Can attribute decomposition improve other interactive reasoning tasks beyond clinical questioning?
This explores whether ALFA's trick — breaking a fuzzy goal like 'ask a good question' into named, separately-trainable attributes — generalizes to interactive reasoning tasks outside the clinical setting where it was demonstrated.
This explores whether attribute decomposition — ALFA's move of splitting 'question quality' into theory-grounded parts like clarity, relevance, and specificity, then training on each separately — can carry beyond clinical questioning into other interactive reasoning tasks. The corpus doesn't test ALFA elsewhere directly, but it does surround the idea with a strong case for why the principle should travel, and where it might break.
The core bet behind ALFA is that a single quality score is too coarse to teach a model anything useful, and that decomposing it into attributes gives the training signal somewhere to land Can models learn to ask genuinely useful clarifying questions?. That same coarse-signal problem shows up elsewhere in the corpus. Supervised fine-tuning raises final-answer accuracy while quietly degrading the reasoning steps, precisely because the reward only looks at the final answer and never at the inferential quality underneath Does supervised fine-tuning improve reasoning or just answers?. Read together, these suggest the real lesson isn't 'clinical questions' — it's that any interactive task scored by a single end-of-task number is a candidate for attribute decomposition, because the missing middle is where the learning signal actually lives.
There's also evidence that the *kind* of decomposition matters, not just that you do it. High-entropy 'forking' tokens turn out to carry most of the reasoning signal — training on the ~20% of decision points that matter matches full training Do high-entropy tokens drive reasoning model improvements?. That's decomposition along a different axis (which *moments* matter rather than which *attributes*), and it points to a transfer caveat: ALFA works because clinical question quality has clean, theory-grounded sub-dimensions. A task without that conceptual scaffolding may not decompose as cleanly.
Two adjacent approaches show the broader family this belongs to. StructRAG routes each query to a task-appropriate knowledge structure — tables, graphs, algorithms — instead of treating all retrieval uniformly, and grounds the move in cognitive-fit theory Can routing queries to task-matched structures improve RAG reasoning?. DialogueReason restructures a single model's monologue into a dialogue between distinct agents, beating monolithic reasoning on tasks that need multiple strategies Can dialogue format help models reason more diversely?. Both are 'decompose the thing the model is doing into named, separately-optimizable pieces' — the same instinct as ALFA, applied to retrieval and to reasoning structure rather than to question quality.
The quiet warning is that decomposition can teach the *form* of good behavior without the substance. Chain-of-thought already imitates the shape of reasoning while failing its logic under distribution shift Does chain-of-thought reasoning actually generalize beyond training data?, and models can compute an answer early then overwrite it with format-compliant filler Do transformers hide reasoning before producing filler tokens?. So the honest answer: attribute decomposition is a portable idea with strong adjacent support for interactive tasks that have clean sub-dimensions — but its payoff depends on whether those attributes capture real reasoning quality or just its surface, a distinction the corpus shows models are very good at faking.
Sources 7 notes
The ALFA framework breaks down question quality into theory-grounded attributes (clarity, relevance, specificity) and trains models on 80K attribute-specific preference pairs. Attribute-specific optimization outperforms single-score training, especially in clinical reasoning where asking the right clarifying question directly impacts decision quality.
Supervised fine-tuning improves final-answer accuracy on benchmarks but cuts Information Gain by 38.9 percent, meaning models generate correct answers through post-hoc rationalization rather than genuine inferential steps. Standard metrics miss this degradation because they only measure final correctness.
Only ~20% of tokens exhibit high entropy as pivotal reasoning decision points; RLVR primarily adjusts these forking tokens. Training exclusively on them matches or exceeds full-gradient performance, revealing that the minority carries the learning signal.
StructRAG demonstrates that selecting knowledge structure type based on query demands—via DPO-trained router choosing among tables, graphs, algorithms, catalogues, and chunks—improves knowledge-intensive reasoning over standard retrieval. The approach grounds this in cognitive load and cognitive fit theory from cognitive science.
DialogueReason, which structures a single model's internal reasoning as dialogue between distinct agents in separate scenes, overcomes monologue reasoning's fixed-strategy and fragmented-attention weaknesses, especially on tasks requiring multiple problem-solving approaches.
DataAlchemy experiments show CoT fails systematically under distributional shifts in task, length, and format. Models produce fluent but logically inconsistent reasoning — imitating reasoning form without valid underlying logic.
Logit lens analysis shows models trained with hidden CoT tokens compute correct answers in layers 1-3, then actively suppress these representations in final layers to produce format-compliant filler output. The reasoning is fully recoverable from lower-ranked token predictions.