diff --git a/src/lerobot/configs/recipes/subtask.yaml b/src/lerobot/configs/recipes/subtask.yaml
index b84a03965..c90ca8f78 100644
--- a/src/lerobot/configs/recipes/subtask.yaml
+++ b/src/lerobot/configs/recipes/subtask.yaml
@@ -1,15 +1,5 @@
-# subtask — subtask + action blend, no memory.
-#
-# Same as subtask_mem but with the memory_update sub-recipe dropped, so
-# the model only learns:
-#
-#   high_level_subtask   — predict the subtask from the task.
-#   low_level_execution  — flow loss with [images, subtask, state].
-#
-# No persistent high-level state (no memory note, no plan): the prompt is
-# just the task / subtask. Requires the dataset to carry `subtask`
-# annotations; samples whose `if_present` bindings are missing don't
-# render.
+# Predicts subtasks from tasks and trains subtask-conditioned action flow without memory or plans.
+# Requires `subtask` annotations; samples with missing `if_present` bindings do not render.
 
 blend:
 
@@ -22,9 +12,5 @@ blend:
   low_level_execution:
     weight: 0.70
     messages:
-      # The action expert is conditioned on the SUBTASK — at inference
-      # `HighLevelSubtaskFwd` generates it via the LM head and feeds it
-      # here. `stream: low_level` flips `predict_actions=True` so the
-      # flow loss fires; no text-CE target (subtask prediction is owned
-      # by `high_level_subtask`).
+      # The low-level stream trains action flow on the generated or annotated subtask.
       - {role: user, content: "${subtask}", stream: low_level, if_present: subtask}
diff --git a/src/lerobot/configs/recipes/subtask_mem.yaml b/src/lerobot/configs/recipes/subtask_mem.yaml
index e47d2c9f5..d12fe5009 100644
--- a/src/lerobot/configs/recipes/subtask_mem.yaml
+++ b/src/lerobot/configs/recipes/subtask_mem.yaml
@@ -1,17 +1,5 @@
-# subtask_mem — compact Hi-Robot blend with memory.
-#
-# Trains the core subtask + action objectives and memory updates:
-#
-#   high_level_subtask         — predict the subtask from the task.
-#   low_level_execution        — flow loss with [images, subtask, state].
-#   memory_update              — compress progress into a memory note.
-#
-# Plan is intentionally left out — memory is the only persistent
-# high-level state here, keeping the prompt short.
-#
-# Requires the dataset to carry `subtask` and `memory` annotations.
-# Sub-recipes whose `if_present` bindings are missing simply don't
-# render for that sample.
+# Trains subtask prediction, subtask-conditioned action flow, and memory updates without plans.
+# Requires `subtask` and `memory`; missing `if_present` bindings skip the affected sub-recipe.
 
 blend:
 
@@ -24,26 +12,12 @@ blend:
   low_level_execution:
     weight: 0.60
     messages:
-      # The action expert is conditioned on the SUBTASK — at inference
-      # `HighLevelSubtaskFwd` generates it via the LM head and feeds it
-      # here. `stream: low_level` flips `predict_actions=True` so the
-      # flow loss fires; no text-CE target (subtask prediction is owned
-      # by `high_level_subtask`).
+      # The low-level stream trains action flow on the generated or annotated subtask.
       - {role: user, content: "${subtask}", stream: low_level, if_present: subtask}
 
