stim, np.zeros([10*int(jitters_math_probe[targ_run][trial_id]), n_dim_pca])], axis=0)
- #Increment of probe index
- probe_id = probe_id + 1
- #Outro
- stim = np.concatenate([stim, np.zeros([outro_dur, n_dim_pca])], axis=0)
- #Exclude Intro
- stim = stim[60:,:]
- stim_math_test = stim
- #Down sampling to 1 Hz
- stim_lang_trn_tmp = np.zeros([np.shape(stim_lang_trn)[0], int(np.shape(stim_lang_trn)[1]/10), np.shape(stim_lang_trn)[2]])
- stim_math_trn_tmp = np.zeros([np.shape(stim_math_trn)[0], int(np.shape(stim_math_trn)[1]/10), np.shape(stim_math_trn)[2]])
- for ff in range(n_dim_pca):
- for rr in range(4):
- stim_lang_trn_tmp[rr][:, ff] = decimate(stim_lang_trn[rr][:, ff], 10, ftype='fir')
- stim_math_trn_tmp[rr][:, ff] = decimate(stim_math_trn[rr][:, ff], 10, ftype='fir')
- stim_lang_trn= copy.copy(stim_lang_trn_tmp)
- stim_math_trn = copy.copy(stim_math_trn_tmp)
- stim_lang_test_tmp = np.zeros([int(np.shape(stim_lang_test)[0]/10), np.shape(stim_lang_test)[1]])
- stim_math_test_tmp = np.zeros([int(np.shape(stim_math_test)[0]/10), np.shape(stim_math_test)[1]])
- for ff in range(n_dim_pca):
- stim_lang_test_tmp[:, ff] = decimate(stim_lang_test[:, ff], 10, ftype='fir')
- stim_math_test_tmp[:, ff] = decimate(stim_math_test[:, ff], 10, ftype='fir')
- stim_lang_test = copy.copy(stim_lang_test_tmp)
- stim_math_test = copy.copy(stim_math_test_tmp)
- np.save(os.path.join(feature_dir, f'Features_{model_name}_Layer{targ_layer:02d}_LangTrn'), stim_lang_trn)
- np.save(os.path.join(feature_dir, f'Features_{model_name}_Layer{targ_layer:02d}_LangTest'), stim_lang_test)
- np.save(os.path.join(feature_dir, f'Features_{model_name}_Layer{targ_layer:02d}_MathTrn'), stim_math_trn)
- np.save(os.path.join(feature_dir, f'Features_{model_name}_Layer{targ_layer:02d}_MathTest'), stim_math_test)
- print("Processing control data based on PCA result ...")
- #Control run
- targ_out_lang_ctrl = []
- with open(os.path.join(stim_dir, "LangStim_Ctrl.csv"), encoding="shift-jis") as f:
- reader = csv.reader(f)
- stim_all = [row for row in reader]
- for targ_stim in range(len(stim_all)):
- print(' ', targ_stim)
- # Temporarily store embeddings within a sentence/trial
- targ_out_sent = []
- text_len = len(stim_all[targ_stim])
- for within_text_pos in range(text_len):
- # Build the incremental context (w1..wk) in the forward direction
- prefix_words = stim_all[targ_stim][0:(within_text_pos + 1)]
- input_text = ' '.join(prefix_words) # Use a fixed concatenation rule
- # Word newly presented at this step
- target_word = stim_all[targ_stim][within_text_pos]
- # Tokenizer
- enc = tokenizer(input_text, return_tensors="pt")
- input_ids = enc["input_ids"].to(model.device)
- attention_mask = enc.get("attention_mask", None)
- if attention_mask is not None:
- attention_mask = attention_mask.to(model.device)
- # Output (requires hidden states to be returned)
- outputs = model(
- input_ids=input_ids,
- attention_mask=attention_mask,
- output_hidden_states=True,
- use_cache=False
- )
- hs = outputs.hidden_states
- # ---- Safely handle the first word (within_text_pos == 0) ----
- T = input_ids.shape[1] # sequence length
- if within_text_pos == 0:
- start, end = 1, T #BOS is shown in Llama3, Gemma2, RakutenAI2, LLM-jp
- #start, end = 0, T #No BOS is shown in GPT-Neox, Qwen25
- else:
- ids_full = input_ids[0].tolist()
- ids_word = tokenizer(" " + target_word, add_special_tokens=False)["input_ids"]
