15 problems
Start from what goes wrong. Each page lists the fixes that studies measured, ranked by effect, and a combined prompt you can use.
LLMs bend verdicts toward the user's stated opinion, praise, premises and pushback. The fixes that measured best remove those cues from the prompt or force the model through criteria, evidence and the opposite assumption before it answers.
27 studiesHallucinations drop most when the prompt gives the model a legitimate way not to answer and when you keep only what repeats across runs, instead of trusting a single fluent reply. Rewarded abstention, named failure states, structured checks and extraction instead of rewriting all have measured effects.
26 studiesIn long chats, early rules lose weight against recent text, the model's own past answers become noise it trusts, and cancelled requirements return. The fixes that measured best replace raw history with a short current-state snapshot, restate active rules close to the question, and restart cleanly instead of arguing with the history.
23 studiesLLM judges react to order, authorship, earlier scores, labels and confident wording, not only to quality. The fixes that measured best are de-anchoring, pairwise comparison in both orders, a fresh context per item and separate criteria.
20 studiesModels drop instructions when a prompt holds too many rules, when a requirement sits far from the point where it must be applied, or when a vague directive competes with the task. The best-measured fixes: put the task last, cut unnecessary constraints, remind at the end, and write few, checkable bans.
18 studiesLLM verdicts shift with demographic cues, source prestige, group labels and wording, even when the facts are identical. Telling the model to ignore a cue often makes things worse; removing the cue before the decision, deciding from an extracted fact list, using neutral labels and arguing opposing perspectives reduce it measurably.
17 studiesLong inputs are used unevenly: facts in the middle or at the end are under-weighted, a question buried in the context gets answered badly, and length alone degrades reasoning. Put the task last (or repeat it there), extract the relevant evidence before reasoning, and split broad requests into narrow passes.
15 studiesLLM confidence numbers cluster at 90-100 because they are habits, not measurements. Prompts that make options compete, shrink the scale, attach a cost to being wrong, or move the final decision outside the model produce confidence signals that track accuracy.
15 studiesLLMs break structured output when a format instruction competes with reasoning, when schema fields are ambiguous, and when a closing format requirement fades by the end of a long answer. The fixes that measured best separate reasoning from formatting and turn the schema into explicit instructions with a self-check.
14 studiesLLMs collapse toward the most typical answer, so repeated requests return one idea in many wordings. The fixes that measured best force contrast: maximally different personas, explicit diversity with a ban on repetition, distributions instead of single answers, and leaps to distant domains.
13 studiesAgents and assistants default to finishing the task, so they silently choose one interpretation or invent missing data. Explicit stop conditions, a clarity check before acting, named blocker categories and a status line for tool failures turn those guesses into questions.
12 studiesLLM reviewers fail in two opposite ways: they invent problems in correct code and they explain real bugs away as intended behavior. Both shrink when the prompt supplies explicit criteria, says that a clean result is a valid outcome, and keeps the reviewer from reading the implementation before it has stated what correct looks like.
11 studiesReasoning and chat models spend far more tokens than most tasks need, and the extra length rarely buys accuracy. Route only hard tasks to step-by-step reasoning, remove phrases that trigger deliberation, give scope and stop conditions, and ask for a short answer rather than writing a short prompt.
11 studiesWhen asked to list, extract or check many things at once, models stop early, return the obvious items and do not say what they skipped. Asking for a rule and checking candidates one by one, splitting the task into named categories or small batches, and giving a fill-in template with a 'not mentioned' option close most of the gaps.
9 studiesModels execute instruction-like text found inside documents, emails and pages, and more capable models do it more readily. A separate scanning pass before the real task, explicit task and data zones, and role-based defenses reinforced with examples reduce the risk; none of them is a complete defense on its own.
4 studiesLooking for a method? Browse techniques →