   memory_update:
-    # At inference, `MemoryUpdateFwd` is triggered only on
-    # `subtask_change` events (sparse). Training densely with
-    # `active_at` — i.e. on every frame inside a subtask interval,
-    # not just the boundary frame — supervises the same
-    # (prior_memory, completed_subtask) → current_memory mapping
-    # against varied observations within the interval. The model
-    # learns a stateless transformation; the *when* to emit lives in
-    # the inference trigger, not the model. Annotations only exist
-    # for ~1% of frames as boundary events, so `emitted_at` would
-    # waste 99% of the blend draws (and silently leak them into a
-    # task-conditioned fallback); `active_at` lifts the renderable
-    # rate to ~87% on this dataset.
+    # `active_at` densifies sparse boundaries while preserving the prior-memory/subtask mapping.
+    # Inference controls update timing through `subtask_change` events.
     weight: 0.15
     bindings:
       prior_memory: "nth_prev(style=memory, offset=1)"
diff --git a/src/lerobot/configs/recipes/subtask_mem_vqa_speech.yaml b/src/lerobot/configs/recipes/subtask_mem_vqa_speech.yaml
index b6424bdf5..1a3d39952 100644
--- a/src/lerobot/configs/recipes/subtask_mem_vqa_speech.yaml
+++ b/src/lerobot/configs/recipes/subtask_mem_vqa_speech.yaml
@@ -1,28 +1,5 @@
-# subtask_mem_vqa_speech — Hi-Robot blend + memory + spoken responses.
-#
-# Extends the compact subtask_mem recipe with VQA and spoken interjection responses:
-#
-#   high_level_subtask         — predict the subtask from the task.
-#   low_level_execution        — flow loss with [images, subtask, state].
-#   memory_update              — compress progress into a memory note.
-#   user_interjection_response — reply to a user interjection with a
-#                                spoken `say` tool call (no plan, no
-#                                subtask text — just the spoken reply).
-#   ask_vqa_{top,wrist}        — camera-grounded VQA.
-#
-# Plan is intentionally left out — memory is the only persistent
-# high-level state here, keeping the prompt short.
-#
-# Requires the dataset to carry `memory`, `interjection` and `say`-tool
-# annotations (the annotation pipeline's memory + interjection modules)
-# in addition to `subtask` and `vqa`. Sub-recipes whose `if_present`
-# bindings are missing simply don't render for that sample, so a
-# dataset without interjections still trains the rest of the blend.
-#
-# Tool-call note: the `say` tool call on the interjection-response turn
-# is flattened to a `<say>...</say>` text marker by the tokenizer step
-# (`_flatten_say_tool_calls`) so the LM head learns to emit exactly the
-# marker the runtime parses back (`_split_plan_and_say`).
+# Adds memory, spoken interjection responses, and camera-grounded VQA to subtask/action training.
+# Missing optional annotations skip only their sub-recipe; `say` tool calls tokenize as `<say>...</say>`.
 
 blend:
 
@@ -35,26 +12,12 @@ blend:
   low_level_execution:
     weight: 0.40
     messages:
-      # The action expert is conditioned on the SUBTASK — at inference
-      # `HighLevelSubtaskFwd` generates it via the LM head and feeds it
-      # here. `stream: low_level` flips `predict_actions=True` so the
-      # flow loss fires; no text-CE target (subtask prediction is owned
-      # by `high_level_subtask`).
+      # The low-level stream trains action flow on the generated or annotated subtask.
       - {role: user, content: "${subtask}", stream: low_level, if_present: subtask}
 
   memory_update:
-    # At inference, `MemoryUpdateFwd` is triggered only on
-    # `subtask_change` events (sparse). Training densely with
-    # `active_at` — i.e. on every frame inside a subtask interval,
-    # not just the boundary frame — supervises the same
-    # (prior_memory, completed_subtask) → current_memory mapping
-    # against varied observations within the interval. The model
-    # learns a stateless transformation; the *when* to emit lives in
-    # the inference trigger, not the model. Annotations only exist
-    # for ~1% of frames as boundary events, so `emitted_at` would
-    # waste 99% of the blend draws (and silently leak them into the
-    # task-conditioned fallback); `active_at` lifts the renderable
-    # rate to ~87% on Hi-Robot-style datasets.
+    # `active_at` densifies sparse boundaries while preserving the prior-memory/subtask mapping.
+    # Inference controls update timing through `subtask_change` events.
     weight: 0.10
     bindings:
       prior_memory: "nth_prev(style=memory, offset=1)"
@@ -74,15 +37,10 @@ blend:
     messages:
       - {role: user, content: "${task}", stream: high_level}
       - {role: user, content: "${interjection}", stream: high_level, if_present: interjection}
-      # Spoken reply only: the assistant turn carries no text content,
-      # just a `say` tool call (`tool_calls_from: speech`). The chat
-      # tokenizer flattens it to a `<say>...</say>` marker, so the
-      # supervised target trains the model to respond to an
-      # interjection with a spoken acknowledgement.
+      # The assistant target is a `say` tool call flattened to a `<say>...</say>` marker.
       - {role: assistant, stream: high_level, target: true, if_present: speech, tool_calls_from: speech}
 
-  # VQA is view-dependent — each camera gets its own sub-recipe so the
-  # resolver disambiguates via `camera=...`. Adjust camera keys to your dataset.
+  # Each camera uses a separate VQA sub-recipe for view-specific binding.
   ask_vqa_top:
     weight: 0.075
     bindings: