Does verbal step-by-step reflection preserve learning signals that abstraction removes?
This explores a tension the corpus frames two ways: whether keeping reasoning in concrete, step-by-step verbal form holds onto useful signal that gets washed out when you compress it into general abstractions — and whether abstraction is purely lossy or sometimes the thing that helps.
This explores whether spelling reasoning out verbally, step by step, preserves learning signal that abstraction strips away — and the corpus answers from both sides, which is the interesting part. On the "verbal preserves signal" side, Reflexion shows agents improve across attempts precisely because they keep their self-diagnoses uncompressed: a binary success/failure signal prevents the model from rationalizing, and storing the full verbal reflection rather than a summary is what keeps it usable next time Can agents learn from failure without updating their weights?. The signal lives in the specifics; squeeze it down and you lose the part you'd actually learn from.
There's a striking mechanism for *why* abstraction quietly removes signal. A WordNet analysis finds that general words (hypernyms) simply occur more often than specific ones (hyponyms), and because LLMs have a frequency bias, preferring the common paraphrase systematically drifts language toward the abstract — erasing exactly the expert-level specificity that carried the information Does word frequency correlate with semantic abstraction?. So abstraction isn't a neutral compression; it has a directional pull toward vagueness. The flip side shows up in tokens: specific reflective words like "Wait" and "Therefore" are mutual-information *peaks* — suppress them and reasoning accuracy drops, while suppressing the same number of random tokens does nothing Do reflection tokens carry more information about correct answers?. The signal is concentrated in the explicit verbal moves, not spread evenly, which is why flattening them costs you.
But the corpus refuses to let abstraction be the villain. RLAD trains a model to generate short reasoning *abstractions* and shows they enforce breadth-first exploration — spending compute on diverse abstractions beats sampling many full solutions, and it cures the "underthinking" failure where depth-only chains tunnel down one path Can abstractions guide exploration better than depth alone?. Here abstraction adds a kind of signal verbose reasoning lacks: structure over the space of strategies. So the honest answer is that the two operate at different altitudes — verbal step-by-step preserves the fine-grained, episode-specific signal; abstraction preserves the strategic, transferable signal. You can even watch them occupy literally distinct regions of the model's activation space, with verbosity steerable as a single linear direction Can we steer reasoning toward brevity without retraining?.
What you didn't know to ask: the deeper question may be *which kind* of knowledge each form carries. One analysis of pretraining documents finds that reasoning generalizes from broad, transferable *procedural* knowledge — how-to patterns — rather than from memorizing specific facts Does procedural knowledge drive reasoning more than factual retrieval?. That suggests abstraction's job is to surface procedure, while verbal step-by-step grounding guards against a different failure: token-level memorization, where "local" copying from the immediately preceding tokens drives up to 67% of reasoning errors as problems get harder Where do memorization errors arise in chain-of-thought reasoning?. And the skeptical reading lurks underneath all of it — if chain-of-thought is partly just constrained imitation of familiar reasoning *forms* rather than genuine inference Does chain-of-thought reasoning reveal genuine inference or pattern matching?, then "preserving the signal" might sometimes mean preserving a convincing imitation. The cleanest way the corpus resolves this: ReAct interleaves verbal reasoning with real external feedback, so each step is grounded against the world rather than against its own abstraction — which is, in the end, the surest way to keep a learning signal real Can interleaving reasoning with real-world feedback prevent hallucination?.
Sources 9 notes
Reflexion demonstrates that unambiguous environmental feedback (success/failure) enables agents to write useful self-diagnoses and improve across episodes without parameter updates. The binary signal prevents rationalization, and keeping reflections uncompressed preserves their usability.
WordNet analysis shows hypernyms (general concepts) occur more frequently than hyponyms (specific ones). Combined with LLMs' frequency bias, this means preferring common paraphrases systematically drifts toward abstraction, erasing expert-level specificity.
Specific tokens like "Wait" and "Therefore" show sharp spikes in mutual information with correct answers. Suppressing them harms reasoning while suppressing equal random tokens does not, and representation recycling improves accuracy 20%.
RLAD jointly trains abstraction and solution generators, showing that allocating test-time compute to diverse abstractions outperforms parallel solution sampling at large budgets. Abstractions create structured breadth-first exploration that prevents the underthinking failure mode of depth-only reasoning chains.
Activation-Steered Compression extracts a single vector from 50 paired examples to reduce chain-of-thought length by 67% while maintaining accuracy and achieving 2.73x speedup. The method is training-free and generalizes across model sizes and domains.
Analysis of 5 million pretraining documents shows reasoning relies on broad, transferable procedural knowledge from diverse sources, unlike factual recall which depends on narrow, document-specific memorization of target facts.
STIM framework identifies local, mid-range, and long-range memorization sources in CoT reasoning. Local memorization—based on preceding tokens—accounts for up to 67% of reasoning errors, especially as complexity increases and distributional shift occurs.
CoT works by constraining models to reproduce familiar reasoning patterns from training, not by enabling novel symbolic reasoning. Performance degrades predictably under distribution shifts—the signature of imitation rather than capability emergence.
ReAct demonstrates that alternating verbal reasoning with external tool queries (Wikipedia API, environment interaction) prevents error propagation by injecting real-world feedback at each step. On knowledge-intensive and interactive tasks, this approach outperforms pure chain-of-thought and reinforcement learning by 10-34% absolute accuracy.