- start = find_subsequence_last(ids_full, ids_word)
- if start == -1:
- print("input_text:", input_text)
- print("target_word:", target_word)
- print("ids_full:", ids_full)
- print("ids_word:", ids_word)
- raise ValueError(f"target_word span not found: {target_word} in {input_text}")
- end = start + len(ids_word)
- # mean pooling
- span_vecs = hs[targ_layer][0, start:end, :]
- targ_out = span_vecs.mean(dim=0).detach().cpu().numpy()
- targ_out_lang_ctrl.append(targ_out)
- targ_out_math_ctrl = []
- with open(os.path.join(stim_dir, "MathStim_Ctrl.csv")) as f:
- reader = csv.reader(f)
- stim_all = [row for row in reader]
- for targ_stim in range(len(stim_all)):
- print(' ', targ_stim)
- # Temporarily store embeddings within a sentence/trial
- targ_out_sent = []
- text_len = len(stim_all[targ_stim])
- for within_text_pos in range(text_len):
- # Build the incremental context (w1..wk) in the forward direction
- prefix_words = stim_all[targ_stim][0:(within_text_pos + 1)]
- input_text = ' '.join(prefix_words)
- # Word newly presented at this step
- target_word = stim_all[targ_stim][within_text_pos]
- # Tokenizer
- enc = tokenizer(input_text, return_tensors="pt")
- input_ids = enc["input_ids"].to(model.device)
- attention_mask = enc.get("attention_mask", None)
- if attention_mask is not None:
- attention_mask = attention_mask.to(model.device)
- # Output (requires hidden states to be returned)
- outputs = model(
- input_ids=input_ids,
- attention_mask=attention_mask,
- output_hidden_states=True,
- use_cache=False
- )
- hs = outputs.hidden_states
- # ---- Safely handle the first word (within_text_pos == 0) ----
- T = input_ids.shape[1] # sequence length
- if within_text_pos == 0:
- start, end = 1, T #BOS is shown in Llama3, Gemma2, RakutenAI2, LLM-jp
- #start, end = 0, T #No BOS is shown in GPT-Neox, Qwen25
- else:
- ids_full = input_ids[0].tolist()
- ids_word = tokenizer(" " + target_word, add_special_tokens=False)["input_ids"]
- start = find_subsequence_last(ids_full, ids_word)
- if start == -1:
- print("input_text:", input_text)
- print("target_word:", target_word)
- print("ids_full:", ids_full)
- print("ids_word:", ids_word)
- raise ValueError(f"target_word span not found: {target_word} in {input_text}")
- end = start + len(ids_word)
- # mean pooling
- span_vecs = hs[targ_layer][0, start:end, :]
- targ_out = span_vecs.mean(dim=0).detach().cpu().numpy()
- targ_out_math_ctrl.append(targ_out)
- # Apply PCA based on fitting with training data (no leakage from test to training data)
- t_targ_st = (targ_out_lang_ctrl - mean) / std
- t_targ_st = np.nan_to_num(t_targ_st)
- lang_ctrl = t_targ_st @ model_pca.components_[use_dimensions].T
- t_targ_st = (targ_out_math_ctrl - mean) / std
- t_targ_st = np.nan_to_num(t_targ_st)
- math_ctrl = t_targ_st @ model_pca.components_[use_dimensions].T
- with open(os.path.join(stim_dir, "LangJitter1_Ctrl.csv")) as f:
- reader = csv.reader(f)
- jitters_lang_ctrl = [row for row in reader]
- with open(os.path.join(stim_dir, "MathJitter1_Ctrl.csv")) as f:
- reader = csv.reader(f)
- jitters_math_ctrl = [row for row in reader]
- with open(os.path.join(stim_dir, "LangJitter2_Ctrl.csv")) as f:
- reader = csv.reader(f)
- jitters_lang_ctrl_probe = [row for row in reader]
- with open(os.path.join(stim_dir, "MathJitter2_Ctrl.csv")) as f:
- reader = csv.reader(f)
- jitters_math_ctrl_probe = [row for row in reader]
- print("Feature downsampling, Control runs ...")
- #Language, control run
- #Make features for 10 Hz (each 100ms)
- intro_dur = 60 #6s
- outro_dur = 60 #6s
- cue_dur = 5 #0.5s
- word_dur = 7 #0.7s for each (5 words in total)
- probe_dur = 20 #2s
- feedback_dur = 10 #1s of T/F feedback
- stim_lang_ctrl = []
- probe_id = 0 #Initialize
- #Intro
- stim = np.zeros([intro_dur, n_dim_pca])
- for trial_id in range(0,40):
- #Cue
- stim = np.concatenate([stim, np.zeros([cue_dur, n_dim_pca])], axis=0)
- #Stimuli
- stim = np.concatenate([stim, np.tile(lang_ctrl[5*trial_id, :], [word_dur, 1])], axis=0)
- stim = np.concatenate([stim, np.tile(lang_ctrl[5*trial_id+1, :], [word_dur, 1])], axis=0)
- stim = np.concatenate([stim, np.tile(lang_ctrl[5*trial_id+2, :], [word_dur, 1])], axis=0)
- stim = np.concatenate([stim, np.tile(lang_ctrl[5*trial_id+3, :], [word_dur, 1])], axis=0)
- stim = np.concatenate([stim, np.tile(lang_ctrl[5*trial_id+4, :], [word_dur, 1])], axis=0)
- #Jitter
- stim = np.concatenate([stim, np.zeros([10+10*int(jitters_lang_ctrl[0][trial_id]), n_dim_pca])], axis=0) #test run id = 4
- #Probe
- if int(jitters_lang_ctrl_probe[0][trial_id]) > 0:
- #Probe
- #stim = np.concatenate([stim, np.tile(lang_probe_test[probe_id, :], [probe_dur, 1])], axis=0)
- stim = np.concatenate([stim, np.zeros([probe_dur, n_dim_pca])], axis=0)
- #Feedback
- stim = np.concatenate([stim, np.zeros([feedback_dur, n_dim_pca])], axis=0)
- #Probe jitter
- stim = np.concatenate([stim, np.zeros([10*int(jitters_lang_ctrl_probe[0][trial_id]), n_dim_pca])], axis=0)
- #Increment of probe index
- probe_id = probe_id + 1
- #Outro
- stim = np.concatenate([stim, np.zeros([outro_dur, n_dim_pca])], axis=0)
- #Exclude Intro
- stim = stim[60:,:]
- stim_lang_ctrl = stim
- #Math, test run
- stim_math_ctrl = []
- probe_id = 0 #Initialize
- #Intro
- stim = np.zeros([intro_dur, n_dim_pca])
- for trial_id in range(0,40):
- #Cue
- stim = np.concatenate([stim, np.zeros([cue_dur, n_dim_pca])], axis=0)
- #Stimuli
- stim = np.concatenate([stim, np.tile(math_ctrl[5*trial_id, :], [word_dur, 1])], axis=0)
- stim = np.concatenate([stim, np.tile(math_ctrl[5*trial_id+1, :], [word_dur, 1])], axis=0)
- stim = np.concatenate([stim, np.tile(math_ctrl[5*trial_id+2, :], [word_dur, 1])], axis=0)
- stim = np.concatenate([stim, np.tile(math_ctrl[5*trial_id+3, :], [word_dur, 1])], axis=0)
- stim = np.concatenate([stim, np.tile(math_ctrl[5*trial_id+4, :], [word_dur, 1])], axis=0)
- #Jitter
- stim = np.concatenate([stim, np.zeros([10+10*int(jitters_math_ctrl[0][trial_id]), n_dim_pca])], axis=0) #test run id = 4
- #Probe
- if int(jitters_math_ctrl_probe[0][trial_id]) > 0:
- #Probe
- #stim = np.concatenate([stim, np.tile(math_probe_test[probe_id, :], [probe_dur, 1])], axis=0)
- stim = np.concatenate([stim, np.zeros([probe_dur, n_dim_pca])], axis=0)
- #Feedback
- stim = np.concatenate([stim, np.zeros([feedback_dur, n_dim_pca])], axis=0)
- #Probe jitter
- stim = np.concatenate([stim, np.zeros([10*int(jitters_math_ctrl_probe[0][trial_id]), n_dim_pca])], axis=0)
- #Increment of probe index
- probe_id = probe_id + 1
- #Outro
- stim = np.concatenate([stim, np.zeros([outro_dur, n_dim_pca])], axis=0)
- #Exclude Intro
- stim = stim[60:,:]
- stim_math_ctrl = stim
- # Downsampling to 1 Hz
- stim_lang_ctrl_tmp = np.zeros([int(np.shape(stim_lang_ctrl)[0]/10), np.shape(stim_lang_ctrl)[1]])
- stim_math_ctrl_tmp = np.zeros([int(np.shape(stim_math_ctrl)[0]/10), np.shape(stim_math_ctrl)[1]])
- for ff in range(n_dim_pca):
- stim_lang_ctrl_tmp[:, ff] = decimate(stim_lang_ctrl[:, ff], 10, ftype='fir')
- stim_math_ctrl_tmp[:, ff] = decimate(stim_math_ctrl[:, ff], 10, ftype='fir')
- stim_lang_ctrl = copy.copy(stim_lang_ctrl_tmp)
- stim_math_ctrl = copy.copy(stim_math_ctrl_tmp)
- np.save(os.path.join(feature_dir, f'Features_{model_name}_Layer{targ_layer:02d}_LangCtrl'), stim_lang_ctrl)
- np.save(os.path.join(feature_dir, f'Features_{model_name}_Layer{targ_layer:02d}_MathCtrl'), stim_math_ctrl)
LM_MakeFeatures_LLM.py, under CC-BY-4.0 · at the source
Overview
- Graduate School of Information Science and Technology, The University of Tokyo, Tokyo, Japan
- Center for Information and Neural Networks, National Institute of Information and Communications Technology, Suita, Japan
- Graduate School of Medical and Dental Sciences, Institute of Science Tokyo, Tokyo, Japan
- Graduate School of Frontier Biosciences, The University of Osaka, Suita, Japan
- Graduate School of Medicine, The University of Osaka, Suita, Japan
Abstract
Whether language and mathematics rely on shared or distinct neural representations remains unresolved in cognitive neuroscience. Here, we combined latent features from a large language model with functional magnetic resonance imaging encoding models to test cross-domain generalization in 32 participants performing sentence comprehension and calculation tasks. Cross-domain prediction identified cortical regions showing partially shared representations, most consistently in the left area 55b, while control analyses suggested that these effects could not be fully explained by low-level visual or task-general factors. Domain-specificity contrasts revealed stronger language-related prediction in the left superior temporal sulcus and angular gyrus and math-related prediction in the left precentral sulcus and bilateral intraparietal sulci. Model-weight analyses revealed distinct weight profiles associated with shared and domain-specific prediction patterns. Cross-domain regions also exhibited task-dependent functional coupling with language- and math-related networks. Together, these findings suggest that language and mathematics involve partially shared neural representations alongside domain-specific cortical organization, helping reconcile previous contrasting views on their neural basis.
Reproduced under the paper's license (CC BY-NC), from the paper cited above.
Repository
Its files are read in the Code ↔ Paper reader above, with 6 matches between paragraphs and lines of code.
Zenodo 20361844
Availability: 1 check, the latest on 21 November 2026: the link answers (HTTP 200)
- 21 November 2026: the link answers (HTTP 200)
13 files
- LM_FC_TestData.py, Python, 371 lines, 1 match
- LM_FC_TestPred.py, Python, 438 lines, 1 match
- LM_GroupStat.py, Python, 104 lines
- LM_MakeFeatures_AlexNet.
py , Python, 482 lines - LM_MakeFeatures_LLM.py, Python, 787 lines, 2 matches
- LM_ModInvar_Bar.py, Python, 167 lines
- LM_PCA.py, Python, 334 lines
- LM_PredAccCorrelation.py
, Python, 254 lines - LM_PredAcc_Bar.py, Python, 183 lines
- LM_Utils.py, Python, 161 lines
- LM_VertexRidge_Test.py, Python, 235 lines
- LM_VertexRidge_Train.py, Python, 70 lines, 2 matches
- LM_WeightCorrelation.py, Python, 218 lines
The paper's code and data availability statement is in the Data section.
Tracing map
Proposed by the machine: these links were found in the paper and verified at the source, without human review. The map will receive a Zenodo DOI once one of the paper's authors has validated it with their ORCID.
What the map holds:
- 1 repository of the authors' code, each at its verified commit, with its license and how the link was found in the paper;
- 13 scripts, each with its path and the digest of its content;
- 6 matches between paragraphs of the paper and lines of the code (method lexical-v1);
- neither the text of the paper nor the code itself.
Its JSON (tracing-map.json) is deposited on Zenodo with its DOI once the map is validated.
Data
No dataset and no data link were found in the paper.
Data, code, and materials availability
All data and code needed to evaluate and reproduce the results in the paper are present in the paper and/
Reproduced under the paper's license (CC BY-NC), from the paper cited above.
Versions
The history of this record: each version stored by the harvester or made by a correction of its authors or of the maintainers of its code, and what changed in its facts. The texts of the paper (its abstract, its availability statements) are not part of it; versions that changed only those are not listed.
Version 1, 21 November 2026: the first record
Recorded: type, language, journal, volume, issue, pages, dates, 3 authors, 11 MeSH terms, 4 funders, 92 references.
Cite
This paper
Nakai, T., Kubo, R., & Nishimoto, S. (2026). Cross-domain encoding models reveal shared and domain-specific neural representations across language and mathematics. Science advances, 12(47), eaej5555. https://
BibTeX
@article{nakai2026cross,
author = {Nakai, Tomoya and Kubo, Rieko and Nishimoto, Shinji},
title = {{Cross-domain encoding models reveal shared and domain-specific neural representations across language and mathematics}},
journal = {Science advances},
year = {2026},
month = nov,
volume = {12},
number = {47},
pages = {eaej5555},
publisher = {American Association for the Advancement of Science},
issn = {2375-2548},
doi = {10.1126/
url = {https://
pmid = {43050544},
pmcid = {PMC13781126}
}
RIS
TY - JOUR
AU - Nakai, Tomoya
AU - Kubo, Rieko
AU - Nishimoto, Shinji
TI - Cross-domain encoding models reveal shared and domain-specific neural representations across language and mathematics
T2 - Science advances
J2 - Sci Adv
PY - 2026
DA - 2026/
VL - 12
IS - 47
SP - eaej5555
SN - 2375-2548
PB - American Association for the Advancement of Science
DO - 10.1126/
UR - https://
LA - en
ER -
CSL-JSON
{
"id": "10.1126/
"type": "article-journal",
"title": "Cross-domain encoding models reveal shared and domain-specific neural representations across language and mathematics",
"container-title": "Science advances",
"author": [
{
"family": "Nakai",
"given": "Tomoya"
},
{
"family": "Kubo",
"given": "Rieko"
},
{
"family": "Nishimoto",
"given": "Shinji"
}
],
"container-title-short":
"volume": "12",
"issue": "47",
"page": "eaej5555",
"DOI": "10.1126/
"PMID": "43050544",
"PMCID": "PMC13781126",
"ISSN": "2375-2548",
"publisher": "American Association for the Advancement of Science",
"URL": "https://
"language": "en",
"issued": {
"date-parts": [
[
2026,
11,
18
]
]
}
}
The tracing map gets a citation of its own once an author has validated it and it has a DOI.
Similar papers
The papers with a page that share the most with this one: the tools found in their code, their categories, datasets, cited references and authors, the rarest counting most.
- [1] doi:10.1016/j.isci.2026.117180 [code]
- Developmental changes in similarity between neural representations of mental arithmetic and artificial neural networks.Journal: iScienceIn common: Hugging Face Transformers, Nilearn, statsmodels, 8 other tools, 17 references, author Tomoya Nakai
- [2] doi:10.1038/s42003-025-08134-4 [code]
- Cortical representational geometry of diverse tasks reveals subject-specific and subject-invariant cognitive structures.Journal: Communications biologyIn common: Nilearn, statsmodels, NiBabel, 6 other tools, 14 references, 2 authors
- [3] doi:10.1038/s41467-024-49173-5 [code]
- Shared functional specialization in transformer-based language models and the human brain.Journal: Nature communicationsIn common: Hugging Face Transformers, Nilearn, statsmodels, 8 other tools, 10 references
- [4] doi:10.1038/s41562-025-02231-4 [code]
- Conversational content is organized across multiple timescales in the brain.Journal: Nature human behaviourIn common: seaborn, scikit-learn, pandas, 3 other tools, 11 references, author Shinji Nishimoto
- [5] doi:10.1038/s41597-025-05462-2 [code]
- The "Podcast" ECoG dataset for modeling neural activity during natural language comprehension.Journal: Scientific dataIn common: Nilearn, statsmodels, NiBabel, 6 other tools, 10 references
- [6] doi:10.1038/s41467-026-75745-8 [code]
- A language network in the individualized functional connectomes of 1199 human brains doing arbitrary tasks.Journal: Nature communicationsIn common: Nilearn, Pillow, NiBabel, 6 other tools, 9 references
- [7] doi:10.1371/journal.pbio.3003820 [code]
- A cortical semantic map integrates fractions and integers in the human brain.Journal: PLoS biologyIn common: Nilearn, statsmodels, NiBabel, 6 other tools, 10 references
- [8] doi:10.1038/s41467-025-66731-7 [code]
- Text-to-music generation models capture musical semantic representations in the human brain.Journal: Nature communicationsIn common: statsmodels, PyTorch, scikit-learn, 3 other tools, 4 references, 2 authors
- [9] doi:10.1126/sciadv.adw1464 [code]
- Mind captioning: Evolving descriptive text of mental content from human brain activity.Journal: Science advancesIn common: Hugging Face Transformers, Pillow, NiBabel, 7 other tools, 8 references
- [10] doi:10.1016/j.dcn.2024.101470
- Encoding models for developmental cognitive computational neuroscience: Promise, challenges, and potential.Journal: Developmental cognitive neuroscienceIn common: 11 references
Contribute
The authors of this paper can claim it, correct its record and validate its tracing map, and the maintainers of its code (its owner, or a public member of its organization) correct what it says of their repository; anyone signed in can ask for its removal. Every request goes to OSCR's own machine, which answers it; your account page follows them.
Sign in with ORCID to claim this paper as one of its authors, correct its record or validate its tracing map: when the paper's metadata lists your ORCID iD, you are recognized at once. Maintainers of its code: sign in with GitHub, then claim the repository on your account page.
Claim this paper
Correct its record
Say what each link of this record is, remove the ones that are not the paper's, add the ones that are missing. The correction becomes a new version of the record, in its Versions section.
Validate its tracing map
You validate the map as this page shows it: 1 repository of the authors' code, each at its verified commit and with its license, 13 scripts, and 6 matches between paragraphs and code (see the Code and Map sections). It then receives a DOI on Zenodo, with you (your ORCID iD) and OSCR as its creators; the code itself is not deposited.
The map's fingerprint: sha256:2b3301bb8355180b…
Add the badge to the code's README
If you are an author or a maintainer of this paper's code, you can add this badge. It says what OSCR knows of the paper (its code, its validated tracing map) and links the code to this page. Paste one of these snippets under the README's title. Keep the two marks around it, so that the badge can be replaced in place and is never added twice. OSCR changes nothing in your repository.
Markdown
<!-- oscr-badge:start -->
[](https://openscicode.org/badge/doi/10.1126/sciadv.aej5555)
<!-- oscr-badge:end -->reStructuredText
.. oscr-badge:start
.. image:: https://openscicode.org/badge/doi/10.1126/sciadv.aej5555.svg
:target: https://openscicode.org/badge/doi/10.1126/sciadv.aej5555
:alt: OSCR
.. oscr-badge:endHTML
<!-- oscr-badge:start -->
<a href="https://openscicode.org/badge/doi/10.1126/sciadv.aej5555"><img src="https://openscicode.org/badge/doi/10.1126/sciadv.aej5555.svg" alt="OSCR"></a>
<!-- oscr-badge:end -->If you choose to, you can have the badge kept up to date for you, with oscr badge --write from the command line or with the GitHub Action in your repository's workflows. Every other form and option (a card, AsciiDoc, LaTeX, shields.io) is on the badge page.
Request its removal
To ask OSCR to remove this record, the copies of its authors' scripts or its tracing map, use the removal request page: signed in, you say who you are, what to remove and why, then review and confirm the request. Published rules decide every request (how).
Discussion, reproductions, activity
Discussion: code errors and mismatches between the code and the paper about this paper and its code, from readers and its authors, beside the paper and its tracing map.
Report a mismatch between the code and the paper · Report a code error (signed in)
Reproductions: reports from readers who ran the authors' code and did not get the paper's result, what they ran, with which environment, commit and data.
Report a reproduction failure (signed in)
Activity: what happens around this paper: new versions of its record, its map's validation, discussions and reproductions. It opens with sign-in.
